From acdef2fa59c1e077138f8e6695113cefee81e03a Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 6 Feb 2024 10:27:54 +0900
Subject: [PATCH 01/28] Add handling empty predicted tensor

---
 .../visual_prompters/zero_shot_segment_anything.py  |  4 ++++
 .../test_zero_shot_segment_anything.py              | 13 ++++++++-----
 2 files changed, 12 insertions(+), 5 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 14c8e5dd6f2..8321714d577 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -163,6 +163,10 @@ def _point_selection(
 
         point_coords = torch.where(mask_sim > threshold)
         fg_coords_scores = torch.stack(point_coords[::-1] + (mask_sim[point_coords],), dim=0).T
+        
+        # to handle empty tensor
+        len_fg_coords_scores = len(fg_coords_scores)
+        fg_coords_scores = F.pad(fg_coords_scores, (0, 0, 0, max(0, 1 - len_fg_coords_scores)), value=-1)
 
         ratio = self.image_size / original_size.max()
         width = (original_size[1] * ratio).to(torch.int64)
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 4437fdc1f42..a03fa4b838a 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -114,18 +114,21 @@ def test_get_prompt_candidates(self, mocker) -> None:
         assert bg_coords == "bg_coords"
 
     @e2e_pytest_unit
-    def test_point_selection(self) -> None:
+    @pytest.mark.parametrize(
+        "mask_sim,expected",
+        [
+            (torch.arange(0.1, 1.0, 0.1).reshape(3, 3), torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]])),
+            (torch.zeros(3, 3), torch.tensor([[-1, -1, -1]]))
+        ])
+    def test_point_selection(self, mask_sim: torch.Tensor, expected: torch.Tensor) -> None:
         """Test _point_selection."""
-        mask_sim = torch.arange(0.1, 1.0, 0.1).reshape(self.prompt_getter.image_size, self.prompt_getter.image_size)
-
         points_scores, bg_coords = self.prompt_getter._point_selection(
             mask_sim=mask_sim,
             original_size=torch.tensor([self.prompt_getter.image_size, self.prompt_getter.image_size]),
             threshold=torch.tensor([[0.5]]),
         )
 
-        assert torch.equal(points_scores, torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]))
-        assert torch.equal(bg_coords, torch.tensor([[0, 0]]))
+        assert torch.equal(points_scores, expected)
 
 
 class TestZeroShotSegmentAnything:

From d6978d56522fe63ff81eb6f4311262feda63e543 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 6 Feb 2024 13:37:50 +0900
Subject: [PATCH 02/28] Update image encoder

---
 .../models/encoders/sam_image_encoder.py      | 23 ++-------
 .../visual_prompters/segment_anything.py      | 35 ++++++-------
 .../models/encoders/test_sam_image_encoder.py | 51 +++++++------------
 .../visual_prompters/test_segment_anything.py | 16 +++---
 .../visual_prompting/test_helpers.py          |  8 +--
 5 files changed, 46 insertions(+), 87 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py
index f823593d592..d56df51fa5f 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py
@@ -5,7 +5,7 @@
 #
 
 from omegaconf import DictConfig
-from torch import Tensor, nn
+from torch import nn
 
 from otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.backbones import (
     build_tiny_vit,
@@ -19,26 +19,13 @@ class SAMImageEncoder(nn.Module):
     Args:
         config (DictConfig): Config for image encoder.
     """
-
-    def __init__(self, config: DictConfig):
-        super().__init__()
+    
+    def __new__(cls, config: DictConfig):
         if "tiny_vit" == config.backbone:
-            self.backbone = build_tiny_vit(config.image_size)
+            return build_tiny_vit(config.image_size)
         elif "vit" in config.backbone:
-            self.backbone = build_vit(config.backbone, config.image_size)
+            return build_vit(config.backbone, config.image_size)
         else:
             raise NotImplementedError(
                 (f"{config.backbone} for image encoder of SAM is not implemented yet. " f"Use vit_b, l, or h.")
             )
-
-    def forward(self, images: Tensor) -> Tensor:
-        """Forward function of image encoder.
-
-        Args:
-            images (Tensor): Input tensor.
-
-        Returns:
-            image_embeddings (Tensor): Output tensor.
-        """
-        image_embeddings = self.backbone(images)
-        return image_embeddings
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index 9581d21ab41..3e8f271394d 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -133,32 +133,26 @@ def set_metrics(self) -> None:
             )
         )
 
-    def load_checkpoint(
-        self,
-        state_dict: Optional[OrderedDict] = None,
-        revise_keys: List = [(r"^image_encoder.", r"image_encoder.backbone.")],
-    ) -> None:
+    def load_checkpoint(self, state_dict: Optional[OrderedDict] = None) -> None:
         """Load checkpoint for SAM.
 
         Args:
             state_dict (Optional[OrderedDict], optional): State dict of SAM. Defaults to None.
-            revise_keys (List, optional): List of tuples of regex patterns to revise keys of state_dict.
-                Defaults to [(r'^image_encoder.', r'image_encoder.backbone.')].
         """
-
-        def replace_state_dict_keys(state_dict, revise_keys):
-            for p, r in revise_keys:
-                state_dict = OrderedDict(
-                    {
-                        re.sub(p, r, k) if re.search(p, k) and not re.search(r, k) else k: v
-                        for k, v in state_dict.items()
-                    }
-                )
-            return state_dict
-
+        def skip_unused_parameters(state_dict):
+            if self.config.model.backbone == "tiny_vit":
+                for key in [
+                    "image_encoder.norm_head.weight",
+                    "image_encoder.norm_head.bias",
+                    "image_encoder.head.weight",
+                    "image_encoder.head.bias",
+                ]:
+                    if key in state_dict:
+                        state_dict.pop(key)
+        
         if state_dict:
             # state_dict from args.load_from
-            state_dict = replace_state_dict_keys(state_dict, revise_keys)
+            skip_unused_parameters(state_dict)
             self.load_state_dict(state_dict)
         elif self.config.model.checkpoint:
             if str(self.config.model.checkpoint).endswith(".ckpt"):
@@ -172,7 +166,8 @@ def replace_state_dict_keys(state_dict, revise_keys):
                     # load checkpoint from local
                     with open(self.config.model.checkpoint, "rb") as f:
                         state_dict = torch.load(f)
-                state_dict = replace_state_dict_keys(state_dict, revise_keys)
+
+                skip_unused_parameters(state_dict)
                 self.load_state_dict(state_dict, strict=False)
 
     ##########################################################
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py
index e80137685d7..5c2c652ed2e 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py
@@ -22,38 +22,21 @@ def forward(self, *args, **kwargs):
 
 
 class TestSAMImageEncoder:
-    @pytest.fixture(autouse=True)
-    def setup(self, mocker) -> None:
-        self.mocker_backbone = mocker.patch(
-            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.encoders.sam_image_encoder.build_vit",
-            return_value=MockBackbone(),
-        )
-
-        self.base_config = DictConfig(dict(backbone="vit_b", image_size=1024))
-
-    @e2e_pytest_unit
-    @pytest.mark.parametrize("backbone", ["vit_b", "resnet"])
-    def test_init(self, backbone: str):
-        """Test init."""
-        self.mocker_backbone.reset_mock()
-
-        config = self.base_config.copy()
-        config.update(dict(backbone=backbone))
-
-        if backbone == "resnet":
-            with pytest.raises(NotImplementedError):
-                sam_image_encoder = SAMImageEncoder(config)
-        else:
-            sam_image_encoder = SAMImageEncoder(config)
-            self.mocker_backbone.assert_called_once()
-
+    @pytest.fixture()
+    def config(self, mocker) -> DictConfig:
+        return DictConfig(dict(image_size=1024))
+        
     @e2e_pytest_unit
-    def test_forward(self, mocker):
-        """Test forward."""
-        self.mocker_backbone.reset_mock()
-
-        sam_image_encoder = SAMImageEncoder(self.base_config)
-        mocker_forward = mocker.patch.object(sam_image_encoder.backbone, "forward")
-        sam_image_encoder.forward(torch.Tensor([1.0]))
-
-        mocker_forward.assert_called_once()
+    @pytest.mark.parametrize("backbone,expected",
+        [
+            ("tiny_vit", "TinyViT"),
+            ("vit_b", "ViT"),
+        ],
+    )
+    def test_new(self, config: DictConfig, backbone: str, expected: str) -> None:
+        """Test __new__."""
+        config.update({"backbone": backbone})
+
+        sam_image_encoder = SAMImageEncoder(config)
+
+        assert sam_image_encoder.__class__.__name__ == expected
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index fed22e060c8..b8fe9e626bc 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -76,7 +76,7 @@ def test_set_models(self, mocker, backbone: str) -> None:
             # backbone == vit_b
             sam = SegmentAnything(config)
 
-            assert isinstance(sam.image_encoder, MockImageEncoder)
+            assert isinstance(sam.image_encoder, nn.Linear)
             assert isinstance(sam.prompt_encoder, MockPromptEncoder)
             assert isinstance(sam.mask_decoder, MockMaskDecoder)
 
@@ -159,8 +159,8 @@ def test_set_metrics(self, mocker, loss_type: str):
                 False,
                 OrderedDict(
                     [
-                        ("image_encoder.weight", Tensor([[0.0]])),
-                        ("image_encoder.bias", Tensor([0.0])),
+                        ("image_encoder.weight", torch.ones(4, 4)),
+                        ("image_encoder.bias", torch.ones(4)),
                         ("prompt_encoder.layer.weight", Tensor([[0.0]])),
                         ("prompt_encoder.layer.bias", Tensor([0.0])),
                         ("mask_decoder.layer.weight", Tensor([[0.0]])),
@@ -172,8 +172,8 @@ def test_set_metrics(self, mocker, loss_type: str):
                 True,
                 OrderedDict(
                     [
-                        ("image_encoder.backbone.weight", Tensor([[1.0]])),
-                        ("image_encoder.backbone.bias", Tensor([1.0])),
+                        ("image_encoder.weight", torch.ones(4, 4)),
+                        ("image_encoder.bias", torch.ones(4)),
                         ("prompt_encoder.layer.weight", Tensor([[1.0]])),
                         ("prompt_encoder.layer.bias", Tensor([1.0])),
                         ("mask_decoder.layer.weight", Tensor([[1.0]])),
@@ -196,10 +196,8 @@ def test_load_checkpoint_with_state_dict(self, mocker, is_backbone_arg: bool, st
         sam_state_dict = sam.state_dict()
 
         for k, v in state_dict.items():
-            if not is_backbone_arg:
-                k = k.replace("image_encoder", "image_encoder.backbone")
             assert k in sam_state_dict
-            assert v == sam_state_dict[k]
+            assert torch.all(v == sam_state_dict[k])
 
     @e2e_pytest_unit
     def test_load_checkpoint_without_checkpoint(self, mocker):
@@ -360,7 +358,7 @@ def test_mask_postprocessing(self, mocker) -> None:
     def test_forward_train(self) -> None:
         """Test forward."""
         sam = SegmentAnything(config=self.base_config)
-        images = torch.zeros((1))
+        images = torch.zeros((1, 3, 4, 4))
         bboxes = torch.zeros((1))
 
         results = sam.forward_train(images=images, bboxes=bboxes, points=None)
diff --git a/tests/unit/algorithms/visual_prompting/test_helpers.py b/tests/unit/algorithms/visual_prompting/test_helpers.py
index c1be0ae3c89..fb585e22a87 100644
--- a/tests/unit/algorithms/visual_prompting/test_helpers.py
+++ b/tests/unit/algorithms/visual_prompting/test_helpers.py
@@ -148,12 +148,8 @@ def __init__(self, use_mask: bool = False):
 
 
 class MockImageEncoder(nn.Module):
-    def __init__(self, *args, **kwargs):
-        super().__init__()
-        self.backbone = nn.Linear(1, 1)
-
-    def forward(self, *args, **kwargs):
-        return torch.ones((1, 2, 4, 4))
+    def __new__(cls, *args, **kwargs):
+        return nn.Linear(4, 4)
 
 
 class MockPromptEncoder(nn.Module):

From 61d85965628abea114a4d326ce5053c8fca64f0a Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 6 Feb 2024 17:57:44 +0900
Subject: [PATCH 03/28] Enable inner postprocessing

---
 .../visual_prompters/segment_anything.py      | 86 ++++++-------------
 .../zero_shot_segment_anything.py             | 26 +++---
 .../visual_prompting/tasks/inference.py       |  6 +-
 .../visual_prompting/tasks/openvino.py        |  2 +-
 .../compressed_decoder.yml                    |  2 +-
 .../visual_prompters/test_segment_anything.py | 38 ++++----
 .../test_zero_shot_segment_anything.py        | 34 ++++----
 .../visual_prompting/tasks/test_inference.py  |  3 +-
 .../visual_prompting/tasks/test_openvino.py   |  2 +-
 9 files changed, 88 insertions(+), 111 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index 3e8f271394d..f6b59922baf 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -181,7 +181,7 @@ def forward(
         point_labels: Tensor,
         mask_input: Tensor,
         has_mask_input: Tensor,
-        # orig_size: Tensor,
+        orig_size: Tensor,
     ):
         """Forward method for SAM inference (export/deploy).
 
@@ -223,18 +223,16 @@ def forward(
         if self.config.model.return_single_mask:
             masks, scores = self.select_masks(masks, scores, point_coords.shape[1])
 
-        return scores, masks
-        # TODO (sungchul): apply inner postprocessing
-        # upscaled_masks = self.mask_postprocessing(masks, orig_size[0])
+        upscaled_masks = self.postprocess_masks(masks, self.config.model.image_size, orig_size[0])
 
-        # if self.config.model.return_extra_metrics:
-        #     stability_scores = self.calculate_stability_score(
-        #         upscaled_masks, self.config.model.mask_threshold, self.config.model.stability_score_offset
-        #     )
-        #     areas = (upscaled_masks > self.config.model.mask_threshold).sum(-1).sum(-1)
-        #     return upscaled_masks, scores, stability_scores, areas, masks
+        if self.config.model.return_extra_metrics:
+            stability_scores = self.calculate_stability_score(
+                upscaled_masks, self.config.model.mask_threshold, self.config.model.stability_score_offset
+            )
+            areas = (upscaled_masks > self.config.model.mask_threshold).sum(-1).sum(-1)
+            return upscaled_masks, scores, stability_scores, areas, masks
 
-        # return upscaled_masks, scores, masks
+        return upscaled_masks, scores, masks
 
     def _embed_points(self, point_coords: Tensor, point_labels: Tensor) -> Tensor:
         """Embed sparse input prompts.
@@ -328,10 +326,10 @@ def select_masks(self, masks: Tensor, iou_preds: Tensor, num_points: int) -> Tup
         iou_preds = iou_preds[torch.arange(masks.shape[0]), best_idx].unsqueeze(1)
 
         return masks, iou_preds
-
-    @staticmethod
-    def mask_postprocessing(masks: Tensor, input_size: int, orig_size: Tensor) -> Tensor:
-        """Postprocesses the predicted masks.
+    
+    @classmethod
+    def postprocess_masks(cls, masks: Tensor, input_size: int, orig_size: Tensor) -> Tensor:
+        """Postprocess the predicted masks.
 
         Args:
             masks (Tensor): A batch of predicted masks with shape Bx1xHxW.
@@ -342,22 +340,20 @@ def mask_postprocessing(masks: Tensor, input_size: int, orig_size: Tensor) -> Te
         Returns:
             masks (Tensor): The postprocessed masks with shape Bx1xHxW.
         """
-
-        def resize_longest_image_size(input_image_size: Tensor, longest_side: int) -> Tensor:
-            scale = longest_side / torch.max(input_image_size)
-            transformed_size = scale * input_image_size
-            transformed_size = torch.floor(transformed_size + 0.5).to(torch.int64)
-            return transformed_size
-
         masks = F.interpolate(masks, size=(input_size, input_size), mode="bilinear", align_corners=False)
 
-        prepadded_size = resize_longest_image_size(orig_size, input_size)
-        masks = masks[..., : prepadded_size[0], : prepadded_size[1]]  # type: ignore
+        prepadded_size = cls.get_prepadded_size(cls, orig_size, input_size)
+        masks = masks[..., : prepadded_size[0], : prepadded_size[1]]
 
         orig_size = orig_size.to(torch.int64)
         h, w = orig_size[0], orig_size[1]
-        masks = F.interpolate(masks, size=(h, w), mode="bilinear", align_corners=False)
-        return masks
+        return F.interpolate(masks, size=(h, w), mode="bilinear", align_corners=False)
+
+    def get_prepadded_size(self, input_image_size: Tensor, longest_side: int) -> Tensor:
+        """Get pre-padded size."""
+        scale = longest_side / torch.max(input_image_size)
+        transformed_size = scale * input_image_size
+        return torch.floor(transformed_size + 0.5).to(torch.int64)
 
     ######################################################
     #     forward for training/validation/prediction     #
@@ -437,9 +433,9 @@ def training_step(self, batch, batch_idx) -> Tensor:
         num_masks = sum(len(pred_mask) for pred_mask in pred_masks)
         for i, (pred_mask, gt_mask, iou_prediction) in enumerate(zip(pred_masks, gt_masks, iou_predictions)):
             pred_mask = self.postprocess_masks(
-                pred_mask, images.shape[2:], batch["padding"][i], batch["original_size"][i]
+                pred_mask, self.config.model.image_size, batch["original_size"][i]
             )
-            pred_mask = pred_mask.sigmoid()
+            pred_mask = pred_mask.sigmoid().squeeze(1)
             self.train_metrics["train_IoU"].update(pred_mask, gt_mask)
             self.train_metrics["train_F1"].update(pred_mask, gt_mask)
             self.train_metrics["train_Dice"].update(pred_mask, gt_mask)
@@ -495,9 +491,9 @@ def validation_step(self, batch, batch_idx) -> MetricCollection:
         pred_masks, _ = self.forward_train(images, bboxes, points)
         for i, (pred_mask, gt_mask) in enumerate(zip(pred_masks, gt_masks)):
             pred_mask = self.postprocess_masks(
-                pred_mask, images.shape[2:], batch["padding"][i], batch["original_size"][i]
+                pred_mask, self.config.model.image_size, batch["original_size"][i]
             )
-            pred_mask = pred_mask.sigmoid()
+            pred_mask = pred_mask.sigmoid().squeeze(1)
             for k, v in self.val_metrics.items():
                 v.update(pred_mask, gt_mask)
 
@@ -527,41 +523,15 @@ def predict_step(self, batch, batch_idx) -> Dict[str, Tensor]:
 
         masks: List[Tensor] = []
         for i, pred_mask in enumerate(pred_masks):
-            mask = self.postprocess_masks(pred_mask, images.shape[2:], batch["padding"][i], batch["original_size"][i])
+            mask = self.postprocess_masks(pred_mask, self.config.model.image_size, batch["original_size"][i])
             if not self.config.model.return_logits:
                 mask = (mask > self.config.model.mask_threshold).to(mask.dtype)
             else:
                 mask = mask.sigmoid()
-            masks.append(mask)
+            masks.append(mask.squeeze(1))
 
         return dict(masks=masks, iou_predictions=iou_predictions, path=batch["path"], labels=batch["labels"])
 
-    @staticmethod
-    def postprocess_masks(
-        masks: Tensor,
-        input_size: Tuple[int, int],
-        padding: Union[Tuple[int, ...], Tensor],
-        original_size: Union[Tuple[int, int], Tensor],
-    ) -> Tensor:
-        """Remove padding and upscale masks to the original image size.
-
-        Args:
-            masks (Tensor): Predicted masks from the mask_decoder with (N, 1, H/downsized_ratio, W/downsized_ratio).
-            input_size (tuple(int, int)): The size of the image input to the model, in (H, W) format.
-                Used to remove padding.
-            padding (tuple(int, int, int, int), Tensor): The padding applied to the image before input to the model,
-                in (left, top, right, bottom) format.
-            original_size (tuple(int, int), Tensor): The original size of the image before resizing
-                for input to the model, in (H, W) format.
-
-        Returns:
-          (Tensor): Postprocessed masks in NxHxW format, where (H, W) is given by original_size.
-        """
-        masks = F.interpolate(masks, input_size, mode="bilinear", align_corners=False)
-        masks = masks[..., : input_size[0] - padding[3], : input_size[1] - padding[2]]
-        masks = F.interpolate(masks, [int(o) for o in original_size], mode="bilinear", align_corners=False)
-        return masks.squeeze(1)
-
     def configure_optimizers(self) -> optim:
         """Configure the optimizer for SAM.
 
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 8321714d577..4fceb645b39 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -132,7 +132,7 @@ def get_prompt_candidates(
 
         sim = self.reference_feats[label].to(device) @ target_feat
         sim = sim.reshape(1, 1, h_feat, w_feat)
-        sim = ZeroShotSegmentAnything.mask_postprocessing(sim, self.image_size, original_size[0])
+        sim = ZeroShotSegmentAnything.postprocess_masks(sim, self.image_size, original_size[0])
 
         threshold = (threshold == 0) * self.default_threshold_target + threshold
         points_scores, bg_coords = self._point_selection(
@@ -321,7 +321,7 @@ def learn(
                         image_embeddings=image_embeddings,
                         point_coords=point_coords,
                         point_labels=point_labels,
-                        original_size=original_size,
+                        original_size=original_size.unsqueeze(0),
                         is_cascade=False,
                     )
                     reference_prompt[masks] += 1
@@ -396,7 +396,7 @@ def infer(
                         image_embeddings=image_embeddings,
                         point_coords=point_coords,
                         point_labels=point_labels,
-                        original_size=original_size[0],
+                        original_size=original_size,
                     )
                     predicted_masks[label].append((mask * points_score[2]).detach().cpu())
                     used_points[label].append(points_score.detach().cpu())
@@ -463,7 +463,7 @@ def _predict_masks(
 
             elif is_cascade and i == 1:
                 # Cascaded Post-refinement-1
-                mask_input, masks = self._postprocess_masks(logits, scores, original_size, is_single=True)  # noqa: F821
+                mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)  # noqa: F821
                 if masks.sum() == 0:
                     return masks
 
@@ -471,7 +471,7 @@ def _predict_masks(
 
             elif is_cascade and i == 2:
                 # Cascaded Post-refinement-2
-                mask_input, masks = self._postprocess_masks(logits, scores, original_size)  # noqa: F821
+                mask_input, masks = self._postprocess_masks(masks, logits, scores)  # noqa: F821
                 if masks.sum() == 0:
                     return masks
 
@@ -480,21 +480,22 @@ def _predict_masks(
                 y, x = coords[:, 0], coords[:, 1]
                 box_coords = ResizeLongestSide.apply_coords(
                     torch.tensor([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device),
-                    original_size,
+                    original_size[0],
                     self.config.model.image_size,
                 )
                 point_coords = torch.cat((point_coords, box_coords), dim=1)
                 point_labels = torch.cat((point_labels, self.point_labels_box.to(self.device)), dim=1)
 
-            scores, logits = self(
+            high_res_masks, scores, logits = self(
                 image_embeddings=image_embeddings,
                 point_coords=point_coords,
                 point_labels=point_labels,
                 mask_input=mask_input,
                 has_mask_input=has_mask_input,
+                orig_size=original_size,
             )
-
-        _, masks = self._postprocess_masks(logits, scores, original_size)
+            masks = high_res_masks > self.config.model.mask_threshold
+        _, masks = self._postprocess_masks(masks, logits, scores)
         return masks
 
     def training_step(self, batch, batch_idx) -> None:
@@ -619,15 +620,12 @@ def _preprocess_masks(self, x: torch.Tensor) -> torch.Tensor:
 
     def _postprocess_masks(
         self,
+        masks: torch.Tensor,
         logits: torch.Tensor,
         scores: torch.Tensor,
-        original_size: torch.Tensor,
         is_single: bool = False,
     ):
         """Post-process masks for cascaded post-refinements."""
-        high_res_masks = self.mask_postprocessing(logits, self.config.model.image_size, original_size)
-        masks = high_res_masks > self.config.model.mask_threshold
-
         if is_single:
             best_idx = 0
         else:
@@ -642,7 +640,7 @@ def _postprocess_masks(
 
             if len(scores[0]) == 0:
                 # all predicted masks were zero masks, ignore them.
-                return None, torch.zeros((self.config.model.image_size, self.config.model.image_size), device="cpu")
+                return None, torch.zeros(masks.shape[-2:], device="cpu")
 
             best_idx = torch.argmax(scores[0])
         return logits[:, best_idx], masks[0, best_idx]
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index 1123cd20c87..725f34aae25 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -305,8 +305,9 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "point_labels": torch.randint(low=0, high=4, size=(1, 2), dtype=torch.float32),
                     "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float32),
                     "has_mask_input": torch.tensor([[1]], dtype=torch.float32),
+                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.float)
                 }
-                output_names = ["iou_predictions", "low_res_masks"]
+                output_names = ["upscaled_masks", "iou_predictions", "low_res_masks"]
                 model_to_export = self.model
 
             with warnings.catch_warnings():
@@ -666,8 +667,9 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "point_labels": torch.randint(low=0, high=4, size=(1, 2), dtype=torch.float32),
                     "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float32),
                     "has_mask_input": torch.tensor([[1]], dtype=torch.float32),
+                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.float)
                 }
-                output_names = ["iou_predictions", "low_res_masks"]
+                output_names = ["upscaled_masks", "iou_predictions", "low_res_masks"]
                 model_to_export = self.model
 
             else:
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 60cc3b25331..02d0c9aa4ce 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -526,7 +526,6 @@ def __getitem__(self, index: int):
             image_embeddings = self.image_encoder(images["images"])
             prompt = prompts[0]  # only use the first prompt
             prompt.pop("label")
-            prompt.pop("orig_size")
             prompt.update({"image_embeddings": image_embeddings["image_embeddings"]})
             return prompt
             # TODO (sungchul): change has_mask_input
@@ -587,6 +586,7 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
             inputs_decoder.update(image_embeddings)
             inputs_decoder.update(
                 {
+                    "orig_size": original_size[None],
                     "mask_input": np.zeros((1, 1, 256, 256), dtype=np.float32),
                     "has_mask_input": np.zeros((1, 1), dtype=np.float32),
                 }
diff --git a/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_decoder.yml b/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_decoder.yml
index 9009e81d953..bbefedd68ef 100644
--- a/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_decoder.yml
+++ b/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_decoder.yml
@@ -1,3 +1,3 @@
 TestToolsZeroShotVisualPrompting:
   ptq:
-    number_of_fakequantizers: 69
+    number_of_fakequantizers: 71
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index b8fe9e626bc..f5dce1d8396 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -344,16 +344,6 @@ def test_select_masks(self) -> None:
         assert masks[:, -1, :, :] == selected_mask
         assert iou_preds[:, -1] == selected_iou_pred
 
-    @e2e_pytest_unit
-    def test_mask_postprocessing(self, mocker) -> None:
-        """Test mask_postprocessing."""
-        masks = torch.empty(1, 1, 2, 2)
-        orig_size = Tensor((8, 8))
-
-        results = SegmentAnything.mask_postprocessing(masks, 6, orig_size)
-
-        assert results[0, 0].shape == tuple(orig_size)
-
     @e2e_pytest_unit
     def test_forward_train(self) -> None:
         """Test forward."""
@@ -494,22 +484,40 @@ def test_predict_step(self, mocker, return_logits: bool, expected: Tensor) -> No
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
-        "input_size,original_size,padding,expected",
+        "input_size,original_size,expected",
         [
-            ((6, 6), (8, 8), (0, 0, 0, 0), (8, 8)),
-            ((6, 6), (8, 8), (0, 0, 2, 2), (8, 8)),
+            (6, torch.tensor((8, 8)), (1, 8, 8)),
+            (6, torch.tensor((8, 8)), (1, 8, 8)),
         ],
     )
     def test_postprocess_masks(
-        self, input_size: Tuple[int], original_size: Tuple[int], padding: Tuple[int], expected: Tuple[int]
+        self, input_size: int, original_size: Tuple[int], expected: Tuple[int]
     ) -> None:
         """Test postprocess_masks."""
         sam = SegmentAnything(config=self.base_config)
         masks = torch.zeros((1, 1, 4, 4))
 
-        results = sam.postprocess_masks(masks, input_size, padding, original_size)
+        results = sam.postprocess_masks(masks, input_size, original_size)
 
         assert results.shape[1:] == expected
+        
+    @e2e_pytest_unit
+    @pytest.mark.parametrize(
+        "input_image_size,expected",
+        [
+            (torch.tensor((2, 4)), torch.tensor((3, 6))),
+            (torch.tensor((4, 2)), torch.tensor((6, 3))),
+        ],
+    )
+    def test_get_prepadded_size(self, input_image_size: Tensor, expected: Tensor) -> None:
+        """Test get_prepadded_size."""
+        sam = SegmentAnything(config=self.base_config)
+
+        longest_side = 6
+
+        results = sam.get_prepadded_size(input_image_size, longest_side)
+
+        assert torch.all(results == expected)
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index a03fa4b838a..3bdb01d79b7 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -201,17 +201,17 @@ def test_learn(self, mocker, set_zero_shot_segment_anything) -> None:
 
         processed_prompts = {MockScoredLabel(label=1, name="label"): [{"box": torch.tensor([[0, 0, 1, 1]])}]}
         zero_shot_segment_anything.learn(
-            images=torch.ones((1, 3, 8, 8)),
+            images=torch.ones((1, 3, 4, 4)),
             processed_prompts=processed_prompts,
             padding=(0, 0, 0, 0),
-            original_size=(8, 8),
+            original_size=torch.tensor((4, 4)),
         )
 
-        assert zero_shot_segment_anything.prompt_getter.reference_feats.shape == (2, 1, 2)
-        assert zero_shot_segment_anything.prompt_getter.reference_prompts.shape == (2, 8, 8)
+        assert zero_shot_segment_anything.prompt_getter.reference_feats.shape == (2, 1, 3)
+        assert zero_shot_segment_anything.prompt_getter.reference_prompts.shape == (2, 4, 4)
 
     @e2e_pytest_unit
-    @pytest.mark.parametrize("expected", [[torch.ones((8, 8)) / 2, torch.tensor([0.0, 0.0, 0.5])]])
+    @pytest.mark.parametrize("expected", [[torch.ones((4, 4)) / 2, torch.tensor([0.0, 0.0, 0.5])]])
     def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expected: torch.Tensor) -> None:
         """Test infer."""
         monkeypatch.setattr(
@@ -221,13 +221,13 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
 
         zero_shot_segment_anything = set_zero_shot_segment_anything()
         zero_shot_segment_anything.prompt_getter.reference_feats = torch.rand(1, 1, 4)
-        zero_shot_segment_anything.prompt_getter.reference_prompts = torch.zeros((8, 8))
+        zero_shot_segment_anything.prompt_getter.reference_prompts = torch.zeros((4, 4))
         mocker.patch.object(
-            SegmentAnything, "forward", return_value=(torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
+            SegmentAnything, "forward", return_value=(torch.ones(1, 4, 4, 4), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
         )
 
         total_results = zero_shot_segment_anything.infer(
-            images=torch.ones((1, 3, 8, 8)), original_size=torch.tensor([[8, 8]], dtype=torch.int64)
+            images=torch.ones((1, 3, 4, 4)), original_size=torch.tensor([[4, 4]], dtype=torch.int64)
         )
 
         for i, results in enumerate(total_results[0]):
@@ -235,11 +235,10 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
                 assert torch.equal(result[0], expected[i])
 
     @e2e_pytest_unit
-    @pytest.mark.parametrize("is_postprocess", [True, False])
-    def test_predict_masks(self, mocker, set_zero_shot_segment_anything, is_postprocess: bool) -> None:
+    def test_predict_masks(self, mocker, set_zero_shot_segment_anything) -> None:
         """Test _predict_masks."""
         mocker.patch.object(
-            SegmentAnything, "forward", return_value=(torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
+            SegmentAnything, "forward", return_value=(torch.ones(1, 4, 8, 8), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
         )
 
         zero_shot_segment_anything = set_zero_shot_segment_anything()
@@ -249,7 +248,7 @@ def test_predict_masks(self, mocker, set_zero_shot_segment_anything, is_postproc
             image_embeddings=torch.rand(1),
             point_coords=torch.rand(1, 2, 2),
             point_labels=torch.randint(low=0, high=2, size=(1, 2)),
-            original_size=torch.tensor((8, 8), dtype=torch.int64),
+            original_size=torch.tensor([[8, 8]], dtype=torch.int64),
         )
         assert mask.shape == (8, 8)
 
@@ -300,22 +299,21 @@ def test_preprocess_masks(self, set_zero_shot_segment_anything) -> None:
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
-        "logits,expected",
+        "masks,logits,expected",
         [
-            (torch.ones(1, 4, 4, 4), torch.ones(4, 4, dtype=torch.bool)),
-            (torch.zeros(1, 4, 4, 4), torch.zeros(4, 4, dtype=torch.bool)),
+            (torch.ones(1, 4, 8, 8), torch.ones(1, 4, 4, 4), torch.ones(8, 8)),
+            (torch.zeros(1, 4, 8, 8), torch.zeros(1, 4, 4, 4), torch.zeros(8, 8)),
         ],
     )
     def test_postprocess_masks(
-        self, set_zero_shot_segment_anything, logits: torch.Tensor, expected: torch.Tensor
+        self, set_zero_shot_segment_anything, masks: torch.Tensor, logits: torch.Tensor, expected: torch.Tensor
     ) -> None:
         """Test _postprocess_masks."""
         zero_shot_segment_anything = set_zero_shot_segment_anything()
         zero_shot_segment_anything.config.model.image_size = 4
         scores = torch.tensor([[0.0, 0.1, 0.2, 0.3]])
-        original_size = torch.tensor([4, 4], dtype=torch.int64)
 
-        _, result = zero_shot_segment_anything._postprocess_masks(logits, scores, original_size)
+        _, result = zero_shot_segment_anything._postprocess_masks(masks, logits, scores)
 
         assert torch.equal(result, expected)
 
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index acd9d0c48ca..8c717a45421 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -353,12 +353,13 @@ def test_export_to_onnx(self):
                 "point_labels": np.random.randint(low=0, high=4, size=(1, 2)).astype(np.float32),
                 "mask_input": np.random.randn(1, 1, *mask_input_size).astype(np.float32),
                 "has_mask_input": np.array([[1]], dtype=np.float32),
+                "orig_size": np.random.randint(low=256, high=2048, size=(1, 2)).astype(np.float32),
             },
         }
         onnx_outputs = {
             "visual_prompting_image_encoder": ["image_embeddings"],
             "visual_prompting_prompt_getter": ["total_points_scores", "total_bg_coords"],
-            "visual_prompting_decoder": ["iou_predictions", "low_res_masks"],
+            "visual_prompting_decoder": ["upscaled_masks", "iou_predictions", "low_res_masks"],
         }
 
         onnx_rt_models = {
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 4181a8b2bab..a7af582cd63 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -365,7 +365,7 @@ def test_getitem(self, mocker, load_dataloader, module_name: str):
             self.mocker_read_model.assert_called_once()
             self.mocker_compile_model.assert_called_once()
             assert "label" not in results
-            assert "orig_size" not in results
+            assert "orig_size" in results
             assert "image_embeddings" in results
 
 

From b4d66b2ddd16ee5c2ca5eddcba4049e6ca3eea3f Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 8 Feb 2024 15:10:09 +0900
Subject: [PATCH 04/28] Update OV inferencer (TODO: zsl)

---
 .../model_wrappers/openvino_models.py         | 45 +------------------
 .../visual_prompting/tasks/inference.py       |  4 +-
 .../visual_prompting/tasks/openvino.py        | 35 +++++++--------
 .../model_wrappers/test_openvino_models.py    | 26 +----------
 .../test_zero_shot_segment_anything.py        | 17 +++----
 .../visual_prompting/tasks/test_inference.py  |  2 +-
 .../visual_prompting/tasks/test_openvino.py   | 35 +++++++--------
 7 files changed, 46 insertions(+), 118 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 1bdc1a473a3..432d0010e09 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -94,7 +94,7 @@ def parameters(cls):  # noqa: D102
         return parameters
 
     def _get_outputs(self):
-        return "low_res_masks"
+        return "upscaled_masks"
 
     def preprocess(self, inputs: Dict[str, Any], meta: Dict[str, Any]) -> List[Dict[str, Any]]:
         """Preprocess prompts."""
@@ -111,7 +111,7 @@ def preprocess(self, inputs: Dict[str, Any], meta: Dict[str, Any]) -> List[Dict[
                     # TODO (sungchul): how to generate mask_input and has_mask_input
                     "mask_input": np.zeros((1, 1, 256, 256), dtype=np.float32),
                     "has_mask_input": np.zeros((1, 1), dtype=np.float32),
-                    "orig_size": np.array(inputs["original_size"], dtype=np.float32).reshape((-1, 2)),
+                    "orig_size": np.array(inputs["original_size"], dtype=np.int64).reshape((-1, 2)),
                     "label": label,
                 }
             )
@@ -159,7 +159,6 @@ def sigmoid(x):
             return np.tanh(x * 0.5) * 0.5 + 0.5  # to avoid overflow
 
         soft_prediction = outputs[self.output_blob_name].squeeze()
-        soft_prediction = self.resize_and_crop(soft_prediction, meta["original_size"][0])
         soft_prediction = sigmoid(soft_prediction)
         meta["soft_prediction"] = soft_prediction
 
@@ -173,43 +172,3 @@ def sigmoid(x):
         meta["label"].probability = probability
 
         return hard_prediction, soft_prediction
-
-    def resize_and_crop(self, soft_prediction: np.ndarray, original_size: np.ndarray) -> np.ndarray:
-        """Resize and crop soft prediction.
-
-        Args:
-            soft_prediction (np.ndarray): Predicted soft prediction with HxW shape.
-            original_size (np.ndarray): The original image size.
-
-        Returns:
-            final_soft_prediction (np.ndarray): Resized and cropped soft prediction for the original image.
-        """
-        resized_soft_prediction = cv2.resize(
-            soft_prediction, (self.image_size, self.image_size), 0, 0, interpolation=cv2.INTER_LINEAR
-        )
-
-        prepadded_size = self.get_padded_size(original_size, self.image_size).astype(np.int64)
-        resized_cropped_soft_prediction = resized_soft_prediction[: prepadded_size[0], : prepadded_size[1], ...]
-
-        original_size = original_size.astype(np.int64)
-        h, w = original_size
-        final_soft_prediction = cv2.resize(
-            resized_cropped_soft_prediction, (w, h), 0, 0, interpolation=cv2.INTER_LINEAR
-        )
-        return final_soft_prediction
-
-    def get_padded_size(self, original_size: np.ndarray, longest_side: int) -> np.ndarray:
-        """Get padded size from original size and longest side of the image.
-
-        Args:
-            original_size (np.ndarray): The original image size with shape Bx2.
-            longest_side (int): The size of the longest side.
-
-        Returns:
-            transformed_size (np.ndarray): The transformed image size with shape Bx2.
-        """
-        original_size = original_size.astype(np.float32)
-        scale = longest_side / np.max(original_size)
-        transformed_size = scale * original_size
-        transformed_size = np.floor(transformed_size + 0.5).astype(np.int64)
-        return transformed_size
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index 725f34aae25..8bff9bd55f7 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -305,7 +305,7 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "point_labels": torch.randint(low=0, high=4, size=(1, 2), dtype=torch.float32),
                     "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float32),
                     "has_mask_input": torch.tensor([[1]], dtype=torch.float32),
-                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.float)
+                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.int64)
                 }
                 output_names = ["upscaled_masks", "iou_predictions", "low_res_masks"]
                 model_to_export = self.model
@@ -667,7 +667,7 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "point_labels": torch.randint(low=0, high=4, size=(1, 2), dtype=torch.float32),
                     "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float32),
                     "has_mask_input": torch.tensor([[1]], dtype=torch.float32),
-                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.float)
+                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.int64)
                 }
                 output_names = ["upscaled_masks", "iou_predictions", "low_res_masks"]
                 model_to_export = self.model
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 02d0c9aa4ce..7751713aa47 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -176,12 +176,11 @@ def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type:
         soft_predictions: List[np.ndarray] = []
         for prompt in prompts:
             label = prompt.pop("label")
-            orig_size = prompt.pop("orig_size")
             prompt.update(image_embeddings)
 
             # forward decoder to get predicted mask
             prediction = self.forward_decoder(prompt)
-            metadata = {"label": label, "original_size": orig_size}
+            metadata = {"label": label}
 
             # set annotation for eval
             annotation, hard_prediction, soft_prediction = self.post_process(prediction, metadata)
@@ -378,42 +377,40 @@ def forward_decoder(  # type: ignore
 
             elif i == 1:
                 # Cascaded Post-refinement-1
-                mask_input, masks, iou_predictions = self._postprocess_masks(
-                    logits, scores, original_size, is_single=True  # noqa: F821
-                )
+                mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)
                 if masks.sum() == 0:
-                    return {"iou_predictions": iou_predictions, "low_res_masks": mask_input}
+                    return {"masks": masks}
 
                 has_mask_input = self.has_mask_inputs[1]
 
             elif i == 2:
                 # Cascaded Post-refinement-2
-                mask_input, masks, iou_predictions = self._postprocess_masks(
-                    logits, scores, original_size  # noqa: F821
-                )
+                mask_input, masks = self._postprocess_masks(masks, logits, scores)
                 if masks.sum() == 0:
-                    return {"iou_predictions": iou_predictions, "low_res_masks": mask_input}
+                    return {"masks": masks}
 
                 has_mask_input = self.has_mask_inputs[1]
                 y, x = np.nonzero(masks)
                 box_coords = self.model["decoder"]._apply_coords(
                     np.array([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=np.float32), original_size
                 )
-                inputs["point_coords"] = np.concatenate((inputs["point_coords"], box_coords), axis=1)
-                inputs["point_labels"] = np.concatenate((inputs["point_labels"], self.point_labels_box), axis=1)
+                inputs.update({
+                    "point_coords": np.concatenate((inputs["point_coords"], box_coords), axis=1),
+                    "point_labels": np.concatenate((inputs["point_labels"], self.point_labels_box), axis=1),
+                })
 
             inputs.update({"mask_input": mask_input, "has_mask_input": has_mask_input})
             prediction = self.model["decoder"].infer_sync(inputs)
-            scores, logits = prediction["iou_predictions"], prediction["low_res_masks"]
+            upscaled_masks, scores, logits = prediction["upscaled_masks"], prediction["iou_predictions"], prediction["low_res_masks"]
+            masks = upscaled_masks > self.model["decoder"].mask_threshold
 
-        return {"iou_predictions": scores[:, mask_slice], "low_res_masks": logits[:, mask_slice, :, :]}
+        _, masks = self._postprocess_masks(masks, logits, scores)
+        return {"masks": masks}
 
     def _postprocess_masks(
-        self, logits: np.ndarray, scores: np.ndarray, original_size: np.ndarray, is_single: bool = False
+        self, masks: np.ndarray, logits: np.ndarray, scores: np.ndarray, is_single: bool = False
     ) -> Tuple[np.ndarray, ...]:
         """Post-process logits for resized masks according to best index based on scores."""
-        high_res_masks = self.model["decoder"].resize_and_crop(logits[0].transpose(1, 2, 0), original_size)
-        masks = high_res_masks > self.model["decoder"].mask_threshold
         masks = masks.transpose(2, 0, 1)[None]
 
         if is_single:
@@ -430,10 +427,10 @@ def _postprocess_masks(
 
             if len(scores[0]) == 0:
                 # all predicted masks were zero masks, ignore them.
-                return None, np.zeros((self.model["decoder"].image_size, self.model["decoder"].image_size)), 0.0
+                return None, np.zeros(masks.shape[-2:])
 
             best_idx = np.argmax(scores[0])
-        return logits[:, [best_idx]], masks[0, best_idx], scores[0, best_idx]
+        return logits[:, [best_idx]], masks[0, best_idx]
 
     def __inspect_overlapping_areas(
         self,
diff --git a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
index 7740de14ab9..9b5093b004a 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
@@ -79,7 +79,7 @@ def test_get_outputs(self):
         """Test _get_outputs."""
         results = self.decoder._get_outputs()
 
-        assert "low_res_masks" == results
+        assert "upscaled_masks" == results
 
     @e2e_pytest_unit
     def test_preprocess(self):
@@ -142,27 +142,3 @@ def test_postprocess(self, mocker):
         returned_value = self.decoder.postprocess(outputs=fake_output, meta=fake_metadata)
 
         assert isinstance(returned_value, tuple)
-        assert np.all(returned_value[0].shape == fake_metadata["original_size"])
-        assert np.all(returned_value[1].shape == fake_metadata["original_size"])
-
-    @e2e_pytest_unit
-    def test_resize_and_crop(self, mocker):
-        """Test resize_and_crop."""
-        mocker.patch.object(self.decoder, "get_padded_size", return_value=np.array((6, 6)))
-
-        masks = np.zeros((2, 2))
-        orig_size = np.array((8, 8))
-
-        results = self.decoder.resize_and_crop(masks, orig_size)
-
-        assert results.shape == tuple(orig_size)
-
-    @e2e_pytest_unit
-    def test_get_padded_size(self):
-        """Test get_padded_size."""
-        original_size = np.array((2, 4))
-        longest_side = 6
-
-        results = self.decoder.get_padded_size(original_size, longest_side)
-
-        assert np.all(results == np.array((3, 6)))
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 3bdb01d79b7..8d145bfcaaa 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -74,13 +74,13 @@ def test_set_reference(self) -> None:
         assert self.prompt_getter.reference_prompts[3].sum() == 9
 
     @e2e_pytest_unit
-    def test_forward(self, mocker) -> None:
+    @pytest.mark.parametrize("result_point_selection", [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])])
+    def test_forward(self, mocker, result_point_selection: torch.Tensor) -> None:
         """Test forward."""
         mocker.patch.object(
             self.prompt_getter,
             "get_prompt_candidates",
-            return_value=(torch.tensor([[[0, 0, 0.5], [1, 1, 0.7]]]), torch.tensor([[[2, 2]]])),
-        )
+            return_value=(result_point_selection, torch.zeros(1, 2)))
         image_embeddings = torch.ones(1, 4, 4, 4)
         self.prompt_getter.reference_feats = torch.rand(1, 1, 4)
         original_size = torch.tensor((self.prompt_getter.image_size, self.prompt_getter.image_size), dtype=torch.int64)
@@ -88,17 +88,18 @@ def test_forward(self, mocker) -> None:
         total_points_scores, total_bg_coords = self.prompt_getter(
             image_embeddings=image_embeddings, original_size=original_size
         )
-
+        
         assert total_points_scores.shape[0] == 1
         assert total_bg_coords.shape[0] == 1
 
     @e2e_pytest_unit
-    def test_get_prompt_candidates(self, mocker) -> None:
+    @pytest.mark.parametrize("result_point_selection", [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])])
+    def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tensor) -> None:
         """Test get_prompt_candidates."""
         mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
         )
-        mocker.patch.object(self.prompt_getter, "_point_selection", return_value=("points_scores", "bg_coords"))
+        mocker.patch.object(self.prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2)))
         image_embeddings = torch.ones(1, 4, 4, 4)
         self.prompt_getter.reference_feats = torch.rand(1, 1, 4)
         label = torch.tensor([[0]], dtype=torch.int64)
@@ -110,8 +111,8 @@ def test_get_prompt_candidates(self, mocker) -> None:
             image_embeddings=image_embeddings, label=label, original_size=original_size
         )
 
-        assert points_scores == "points_scores"
-        assert bg_coords == "bg_coords"
+        assert torch.all(points_scores == result_point_selection)
+        assert torch.all(bg_coords == torch.zeros(1, 2))
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index 8c717a45421..051ddcdfe8f 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -353,7 +353,7 @@ def test_export_to_onnx(self):
                 "point_labels": np.random.randint(low=0, high=4, size=(1, 2)).astype(np.float32),
                 "mask_input": np.random.randn(1, 1, *mask_input_size).astype(np.float32),
                 "has_mask_input": np.array([[1]], dtype=np.float32),
-                "orig_size": np.random.randint(low=256, high=2048, size=(1, 2)).astype(np.float32),
+                "orig_size": np.random.randint(low=256, high=2048, size=(1, 2)).astype(np.int64),
             },
         }
         onnx_outputs = {
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index a7af582cd63..459aba889f5 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -202,6 +202,7 @@ def setup(self, mocker):
         self.visual_prompting_ov_inferencer.model["decoder"] = mocker.patch(
             "otx.algorithms.visual_prompting.tasks.openvino.model_wrappers.Decoder", autospec=True
         )
+        self.visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.3
         self.visual_prompting_ov_inferencer.model["decoder"]._apply_coords.return_value = np.array([[1, 1]])
 
     @e2e_pytest_unit
@@ -243,14 +244,14 @@ def test_predict(self, mocker):
         "postprocess_output,infer_sync_output,expected",
         [
             (
-                (np.ones((1, 1)), np.ones((3, 3)), 0.9),
-                {"iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
-                {"iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
+                (np.ones((1, 1)), np.ones((3, 3))),
+                {"upscaled_masks": np.ones((3, 3)), "iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
+                {"masks": np.ones((3, 3))},
             ),
             (
-                (np.zeros((2, 2)), np.zeros((3, 3)), 0.0),
-                {"iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
-                {"iou_predictions": 0.0, "low_res_masks": np.zeros((2, 2))},
+                (np.zeros((2, 2)), np.zeros((3, 3))),
+                {"upscaled_masks": np.zeros((3, 3)), "iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
+                {"masks": np.zeros((3, 3))},
             ),
         ],
     )
@@ -281,17 +282,15 @@ def test_forward_decoder(
             original_size=np.array([3, 3]),
         )
 
-        assert np.all(result["iou_predictions"] == expected["iou_predictions"])
-        assert np.all(result["low_res_masks"] == expected["low_res_masks"])
+        assert np.all(result["masks"] == expected["masks"])
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
-        "high_res_masks,expected_masks,expected_scores",
+        "masks,expected_masks",
         [
             (
                 np.repeat(np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])[..., None], 4, axis=-1),
-                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_),
-                0.9,
+                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_)
             ),
             (
                 np.concatenate(
@@ -301,25 +300,21 @@ def test_forward_decoder(
                     ),
                     axis=-1,
                 ),
-                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_),
-                0.8,
+                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_)
             ),
-            (np.zeros((3, 3, 4)), np.zeros((3, 3)), 0.0),
+            (np.zeros((3, 3, 4)), np.zeros((3, 3))),
         ],
     )
-    def test_postprocess_masks(self, high_res_masks: np.ndarray, expected_masks: np.ndarray, expected_scores: float):
+    def test_postprocess_masks(self, masks: np.ndarray, expected_masks: np.ndarray):
         """Test _postprocess_masks."""
-        self.visual_prompting_ov_inferencer.model["decoder"].resize_and_crop.return_value = high_res_masks
         self.visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.0
         self.visual_prompting_ov_inferencer.model["decoder"].image_size = 3
 
-        _, result_masks, result_scores = self.visual_prompting_ov_inferencer._postprocess_masks(
-            logits=np.empty((1, 4, 2, 2)), scores=np.array([[0.5, 0.7, 0.8, 0.9]]), original_size=np.array([3, 3])
-        )
+        _, result_masks = self.visual_prompting_ov_inferencer._postprocess_masks(
+            masks=masks, logits=np.empty((1, 4, 2, 2)), scores=np.array([[0.5, 0.7, 0.8, 0.9]]))
 
         assert result_masks.shape == (3, 3)
         assert np.all(result_masks == expected_masks)
-        assert result_scores == expected_scores
 
 
 class TestOTXOpenVinoDataLoader:

From 205c7635912072460dca6d411e33f7edb955a301 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 13 Feb 2024 11:44:42 +0900
Subject: [PATCH 05/28] Update `learn` pipeline

---
 .../datasets/pipelines/transforms.py          |   4 -
 .../zero_shot_segment_anything.py             | 390 +++++++++---------
 .../visual_prompting/tasks/inference.py       |  11 +-
 .../datasets/pipelines/test_transforms.py     |  11 +-
 .../test_zero_shot_segment_anything.py        |  81 ++--
 .../visual_prompting/tasks/test_inference.py  |   5 +-
 .../visual_prompting/test_helpers.py          |   1 +
 7 files changed, 226 insertions(+), 277 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py
index f53fb4b3457..05b0693a39a 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py
@@ -43,7 +43,6 @@ def _convert_empty_to_none(x: str, dtype: torch.dtype = torch.float32) -> List:
     points = None  # TBD
     gt_masks = _convert_empty_to_none("gt_masks", torch.int32)
     original_size = _convert_empty_to_none("original_size")
-    padding = [item["padding"] for item in batch]
     path = [item["path"] for item in batch]
     labels = [item["labels"] for item in batch]
     if gt_masks:
@@ -56,7 +55,6 @@ def _convert_empty_to_none(x: str, dtype: torch.dtype = torch.float32) -> List:
             "original_size": original_size,
             "path": path,
             "labels": labels,
-            "padding": padding,
         }
     return {
         "index": -1,
@@ -67,7 +65,6 @@ def _convert_empty_to_none(x: str, dtype: torch.dtype = torch.float32) -> List:
         "original_size": [],
         "path": [],
         "labels": [],
-        "padding": [],
     }
 
 
@@ -89,7 +86,6 @@ def __call__(self, item: Dict[str, Union[List[Any], Tensor]]) -> Dict[str, Union
         pad_h = max_dim - h
         padding = (0, 0, pad_w, pad_h)
 
-        item["padding"] = padding
         item["images"] = pad(item["images"], padding, fill=0, padding_mode="constant")
 
         return item
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 4fceb645b39..1f26ae73b28 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -10,7 +10,8 @@
 
 import torch
 from omegaconf import DictConfig
-from torch import nn
+from torch import nn, Tensor
+from torch.nn import Parameter, ParameterDict
 from torch.nn import functional as F
 
 from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.pipelines import (
@@ -30,107 +31,73 @@ class PromptGetter(nn.Module):
     default_threshold_reference = 0.3
     default_threshold_target = 0.65
 
-    def __init__(
-        self,
-        image_size: int,
-        reference_feats: Optional[torch.Tensor] = None,
-        reference_prompts: Optional[torch.Tensor] = None,
-        downsizing: int = 64,
-    ) -> None:
+    def __init__(self, image_size: int, downsizing: int = 64) -> None:
         super().__init__()
         self.image_size = image_size
         self.downsizing = downsizing
-        self.initialize(reference_feats, reference_prompts)
 
-        self.zero_tensor = torch.tensor(0)
-
-    def initialize(
-        self, reference_feats: Optional[torch.Tensor] = None, reference_prompts: Optional[torch.Tensor] = None
-    ) -> None:
-        """Initialize reference features and prompts."""
-        self.reference_feats = reference_feats
-        self.reference_prompts = reference_prompts
+        self.zero_tensor = torch.as_tensor(0)
 
     def set_default_thresholds(self, default_threshold_reference: float, default_threshold_target: float) -> None:
         """Set default thresholds."""
         self.default_threshold_reference = default_threshold_reference
         self.default_threshold_target = default_threshold_target
 
-    def set_reference(self, label: ScoredLabel, reference_feats: torch.Tensor, reference_prompts: torch.Tensor) -> None:
-        """Set reference features and prompts."""
-        if self.reference_feats is None:
-            self.reference_feats = torch.zeros_like(reference_feats).unsqueeze(0)
-        if self.reference_prompts is None:
-            self.reference_prompts = torch.zeros_like(reference_prompts).unsqueeze(0)
-
-        for idx in range(int(label.id_) + 1):
-            if idx == int(label.id_):
-                while self.reference_feats.shape[0] - 1 < idx:
-                    self.reference_feats = torch.cat(
-                        (self.reference_feats, torch.zeros_like(reference_feats).unsqueeze(0)), dim=0
-                    )
-                    self.reference_prompts = torch.cat(
-                        (self.reference_prompts, torch.zeros_like(reference_prompts).unsqueeze(0)), dim=0
-                    )
-                self.reference_feats[idx] = reference_feats
-                self.reference_prompts[idx] = reference_prompts
-
     def forward(
         self,
-        image_embeddings: torch.Tensor,
-        original_size: torch.Tensor,
-        threshold: torch.Tensor = torch.tensor([[0.0]], dtype=torch.float32),
-        num_bg_points: torch.Tensor = torch.tensor([[1]], dtype=torch.int64),
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        image_embedding: Tensor,
+        reference_feats: Tensor,
+        used_indices: Tensor,
+        original_size: Tensor,
+        threshold: Tensor = torch.as_tensor([[0.0]], dtype=torch.float32),
+        num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
+    ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates."""
-        total_points_scores: torch.Tensor
-        total_bg_coords: torch.Tensor
-
-        device = image_embeddings.device
+        device = image_embedding.device
         threshold = threshold.to(device)
-        for label in torch.arange(self.reference_feats.shape[0]):
+
+        total_points_scores: Tensor = torch.zeros(used_indices.max() + 1, 0, 3, device=device)
+        total_bg_coords: Tensor = torch.zeros(used_indices.max() + 1, num_bg_points, 2, device=device)
+        for label in used_indices[0]:
             points_scores, bg_coords = self.get_prompt_candidates(
-                image_embeddings=image_embeddings,
-                label=label,
+                image_embedding=image_embedding,
+                reference_feat=reference_feats[label],
                 original_size=original_size,
                 threshold=threshold,
                 num_bg_points=num_bg_points,
                 device=device,
             )
-            if label == 0:
-                total_points_scores = points_scores.unsqueeze(0)
-                total_bg_coords = bg_coords.unsqueeze(0)
-            else:
-                pad_size = torch.tensor(points_scores.shape[0] - total_points_scores.shape[1])
-                pad_tot = torch.max(self.zero_tensor, pad_size)
-                pad_cur = torch.max(self.zero_tensor, -pad_size)
+                
+            pad_size = torch.as_tensor(points_scores.shape[0] - total_points_scores.shape[1])
+            pad_tot = torch.max(self.zero_tensor, pad_size)
+            pad_cur = torch.max(self.zero_tensor, -pad_size)
 
-                total_points_scores = F.pad(total_points_scores, (0, 0, 0, pad_tot, 0, 0), value=-1)
-                points_scores = F.pad(points_scores, (0, 0, 0, pad_cur), value=-1)
+            total_points_scores = F.pad(total_points_scores, (0, 0, 0, pad_tot, 0, 0), value=-1)
+            points_scores = F.pad(points_scores, (0, 0, 0, pad_cur), value=-1)
 
-                total_points_scores = torch.cat((total_points_scores, points_scores.unsqueeze(0)), dim=0)
-                total_bg_coords = torch.cat((total_bg_coords, bg_coords.unsqueeze(0)), dim=0)
+            total_points_scores[label] = points_scores
+            total_bg_coords[label] = bg_coords
 
         return total_points_scores, total_bg_coords
 
     def get_prompt_candidates(
         self,
-        image_embeddings: torch.Tensor,
-        label: torch.Tensor,
-        original_size: torch.Tensor,
-        threshold: torch.Tensor = torch.tensor([[0.0]], dtype=torch.float32),
-        num_bg_points: torch.Tensor = torch.tensor([[1]], dtype=torch.int64),
-        device: torch.device = torch.device("cpu"),
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        image_embedding: Tensor,
+        reference_feat: Tensor,
+        original_size: Tensor,
+        threshold: Tensor = torch.as_tensor([[0.0]], dtype=torch.float32),
+        num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
+        device: Union[torch.device, str] = torch.device("cpu"),
+    ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates from given reference and target features."""
         assert original_size.dim() == 2 and threshold.dim() == 2 and num_bg_points.dim() == 2
 
-        target_feat = image_embeddings.squeeze()
+        target_feat = image_embedding.squeeze()
         c_feat, h_feat, w_feat = target_feat.shape
         target_feat = target_feat / target_feat.norm(dim=0, keepdim=True)
         target_feat = target_feat.reshape(c_feat, h_feat * w_feat)
 
-        sim = self.reference_feats[label].to(device) @ target_feat
+        sim = reference_feat.to(device) @ target_feat
         sim = sim.reshape(1, 1, h_feat, w_feat)
         sim = ZeroShotSegmentAnything.postprocess_masks(sim, self.image_size, original_size[0])
 
@@ -146,11 +113,11 @@ def get_prompt_candidates(
 
     def _point_selection(
         self,
-        mask_sim: torch.Tensor,
-        original_size: torch.Tensor,
-        threshold: torch.Tensor,
-        num_bg_points: torch.Tensor = torch.tensor([[1]], dtype=torch.int64),
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        mask_sim: Tensor,
+        original_size: Tensor,
+        threshold: Tensor,
+        num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
+    ) -> Tuple[Tensor, Tensor]:
         """Select point used as point prompts."""
         _, w_sim = mask_sim.shape
 
@@ -202,40 +169,36 @@ def __init__(self, config: Optional[DictConfig] = None, state_dict: Optional[Ord
         if config is None:
             config = self.set_default_config()
 
-        if not config.model.freeze_image_encoder:
-            logger.warning("config.model.freeze_image_encoder(=False) must be set to True, changed.")
-            config.model.freeze_image_encoder = True
-
-        if not config.model.freeze_prompt_encoder:
-            logger.warning("config.model.freeze_prompt_encoder(=False) must be set to True, changed.")
-            config.model.freeze_prompt_encoder = True
-
-        if not config.model.freeze_mask_decoder:
-            logger.warning("config.model.freeze_mask_decoder(=False) must be set to True, changed.")
-            config.model.freeze_mask_decoder = True
-
-        prompt_getter_reference_feats = None
-        prompt_getter_reference_prompts = None
-        if state_dict:
-            if "prompt_getter.reference_feats" in state_dict:
-                prompt_getter_reference_feats = state_dict.pop("prompt_getter.reference_feats")
-            if "prompt_getter.reference_prompts" in state_dict:
-                prompt_getter_reference_prompts = state_dict.pop("prompt_getter.reference_prompts")
+        # check freeze conditions
+        for condition in ["freeze_image_encoder", "freeze_prompt_encoder", "freeze_mask_decoder"]:
+            if not getattr(config.model, condition, False):
+                logger.warning(f"config.model.{condition}(=False) must be set to True, changed.")
+                setattr(config.model, condition, True)
 
-        super().__init__(config, state_dict)
+        super().__init__(config, None)
+        self.set_empty_reference_info()
+        self._register_load_state_dict_pre_hook(self.load_state_dict_pre_hook)
+        self.load_checkpoint(state_dict)
 
-        self.prompt_getter = PromptGetter(
-            image_size=config.model.image_size,
-            reference_feats=prompt_getter_reference_feats,
-            reference_prompts=prompt_getter_reference_prompts,
-        )
+        self.prompt_getter = PromptGetter(image_size=config.model.image_size)
         self.prompt_getter.set_default_thresholds(
             default_threshold_reference=config.model.default_threshold_reference,
             default_threshold_target=config.model.default_threshold_target,
         )
 
-        self.point_labels_box = torch.tensor([[2, 3]], dtype=torch.float32)
-        self.has_mask_inputs = [torch.tensor([[0.0]]), torch.tensor([[1.0]])]
+        self.point_labels_box = torch.as_tensor([[2, 3]], dtype=torch.float32)
+        self.has_mask_inputs = [torch.as_tensor([[0.0]]), torch.as_tensor([[1.0]])]
+        
+    def load_state_dict_pre_hook(self, state_dict: dict[str, Any], prefix: str = "", *args, **kwargs) -> None:
+        """Load reference info manually."""
+        _reference_feats: Tensor = state_dict.get("reference_info.reference_feats", torch.as_tensor([], dtype=torch.float32))
+        _used_indices: Tensor = state_dict.get("reference_info.used_indices", torch.as_tensor([], dtype=torch.int64))
+        self.reference_info = ParameterDict(
+            {
+                "reference_feats": Parameter(_reference_feats, requires_grad=False),
+                "used_indices": Parameter(_used_indices, requires_grad=False),
+            },
+        )
 
     def set_default_config(self) -> DictConfig:
         """Set default config when using independently."""
@@ -254,14 +217,42 @@ def set_default_config(self) -> DictConfig:
                 }
             }
         )
+        
+    def set_empty_reference_info(self) -> None:
+        """Set empty reference information."""
+        reference_feats: Parameter = Parameter(torch.as_tensor([], dtype=torch.float32), requires_grad=False)
+        used_indices: Parameter = Parameter(torch.as_tensor([], dtype=torch.int64), requires_grad=False)
+        self.reference_info = ParameterDict(
+            {
+                "reference_feats": reference_feats,
+                "used_indices": used_indices,
+            },
+        )
+        self.is_reference_info_empty = True
+        
+    def initialize_reference_info(self, largest_label: int) -> None:
+        """Initialize reference information."""
+        self.reference_info["reference_feats"] = Parameter(torch.zeros(largest_label + 1, 1, 256), requires_grad=False)
+        self.reference_info["used_indices"] = Parameter(torch.as_tensor([], dtype=torch.int64), requires_grad=False)
+        self.is_reference_info_empty = False
+        
+    def expand_reference_info(self, new_largest_label: int) -> None:
+        """Expand reference info dimensions if newly given processed prompts have more lables."""
+        if new_largest_label > (cur_largest_label := len(self.reference_info["reference_feats"]) - 1):
+            diff = new_largest_label - cur_largest_label
+            self.reference_info["reference_feats"] = F.pad(
+                self.reference_info["reference_feats"],
+                (0, 0, 0, 0, 0, diff),
+                value=0.0,
+            )
 
     @torch.no_grad()
     def learn(
         self,
-        images: torch.Tensor,
-        processed_prompts: Dict[ScoredLabel, List[Dict[str, torch.Tensor]]],
-        padding: Union[Tuple[int, ...], torch.Tensor],
-        original_size: torch.Tensor,
+        images: Tensor,
+        processed_prompts: Dict[ScoredLabel, List[Dict[str, Tensor]]],
+        original_size: Tensor,
+        return_outputs: bool = False,
     ) -> None:
         """Get reference features.
 
@@ -270,19 +261,25 @@ def learn(
         Currently, single batch is only supported.
 
         Args:
-            images (torch.Tensor): Given images for reference features.
-            processed_prompts (Dict[ScoredLabel, List[Dict[str, torch.Tensor]]]): The whole class-wise prompts
+            images (Tensor): Given images for reference features.
+            processed_prompts (Dict[ScoredLabel, List[Dict[str, Tensor]]]): The whole class-wise prompts
                 processed at _preprocess_prompts.
-            padding (Union[Tuple[int, ...], torch.Tensor]): Padding size.
-            original_size (torch.Tensor): Original image size.
+            original_size (Tensor): Original image size.
         """
         assert images.shape[0] == 1, "Only single batch is supported."
-
-        self.prompt_getter.initialize()
+        
+        # initialize tensors to contain reference features and prompts
+        largest_label = max([int(label.id) for label in processed_prompts.keys()])
+        if self.is_reference_info_empty:
+            self.initialize_reference_info(largest_label)
+        else:
+            self.expand_reference_info(largest_label)
+            # TODO(sungchul): consider who to handle multiple reference features, currently replace it # noqa: TD003
 
         image_embeddings = self.image_encoder(images)
-        ref_feat = image_embeddings.squeeze().permute(1, 2, 0)
-
+        processed_embedding = image_embeddings.squeeze().permute(1, 2, 0)
+        
+        ref_masks = torch.zeros(largest_label + 1, *map(int, original_size))
         for label, input_prompts in processed_prompts.items():
             if label.name.lower() == "background":
                 # skip background
@@ -291,32 +288,28 @@ def learn(
 
             # generate reference mask
             # TODO (sungchul): ensemble multi reference features (current : use merged masks)
-            reference_prompt = torch.zeros(*map(int, original_size), dtype=torch.uint8, device=self.device)
+            ref_mask = torch.zeros(*map(int, original_size), dtype=torch.uint8, device=self.device)
             for input_prompt in input_prompts:
-                if "annotation" in input_prompt:
+                if (anno_prompt := input_prompt.get("annotation", None)) is not None:
                     # directly use annotation information as a mask
-                    reference_prompt[input_prompt.get("annotation") == 1] += 1
+                    ref_mask[anno_prompt == 1] += 1
                 else:
-                    merged_input_prompts = self._merge_prompts(label, input_prompt, processed_prompts)
+                    merged_input_prompts: Dict[str, Tensor] = self._merge_prompts(label, input_prompt, processed_prompts)
                     # TODO (sungchul): they must be processed in `_merge_prompts`
                     # and it is required to be expanded to other prompts.
                     point_coords = []
                     point_labels = []
-                    if "box" in merged_input_prompts:
-                        for box in merged_input_prompts["box"]:
-                            point_coords.append(box[:2])
-                            point_labels.append(2)
-                            point_coords.append(box[2:])
-                            point_labels.append(3)
-
-                    if "points" in merged_input_prompts:
-                        raise NotImplementedError()
+                    if (input_prompt := merged_input_prompts.get("box", None)) is not None:
+                        point_coords = input_prompt.reshape(1, 2, 2)
+                        point_labels = torch.as_tensor([[2, 3]], device=self.device)
+
+                    if (input_prompt := merged_input_prompts.get("points", None)) is not None:
+                        point_coords = input_prompt.reshape(1, 1, 2)
+                        point_labels = torch.as_tensor([[1]], device=self.device)
 
-                    if "annotations" in merged_input_prompts:
+                    if (input_prompt := merged_input_prompts.get("annotations", None)) is not None:
                         raise NotImplementedError()
 
-                    point_coords = torch.stack(point_coords, dim=0).unsqueeze(0)
-                    point_labels = torch.tensor([point_labels], device=self.device)
                     masks = self._predict_masks(
                         image_embeddings=image_embeddings,
                         point_coords=point_coords,
@@ -324,35 +317,38 @@ def learn(
                         original_size=original_size.unsqueeze(0),
                         is_cascade=False,
                     )
-                    reference_prompt[masks] += 1
-            reference_prompt = torch.clip(reference_prompt, 0, 1)
+                    ref_mask[masks] += 1
+            ref_mask = torch.clip(ref_mask, 0, 1).to(torch.float32)
 
-            ref_mask = reference_prompt.to(torch.float32)
-            reference_feat = None
+            ref_feat = None
             default_threshold_reference = deepcopy(self.prompt_getter.default_threshold_reference)
-            while reference_feat is None:
+            while ref_feat is None:
                 logger.info(f"[*] default_threshold_reference : {default_threshold_reference:.4f}")
-                reference_feat = self._generate_masked_features(
-                    ref_feat, ref_mask, default_threshold_reference, padding=padding
-                )
+                ref_feat = self._generate_masked_features(processed_embedding, ref_mask, default_threshold_reference)
                 default_threshold_reference -= 0.05
 
-            self.prompt_getter.set_reference(label, reference_feat, reference_prompt)
+            self.reference_info["reference_feats"][int(label.id)] = ref_feat.detach().cpu()
+            self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([int(label.id)]))), requires_grad=False)
+            ref_masks[int(label.id)] = ref_mask.detach().cpu()
+            
+        self.reference_info["used_indices"] = Parameter(self.reference_info["used_indices"].unsqueeze(0), requires_grad=False)
+        if return_outputs:
+            return self.reference_info, ref_masks
 
     @torch.no_grad()
     def infer(
-        self, images: torch.Tensor, original_size: torch.Tensor
-    ) -> List[List[DefaultDict[int, List[torch.Tensor]]]]:
+        self, images: Tensor, original_size: Tensor
+    ) -> List[List[DefaultDict[int, List[Tensor]]]]:
         """Zero-shot inference with reference features.
 
         Get target results by using reference features and target images' features.
 
         Args:
-            images (torch.Tensor): Given images for target results.
-            original_size (torch.Tensor): Original image size.
+            images (Tensor): Given images for target results.
+            original_size (Tensor): Original image size.
 
         Returns:
-            (List[List[DefaultDict[int, List[torch.Tensor]]]]): Target results.
+            (List[List[DefaultDict[int, List[Tensor]]]]): Target results.
                 Lists wrapping results is following this order:
                     1. Target images
                     2. Tuple of predicted masks and used points gotten by point selection
@@ -364,10 +360,10 @@ def infer(
             if image.ndim == 3:
                 image = image.unsqueeze(0)
 
-            image_embeddings = self.image_encoder(images)
+            image_embedding = self.image_encoder(images)
 
             total_points_scores, total_bg_coords = self.prompt_getter(
-                image_embeddings=image_embeddings, original_size=original_size
+                image_embedding=image_embedding, original_size=original_size
             )
             predicted_masks: defaultdict = defaultdict(list)
             used_points: defaultdict = defaultdict(list)
@@ -389,11 +385,11 @@ def infer(
                     point_coords = ResizeLongestSide.apply_coords(
                         point_coords, original_size[0], self.config.model.image_size
                     )
-                    point_labels = torch.tensor(
+                    point_labels = torch.as_tensor(
                         [1] + [0] * len(bg_coords), dtype=torch.float32, device=self.device
                     ).unsqueeze(0)
                     mask = self._predict_masks(
-                        image_embeddings=image_embeddings,
+                        image_embeddings=image_embedding,
                         point_coords=point_coords,
                         point_labels=point_labels,
                         original_size=original_size,
@@ -408,11 +404,11 @@ def infer(
 
     def __inspect_overlapping_areas(
         self,
-        predicted_masks: Dict[int, List[torch.Tensor]],
-        used_points: Dict[int, List[torch.Tensor]],
+        predicted_masks: Dict[int, List[Tensor]],
+        used_points: Dict[int, List[Tensor]],
         threshold_iou: float = 0.8,
     ):
-        def __calculate_mask_iou(mask1: torch.Tensor, mask2: torch.Tensor):
+        def __calculate_mask_iou(mask1: Tensor, mask2: Tensor):
             assert mask1.ndim == 2 and mask2.ndim == 2
             intersection = torch.logical_and(mask1, mask2).sum().item()
             union = torch.logical_or(mask1, mask2).sum().item()
@@ -446,15 +442,15 @@ def __calculate_mask_iou(mask1: torch.Tensor, mask2: torch.Tensor):
 
     def _predict_masks(
         self,
-        image_embeddings: torch.Tensor,
-        point_coords: torch.Tensor,
-        point_labels: torch.Tensor,
-        original_size: torch.Tensor,
+        image_embeddings: Tensor,
+        point_coords: Tensor,
+        point_labels: Tensor,
+        original_size: Tensor,
         is_cascade: bool = True,
-    ) -> torch.Tensor:
+    ) -> Tensor:
         """Predict target masks."""
-        logits: torch.Tensor
-        scores: torch.Tensor
+        logits: Tensor
+        scores: Tensor
         for i in range(3):
             if i == 0:
                 # First-step prediction
@@ -479,7 +475,7 @@ def _predict_masks(
                 coords = torch.nonzero(masks)
                 y, x = coords[:, 0], coords[:, 1]
                 box_coords = ResizeLongestSide.apply_coords(
-                    torch.tensor([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device),
+                    torch.as_tensor([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device),
                     original_size[0],
                     self.config.model.image_size,
                 )
@@ -513,7 +509,6 @@ def training_step(self, batch, batch_idx) -> None:
         self.learn(
             images=batch["images"],
             processed_prompts=processed_prompts,
-            padding=batch.get("padding")[0],
             original_size=batch.get("original_size")[0],
         )
 
@@ -524,24 +519,24 @@ def predict_step(self, batch, batch_idx):
 
     def _preprocess_prompts(
         self,
-        bboxes: Optional[torch.Tensor] = None,
-        points: Optional[torch.Tensor] = None,
-        annotations: Optional[torch.Tensor] = None,
-        labels: Optional[torch.Tensor] = None,
-    ) -> Dict[ScoredLabel, List[Dict[str, torch.Tensor]]]:
+        bboxes: Optional[Tensor] = None,
+        points: Optional[Tensor] = None,
+        annotations: Optional[Tensor] = None,
+        labels: Optional[Tensor] = None,
+    ) -> Dict[ScoredLabel, List[Dict[str, Tensor]]]:
         """Preprocess prompts.
 
         Currently, preprocessing for bounding boxes is only supported.
 
         Args:
-            bboxes (torch.Tensor, optional): Bounding box prompts to be preprocessed.
-            points (torch.Tensor, optional): Point prompts to be preprocessed, to be supported.
-            annotations (torch.Tensor, optional): annotation prompts to be preprocessed, to be supported.
-            labels (torch.Tensor, optional): Assigned labels according to given prompts.
+            bboxes (Tensor, optional): Bounding box prompts to be preprocessed.
+            points (Tensor, optional): Point prompts to be preprocessed, to be supported.
+            annotations (Tensor, optional): annotation prompts to be preprocessed, to be supported.
+            labels (Tensor, optional): Assigned labels according to given prompts.
                 Currently, it is only matched to bboxes, and it will be deprecated.
 
         Returns:
-            (defaultdict[ScoredLabel, List[Dict[str, torch.Tensor]]]): Processed and arranged each single prompt
+            (defaultdict[ScoredLabel, List[Dict[str, Tensor]]]): Processed and arranged each single prompt
                 using label information as keys. Unlike other prompts, `annotation` prompts will be aggregated
                 as single annotation.
         """
@@ -562,39 +557,31 @@ def _preprocess_prompts(
 
     def _generate_masked_features(
         self,
-        feats: torch.Tensor,
-        masks: torch.Tensor,
+        feats: Tensor,
+        masks: Tensor,
         threshold_mask: float,
-        padding: Optional[Union[Tuple[int, ...], torch.Tensor]] = None,
-    ) -> Tuple[torch.Tensor, ...]:
+    ) -> Tuple[Tensor, ...]:
         """Generate masked features.
 
         Args:
-            feats (torch.Tensor): Raw reference features. It will be filtered with masks.
-            masks (torch.Tensor): Reference masks used to filter features.
+            feats (Tensor): Raw reference features. It will be filtered with masks.
+            masks (Tensor): Reference masks used to filter features.
             threshold_mask (float): Threshold to control masked region.
-            padding (Union[Tuple[int, ...], torch.Tensor], optional): Padding size.
 
         Returns:
-            (torch.Tensor): Masked features.
+            (Tensor): Masked features.
         """
-        if padding:
-            resized_size = (
-                self.config.model.image_size - padding[1] - padding[3],
-                self.config.model.image_size - padding[0] - padding[2],
-            )
-        else:
-            resized_size = (self.config.model.image_size, self.config.model.image_size)
+        scale_factor = self.config.model.image_size / max(masks.shape)
 
         # Post-process masks
-        masks = F.interpolate(masks.unsqueeze(0).unsqueeze(0), size=resized_size, mode="bilinear").squeeze()
-        masks = self._preprocess_masks(masks)
+        masks = F.interpolate(masks.unsqueeze(0).unsqueeze(0), scale_factor=scale_factor, mode="bilinear").squeeze()
+        masks = self._pad_to_square(masks)
         masks = F.interpolate(masks.unsqueeze(0).unsqueeze(0), size=feats.shape[0:2], mode="bilinear").squeeze()
 
         # Target feature extraction
         if (masks > threshold_mask).sum() == 0:
             # (for stability) there is no area to be extracted
-            return None, None
+            return None
 
         masked_feat = feats[masks > threshold_mask]
         masked_feat = masked_feat.mean(0).unsqueeze(0)
@@ -602,16 +589,15 @@ def _generate_masked_features(
 
         return masked_feat
 
-    def _preprocess_masks(self, x: torch.Tensor) -> torch.Tensor:
-        """Normalize pixel values and pad to a square input.
+    def _pad_to_square(self, x: Tensor) -> Tensor:
+        """Pad to a square input.
 
         Args:
-            x (torch.Tensor): Mask to be padded.
+            x (Tensor): Mask to be padded.
 
         Returns:
-            (torch.Tensor): Padded mask.
+            (Tensor): Padded mask.
         """
-        # Pad
         h, w = x.shape[-2:]
         padh = self.config.model.image_size - h
         padw = self.config.model.image_size - w
@@ -620,9 +606,9 @@ def _preprocess_masks(self, x: torch.Tensor) -> torch.Tensor:
 
     def _postprocess_masks(
         self,
-        masks: torch.Tensor,
-        logits: torch.Tensor,
-        scores: torch.Tensor,
+        masks: Tensor,
+        logits: Tensor,
+        scores: Tensor,
         is_single: bool = False,
     ):
         """Post-process masks for cascaded post-refinements."""
@@ -645,13 +631,13 @@ def _postprocess_masks(
             best_idx = torch.argmax(scores[0])
         return logits[:, best_idx], masks[0, best_idx]
 
-    def _update_value(self, target: Dict[str, Any], key: str, value: torch.Tensor) -> None:
+    def _update_value(self, target: Dict[str, Any], key: str, value: Tensor) -> None:
         """Update tensor to target dictionary.
 
         Args:
             target (Dict[str, Any]): Target dictionary to be updated.
             key (str): Key to be used for update.
-            value (torch.Tensor): Value to be used for update.
+            value (Tensor): Value to be used for update.
         """
         if key in target:
             target[key] = torch.cat((target[key], value))
@@ -661,24 +647,24 @@ def _update_value(self, target: Dict[str, Any], key: str, value: torch.Tensor) -
     def _merge_prompts(
         self,
         label: ScoredLabel,
-        input_prompts: Dict[str, torch.Tensor],
-        processed_prompts: Dict[ScoredLabel, List[Dict[str, torch.Tensor]]],
+        input_prompts: Dict[str, Tensor],
+        processed_prompts: Dict[ScoredLabel, List[Dict[str, Tensor]]],
         use_only_background: bool = True,
-    ) -> Dict[str, torch.Tensor]:
+    ) -> Dict[str, Tensor]:
         """Merge target prompt and other prompts.
 
         Merge a foreground prompt and other prompts (background or prompts with other classes).
 
         Args:
             label (ScoredLabel): Label information. Background is 0 and other foregrounds are >= 0.
-            input_prompts (Dict[str, torch.Tensor]): A foreground prompt to be merged with other prompts.
-            processed_prompts (Dict[ScoredLabel, List[Dict[str, torch.Tensor]]]): The whole class-wise prompts
+            input_prompts (Dict[str, Tensor]): A foreground prompt to be merged with other prompts.
+            processed_prompts (Dict[ScoredLabel, List[Dict[str, Tensor]]]): The whole class-wise prompts
                 processed at _preprocess_prompts.
             use_only_background (bool): Whether merging only background prompt, defaults to True.
                 It is applied to only point_coords.
 
         Returns:
-            (Dict[str, torch.Tensor]): Merged prompts.
+            (Dict[str, Tensor]): Merged prompts.
         """
         merged_input_prompts = deepcopy(input_prompts)
         for other_label, other_input_prompts in processed_prompts.items():
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index 8bff9bd55f7..aca6039d678 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -643,12 +643,16 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
             elif module == "visual_prompting_prompt_getter":
                 dummy_inputs = {
                     "image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float32),
+                    "reference_feats": torch.randn(1, 1, 256, dtype=torch.float32),
+                    "used_indices": torch.as_tensor([[0]], dtype=torch.int64),
                     "original_size": torch.randint(low=0, high=image_size * 2, size=(1, 2), dtype=torch.int64),
                     "threshold": torch.tensor([[0.1]], dtype=torch.float32),
                     "num_bg_points": torch.randint(low=1, high=image_size, size=(1, 1), dtype=torch.int64),
                 }
                 output_names = ["total_points_scores", "total_bg_coords"]
                 dynamic_axes = {
+                    "reference_feats": {0: "num_labels"},
+                    "used_indices": {1: "num_labels"},
                     "total_points_scores": {0: "num_labels", 1: "num_points"},
                     "total_bg_coords": {0: "num_labels", 1: "num_points"},
                 }
@@ -706,13 +710,6 @@ def save_model(self, output_model: ModelEntity) -> None:
         logger.info("Saving the model weights and reference features.")
 
         model_info = self.model.state_dict()
-        # TODO (sungchul): is there more efficient way not to manually add properties?
-        model_info.update(
-            {
-                "prompt_getter.reference_feats": self.model.prompt_getter.reference_feats,
-                "prompt_getter.reference_prompts": self.model.prompt_getter.reference_prompts,
-            }
-        )
 
         buffer = io.BytesIO()
         torch.save(model_info, buffer)
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
index 36225af9d4a..a703c7512c6 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
@@ -31,7 +31,6 @@ def test_collate_fn():
             "points": [],
             "gt_masks": [Tensor([1, 2, 3])],
             "original_size": np.array([1, 3]),
-            "padding": [],
             "path": [],
             "labels": [],
         },
@@ -42,7 +41,6 @@ def test_collate_fn():
             "points": [],
             "gt_masks": [Tensor([4, 5, 6])],
             "original_size": np.array([1, 3]),
-            "padding": [],
             "path": [],
             "labels": [],
         },
@@ -56,7 +54,6 @@ def test_collate_fn():
         "original_size": [Tensor([1, 3]), Tensor([1, 3])],
         "path": [[], []],
         "labels": [[], []],
-        "padding": [[], []],
     }
 
     results = collate_fn(batch)
@@ -73,7 +70,6 @@ def test_collate_fn():
         assert torch.all(r == e)
     assert results["path"] == expected["path"]
     assert results["labels"] == expected["labels"]
-    assert results["padding"] == expected["padding"]
 
 
 class TestPad:
@@ -88,22 +84,21 @@ class TestPad:
                     bboxes=[[1, 1, 3, 3]],
                     points=[[1, 1, 2, 2]],
                 ),
-                ((0, 0, 0, 2), (3, 6, 6), [(4, 6)], [[1, 1, 3, 3]], [[1, 1, 2, 2]]),
+                ((3, 6, 6), [(4, 6)], [[1, 1, 3, 3]], [[1, 1, 2, 2]]),
             ),
             (
                 dict(images=torch.zeros((3, 4, 6)), gt_masks=[torch.zeros((4, 6))], bboxes=[[1, 1, 3, 3]], points=None),
-                ((0, 0, 0, 2), (3, 6, 6), [(4, 6)], [[1, 1, 3, 3]], None),
+                ((3, 6, 6), [(4, 6)], [[1, 1, 3, 3]], None),
             ),
         ],
     )
     def test_call(self, item: Dict[str, Any], expected: Tuple[Any]):
         """Test __call__."""
         pad_transform = Pad()
-        expected_padding, expected_images_shape, expected_gt_masks_shape, expected_bboxes, expected_points = expected
+        expected_images_shape, expected_gt_masks_shape, expected_bboxes, expected_points = expected
 
         result = pad_transform(item)
 
-        assert result["padding"] == expected_padding
         assert result["images"].shape == expected_images_shape
         assert len(result["gt_masks"]) == len(expected_gt_masks_shape)
         assert all(gt_mask.shape == shape for gt_mask, shape in zip(result["gt_masks"], expected_gt_masks_shape))
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 8d145bfcaaa..a3478ccb59a 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -9,6 +9,7 @@
 from collections import OrderedDict
 from tests.test_suite.e2e_test_system import e2e_pytest_unit
 import torch
+from torch import nn
 from omegaconf import DictConfig
 
 from otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything import (
@@ -31,12 +32,6 @@ class TestPromptGetter:
     def setup(self) -> None:
         self.prompt_getter = PromptGetter(image_size=3, downsizing=1)
 
-    @e2e_pytest_unit
-    def test_initialize(self) -> None:
-        """Test initialize."""
-        assert not self.prompt_getter.reference_feats
-        assert not self.prompt_getter.reference_prompts
-
     @e2e_pytest_unit
     def test_set_default_thresholds(self) -> None:
         """Test set_default_thresholds."""
@@ -48,31 +43,6 @@ def test_set_default_thresholds(self) -> None:
         assert self.prompt_getter.default_threshold_reference == 0.5
         assert self.prompt_getter.default_threshold_target == 0.7
 
-    @e2e_pytest_unit
-    def test_set_reference(self) -> None:
-        """Test set_reference."""
-        self.prompt_getter.set_reference(
-            label=MockScoredLabel(label=1),
-            reference_feats=torch.ones((self.prompt_getter.image_size, self.prompt_getter.image_size)),
-            reference_prompts=torch.ones((self.prompt_getter.image_size, self.prompt_getter.image_size)),
-        )
-
-        assert self.prompt_getter.reference_feats[0].sum() == 0
-        assert self.prompt_getter.reference_prompts[0].sum() == 0
-        assert self.prompt_getter.reference_feats[1].sum() == 9
-        assert self.prompt_getter.reference_prompts[1].sum() == 9
-
-        self.prompt_getter.set_reference(
-            label=MockScoredLabel(label=3),
-            reference_feats=torch.ones((self.prompt_getter.image_size, self.prompt_getter.image_size)),
-            reference_prompts=torch.ones((self.prompt_getter.image_size, self.prompt_getter.image_size)),
-        )
-
-        assert self.prompt_getter.reference_feats[2].sum() == 0
-        assert self.prompt_getter.reference_prompts[2].sum() == 0
-        assert self.prompt_getter.reference_feats[3].sum() == 9
-        assert self.prompt_getter.reference_prompts[3].sum() == 9
-
     @e2e_pytest_unit
     @pytest.mark.parametrize("result_point_selection", [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])])
     def test_forward(self, mocker, result_point_selection: torch.Tensor) -> None:
@@ -81,12 +51,13 @@ def test_forward(self, mocker, result_point_selection: torch.Tensor) -> None:
             self.prompt_getter,
             "get_prompt_candidates",
             return_value=(result_point_selection, torch.zeros(1, 2)))
-        image_embeddings = torch.ones(1, 4, 4, 4)
-        self.prompt_getter.reference_feats = torch.rand(1, 1, 4)
+        image_embedding = torch.ones(1, 4, 4, 4)
+        reference_feats = torch.rand(1, 1, 4)
+        used_indices = torch.as_tensor([[0]])
         original_size = torch.tensor((self.prompt_getter.image_size, self.prompt_getter.image_size), dtype=torch.int64)
 
         total_points_scores, total_bg_coords = self.prompt_getter(
-            image_embeddings=image_embeddings, original_size=original_size
+            image_embedding=image_embedding, reference_feats=reference_feats, used_indices=used_indices, original_size=original_size
         )
         
         assert total_points_scores.shape[0] == 1
@@ -100,15 +71,14 @@ def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tenso
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
         )
         mocker.patch.object(self.prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2)))
-        image_embeddings = torch.ones(1, 4, 4, 4)
-        self.prompt_getter.reference_feats = torch.rand(1, 1, 4)
-        label = torch.tensor([[0]], dtype=torch.int64)
+        image_embedding = torch.ones(1, 4, 4, 4)
+        reference_feat = torch.rand(1, 4)
         original_size = torch.tensor(
             [[self.prompt_getter.image_size, self.prompt_getter.image_size]], dtype=torch.int64
         )
 
         points_scores, bg_coords = self.prompt_getter.get_prompt_candidates(
-            image_embeddings=image_embeddings, label=label, original_size=original_size
+            image_embedding=image_embedding, reference_feat=reference_feat, original_size=original_size
         )
 
         assert torch.all(points_scores == result_point_selection)
@@ -154,13 +124,18 @@ def zero_shot_segment_anything(state_dict: Optional[OrderedDict] = None):
         [
             None,
             {
-                "prompt_getter.reference_feats": "prompt_getter.reference_feats",
-                "prompt_getter.reference_prompts": "prompt_getter.reference_prompts",
+                "reference_info": nn.ParameterDict({
+                    "reference_feats": "reference_feats",
+                    "used_indices": "used_indices"})
             },
         ],
     )
-    def test_init(self, set_zero_shot_segment_anything, state_dict: Dict[str, Any]) -> None:
+    def test_init(self, set_zero_shot_segment_anything, state_dict: Optional[Dict[str, Any]]) -> None:
         """Test __init__."""
+        if state_dict is not None:
+            zero_shot_segment_anything_for_init_weights = set_zero_shot_segment_anything()
+            state_dict.update(zero_shot_segment_anything_for_init_weights.state_dict())
+        
         zero_shot_segment_anything = set_zero_shot_segment_anything(state_dict=state_dict)
 
         assert zero_shot_segment_anything.config.model.freeze_image_encoder
@@ -168,8 +143,8 @@ def test_init(self, set_zero_shot_segment_anything, state_dict: Dict[str, Any])
         assert zero_shot_segment_anything.config.model.freeze_mask_decoder
 
         if state_dict:
-            zero_shot_segment_anything.prompt_getter.reference_feats = "prompt_getter.reference_feats"
-            zero_shot_segment_anything.prompt_getter.reference_prompts = "prompt_getter.reference_prompts"
+            zero_shot_segment_anything.reference_info.reference_feats = "reference_feats"
+            zero_shot_segment_anything.reference_info.used_indices = "used_indices"
 
     @e2e_pytest_unit
     def test_set_default_config(self, set_zero_shot_segment_anything) -> None:
@@ -197,19 +172,19 @@ def test_learn(self, mocker, set_zero_shot_segment_anything) -> None:
         mocker.patch.object(
             zero_shot_segment_anything,
             "_predict_masks",
-            return_value=torch.tensor([[[[0, 0, 0], [0, 1, 0], [0, 0, 0]]]]),
+            return_value=torch.tensor([[[[0, 0, 0, 0], [0, 1, 1, 0], [0, 1, 1, 0], [0, 0, 0, 0]]]]),
         )
+        mocker.patch.object(zero_shot_segment_anything, "_generate_masked_features", return_value=torch.ones(1, 256))
 
-        processed_prompts = {MockScoredLabel(label=1, name="label"): [{"box": torch.tensor([[0, 0, 1, 1]])}]}
+        processed_prompts = {MockScoredLabel(label=0, name="label"): [{"box": torch.tensor([[0, 0, 1, 1]])}]}
         zero_shot_segment_anything.learn(
             images=torch.ones((1, 3, 4, 4)),
             processed_prompts=processed_prompts,
-            padding=(0, 0, 0, 0),
             original_size=torch.tensor((4, 4)),
         )
 
-        assert zero_shot_segment_anything.prompt_getter.reference_feats.shape == (2, 1, 3)
-        assert zero_shot_segment_anything.prompt_getter.reference_prompts.shape == (2, 4, 4)
+        assert zero_shot_segment_anything.reference_info.reference_feats.shape == (1, 1, 256)
+        assert zero_shot_segment_anything.reference_info.used_indices == torch.as_tensor([0])
 
     @e2e_pytest_unit
     @pytest.mark.parametrize("expected", [[torch.ones((4, 4)) / 2, torch.tensor([0.0, 0.0, 0.5])]])
@@ -221,8 +196,8 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
         )
 
         zero_shot_segment_anything = set_zero_shot_segment_anything()
-        zero_shot_segment_anything.prompt_getter.reference_feats = torch.rand(1, 1, 4)
-        zero_shot_segment_anything.prompt_getter.reference_prompts = torch.zeros((4, 4))
+        zero_shot_segment_anything.reference_info.reference_feats = torch.rand(1, 1, 256)
+        zero_shot_segment_anything.reference_info.used_indices = {0}
         mocker.patch.object(
             SegmentAnything, "forward", return_value=(torch.ones(1, 4, 4, 4), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
         )
@@ -286,12 +261,12 @@ def test_generate_masked_features(self, set_zero_shot_segment_anything) -> None:
         assert masked_feat.shape == (1, 1)
 
     @e2e_pytest_unit
-    def test_preprocess_masks(self, set_zero_shot_segment_anything) -> None:
-        """Test _preprocess_masks."""
+    def test_pad_to_square(self, set_zero_shot_segment_anything) -> None:
+        """Test _pad_to_square."""
         zero_shot_segment_anything = set_zero_shot_segment_anything()
         zero_shot_segment_anything.config.model.image_size = 16
 
-        result = zero_shot_segment_anything._preprocess_masks(x=torch.ones(1, 1, 8, 8))
+        result = zero_shot_segment_anything._pad_to_square(x=torch.ones(1, 1, 8, 8))
 
         assert result[:8, :8].sum() == 8**2
         assert result[:8, 8:].sum() == 0
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index 051ddcdfe8f..d2a5852e0af 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -379,10 +379,9 @@ def test_save_model(self, mocker):
         mocker_otx_model = mocker.patch("otx.api.entities.model.ModelEntity")
         mocker_io_bytes_io = mocker.patch("io.BytesIO")
         mocker_torch_save = mocker.patch("torch.save")
+        mocker.patch.object(self.zero_shot_task.model, "state_dict", return_value={"reference_info.reference_feats": None})
 
-        self.zero_shot_task.model.prompt_getter = mocker.MagicMock()
-        self.zero_shot_task.model.prompt_getter.reference_feats.return_value = "reference_feats"
-        self.zero_shot_task.model.prompt_getter.reference_prompts.return_value = "reference_prompts"
+        self.zero_shot_task.model.reference_info = "reference_info"
 
         self.zero_shot_task.save_model(mocker_otx_model)
 
diff --git a/tests/unit/algorithms/visual_prompting/test_helpers.py b/tests/unit/algorithms/visual_prompting/test_helpers.py
index fb585e22a87..ac16b4cea4e 100644
--- a/tests/unit/algorithms/visual_prompting/test_helpers.py
+++ b/tests/unit/algorithms/visual_prompting/test_helpers.py
@@ -185,6 +185,7 @@ class MockScoredLabel:
     def __init__(self, label: int, name: str = "background"):
         self.name = name
         self.id_ = label
+        self.id = label
 
 
 class MockPromptGetter(nn.Module):

From e661f0aae45544381cb19ed6e79d3d4b4508efc6 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 13 Feb 2024 13:52:57 +0900
Subject: [PATCH 06/28] Update `infer` pipeline

---
 .../zero_shot_segment_anything.py             |  69 +++++++---
 .../test_zero_shot_segment_anything.py        | 128 ++++++++++++++++--
 2 files changed, 171 insertions(+), 26 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 1f26ae73b28..8330b312b22 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -221,7 +221,7 @@ def set_default_config(self) -> DictConfig:
     def set_empty_reference_info(self) -> None:
         """Set empty reference information."""
         reference_feats: Parameter = Parameter(torch.as_tensor([], dtype=torch.float32), requires_grad=False)
-        used_indices: Parameter = Parameter(torch.as_tensor([], dtype=torch.int64), requires_grad=False)
+        used_indices: Parameter = Parameter(torch.as_tensor([[]], dtype=torch.int64), requires_grad=False)
         self.reference_info = ParameterDict(
             {
                 "reference_feats": reference_feats,
@@ -233,7 +233,7 @@ def set_empty_reference_info(self) -> None:
     def initialize_reference_info(self, largest_label: int) -> None:
         """Initialize reference information."""
         self.reference_info["reference_feats"] = Parameter(torch.zeros(largest_label + 1, 1, 256), requires_grad=False)
-        self.reference_info["used_indices"] = Parameter(torch.as_tensor([], dtype=torch.int64), requires_grad=False)
+        self.reference_info["used_indices"] = Parameter(torch.as_tensor([[]], dtype=torch.int64), requires_grad=False)
         self.is_reference_info_empty = False
         
     def expand_reference_info(self, new_largest_label: int) -> None:
@@ -328,16 +328,20 @@ def learn(
                 default_threshold_reference -= 0.05
 
             self.reference_info["reference_feats"][int(label.id)] = ref_feat.detach().cpu()
-            self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([int(label.id)]))), requires_grad=False)
+            self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([[int(label.id)]])), dim=1), requires_grad=False)
             ref_masks[int(label.id)] = ref_mask.detach().cpu()
             
-        self.reference_info["used_indices"] = Parameter(self.reference_info["used_indices"].unsqueeze(0), requires_grad=False)
         if return_outputs:
             return self.reference_info, ref_masks
 
     @torch.no_grad()
     def infer(
-        self, images: Tensor, original_size: Tensor
+        self,
+        images: Tensor,
+        reference_feats: Tensor,
+        used_indices: Tensor,
+        original_size: Tensor,
+        is_cascade: bool = False,
     ) -> List[List[DefaultDict[int, List[Tensor]]]]:
         """Zero-shot inference with reference features.
 
@@ -345,7 +349,10 @@ def infer(
 
         Args:
             images (Tensor): Given images for target results.
+            reference_feats (Tensor): Reference features for target prediction.
+            used_indices (Tensor): To check which indices of reference features are validate.
             original_size (Tensor): Original image size.
+            is_cascade (bool): Whether use cascade inference. Defaults to False.
 
         Returns:
             (List[List[DefaultDict[int, List[Tensor]]]]): Target results.
@@ -363,7 +370,10 @@ def infer(
             image_embedding = self.image_encoder(images)
 
             total_points_scores, total_bg_coords = self.prompt_getter(
-                image_embedding=image_embedding, original_size=original_size
+                image_embedding=image_embedding,
+                reference_feats=reference_feats,
+                used_indices=used_indices,
+                original_size=original_size
             )
             predicted_masks: defaultdict = defaultdict(list)
             used_points: defaultdict = defaultdict(list)
@@ -382,9 +392,7 @@ def infer(
                         continue
 
                     point_coords = torch.cat((points_score[:2].unsqueeze(0), bg_coords), dim=0).unsqueeze(0)
-                    point_coords = ResizeLongestSide.apply_coords(
-                        point_coords, original_size[0], self.config.model.image_size
-                    )
+                    point_coords = self._preprocess_coords(point_coords, original_size[0], self.config.model.image_size)
                     point_labels = torch.as_tensor(
                         [1] + [0] * len(bg_coords), dtype=torch.float32, device=self.device
                     ).unsqueeze(0)
@@ -393,22 +401,23 @@ def infer(
                         point_coords=point_coords,
                         point_labels=point_labels,
                         original_size=original_size,
+                        is_cascade=is_cascade,
                     )
                     predicted_masks[label].append((mask * points_score[2]).detach().cpu())
                     used_points[label].append(points_score.detach().cpu())
 
             # check overlapping area between different label masks
-            self.__inspect_overlapping_areas(predicted_masks, used_points)
+            self._inspect_overlapping_areas(predicted_masks, used_points)
             total_results.append([predicted_masks, used_points])
         return total_results
 
-    def __inspect_overlapping_areas(
+    def _inspect_overlapping_areas(
         self,
         predicted_masks: Dict[int, List[Tensor]],
         used_points: Dict[int, List[Tensor]],
         threshold_iou: float = 0.8,
     ):
-        def __calculate_mask_iou(mask1: Tensor, mask2: Tensor):
+        def _calculate_mask_iou(mask1: Tensor, mask2: Tensor):
             assert mask1.ndim == 2 and mask2.ndim == 2
             intersection = torch.logical_and(mask1, mask2).sum().item()
             union = torch.logical_or(mask1, mask2).sum().item()
@@ -426,17 +435,17 @@ def __calculate_mask_iou(mask1: Tensor, mask2: Tensor):
             overlapped_label = []
             overlapped_other_label = []
             for (im, mask), (jm, other_mask) in product(enumerate(masks), enumerate(other_masks)):
-                if __calculate_mask_iou(mask, other_mask) > threshold_iou:
+                if _calculate_mask_iou(mask, other_mask) > threshold_iou:
                     if used_points[label][im][2] > used_points[other_label][jm][2]:
                         overlapped_other_label.append(jm)
                     else:
                         overlapped_label.append(im)
 
-            for im in overlapped_label[::-1]:
+            for im in sorted(list(set(overlapped_label)), reverse=True):
                 masks.pop(im)
                 used_points[label].pop(im)
 
-            for jm in overlapped_other_label[::-1]:
+            for jm in sorted(list(set(overlapped_other_label)), reverse=True):
                 other_masks.pop(jm)
                 used_points[other_label].pop(jm)
 
@@ -514,7 +523,11 @@ def training_step(self, batch, batch_idx) -> None:
 
     def predict_step(self, batch, batch_idx):
         """Predict step for `infer`."""
-        results = self.infer(images=batch["images"], original_size=batch.get("original_size")[0].unsqueeze(0))
+        results = self.infer(
+            images=batch["images"],
+            reference_feats=self.reference_info["reference_feats"],
+            used_indices=self.reference_info["used_indices"],
+            original_size=batch.get("original_size")[0].unsqueeze(0))
         return [result[0] for result in results]  # tmp: only mask
 
     def _preprocess_prompts(
@@ -554,6 +567,30 @@ def _preprocess_prompts(
 
         processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x[0].id_))  # type: ignore[assignment]
         return processed_prompts
+    
+    def _preprocess_coords(
+        self,
+        coords: Tensor,
+        ori_shape: Union[list[int], tuple[int, int], Tensor],
+        target_length: int,
+    ) -> Tensor:
+        """Expects a torch tensor of length 2 in the final dimension.
+
+        Requires the original image size in (H, W) format.
+
+        Args:
+            coords (Tensor): Coordinates tensor.
+            ori_shape (Union[list[int], tuple[int, int], Tensor]): Original size of image.
+            target_length (int): The length of the longest side of the image.
+
+        Returns:
+            (Tensor): Resized coordinates.
+        """
+        old_h, old_w = ori_shape
+        new_h, new_w = self.get_prepadded_size(ori_shape, target_length)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
 
     def _generate_masked_features(
         self,
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index a3478ccb59a..58fa1554ed7 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -124,17 +124,17 @@ def zero_shot_segment_anything(state_dict: Optional[OrderedDict] = None):
         [
             None,
             {
-                "reference_info": nn.ParameterDict({
-                    "reference_feats": "reference_feats",
-                    "used_indices": "used_indices"})
+                "reference_info.reference_feats": torch.zeros(1),
+                "reference_info.used_indices": torch.zeros(1, dtype=torch.int64),
             },
         ],
     )
     def test_init(self, set_zero_shot_segment_anything, state_dict: Optional[Dict[str, Any]]) -> None:
         """Test __init__."""
         if state_dict is not None:
-            zero_shot_segment_anything_for_init_weights = set_zero_shot_segment_anything()
-            state_dict.update(zero_shot_segment_anything_for_init_weights.state_dict())
+            zero_shot_segment_anything_for_init_weights = set_zero_shot_segment_anything().state_dict()
+            zero_shot_segment_anything_for_init_weights.update(state_dict)
+            state_dict = zero_shot_segment_anything_for_init_weights
         
         zero_shot_segment_anything = set_zero_shot_segment_anything(state_dict=state_dict)
 
@@ -143,8 +143,11 @@ def test_init(self, set_zero_shot_segment_anything, state_dict: Optional[Dict[st
         assert zero_shot_segment_anything.config.model.freeze_mask_decoder
 
         if state_dict:
-            zero_shot_segment_anything.reference_info.reference_feats = "reference_feats"
-            zero_shot_segment_anything.reference_info.used_indices = "used_indices"
+            assert zero_shot_segment_anything.reference_info.reference_feats == torch.zeros(1)
+            assert zero_shot_segment_anything.reference_info.used_indices == torch.zeros(1, dtype=torch.int64)
+
+        assert zero_shot_segment_anything.reference_info.reference_feats.dtype == torch.float32
+        assert zero_shot_segment_anything.reference_info.used_indices.dtype == torch.int64
 
     @e2e_pytest_unit
     def test_set_default_config(self, set_zero_shot_segment_anything) -> None:
@@ -196,19 +199,124 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
         )
 
         zero_shot_segment_anything = set_zero_shot_segment_anything()
-        zero_shot_segment_anything.reference_info.reference_feats = torch.rand(1, 1, 256)
-        zero_shot_segment_anything.reference_info.used_indices = {0}
+        reference_feats = nn.Parameter(torch.rand(1, 1, 256), requires_grad=False)
+        used_indices = nn.Parameter(torch.as_tensor([[0]], dtype=torch.int64), requires_grad=False)
         mocker.patch.object(
             SegmentAnything, "forward", return_value=(torch.ones(1, 4, 4, 4), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
         )
 
         total_results = zero_shot_segment_anything.infer(
-            images=torch.ones((1, 3, 4, 4)), original_size=torch.tensor([[4, 4]], dtype=torch.int64)
+            images=torch.ones((1, 3, 4, 4)), 
+            reference_feats=reference_feats,
+            used_indices=used_indices,
+            original_size=torch.tensor([[4, 4]], dtype=torch.int64)
         )
 
         for i, results in enumerate(total_results[0]):
             for _, result in results.items():
                 assert torch.equal(result[0], expected[i])
+                
+    @e2e_pytest_unit
+    def test_inspect_overlapping_areas(self, mocker, set_zero_shot_segment_anything) -> None:
+        """Test _inspect_overlapping_areas."""
+        mocker.patch("otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_checkpoint")
+        zero_shot_segment_anything = set_zero_shot_segment_anything()
+        predicted_masks = {
+            0: [
+                torch.tensor(
+                    [
+                        [1, 1, 0, 0, 0, 0],
+                        [1, 1, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+                torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 1, 1, 1, 0],
+                        [0, 0, 1, 1, 1, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+                torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 1, 1, 1, 0, 0],
+                        [0, 1, 1, 1, 0, 0],
+                    ],
+                ),
+            ],
+            1: [
+                torch.tensor(
+                    [
+                        [0, 0, 0, 1, 1, 0],
+                        [0, 0, 0, 1, 1, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+                torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 1, 1],
+                        [0, 0, 0, 0, 1, 1],
+                    ],
+                ),
+                torch.tensor(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 1, 1, 0, 0, 0],
+                        [0, 1, 1, 0, 0, 0],
+                    ],
+                ),
+                torch.tensor(
+                    [
+                        [1, 1, 0, 0, 0, 0],
+                        [1, 1, 0, 0, 0, 0],
+                        [1, 1, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+            ],
+        }
+        used_points = {
+            0: [
+                torch.tensor([0, 0, 0.5]),  # to be removed
+                torch.tensor([2, 2, 0.5]),
+                torch.tensor([1, 4, 0.5]),
+            ],
+            1: [
+                torch.tensor([3, 0, 0.5]),
+                torch.tensor([4, 4, 0.5]),
+                torch.tensor([1, 4, 0.3]),  # to be removed
+                torch.tensor([0, 0, 0.7]),
+            ],
+        }
+
+        zero_shot_segment_anything._inspect_overlapping_areas(predicted_masks, used_points, threshold_iou=0.5)
+
+        assert len(predicted_masks[0]) == 2
+        assert len(predicted_masks[1]) == 3
+        assert all(torch.tensor([2, 2, 0.5]) == used_points[0][0])
+        assert all(torch.tensor([0, 0, 0.7]) == used_points[1][2])
 
     @e2e_pytest_unit
     def test_predict_masks(self, mocker, set_zero_shot_segment_anything) -> None:

From a6add5d58c65578d7a35cff088f0cf4ae639615c Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 13 Feb 2024 16:35:19 +0900
Subject: [PATCH 07/28] [WIP] Update for export

---
 .../model_wrappers/openvino_models.py         | 15 ++++++++++++-
 .../zero_shot_segment_anything.py             | 21 ++++++++++---------
 .../visual_prompting/tasks/inference.py       |  2 +-
 3 files changed, 26 insertions(+), 12 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 432d0010e09..2ed3bcec381 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -17,7 +17,6 @@
 from copy import deepcopy
 from typing import Any, Dict, List, Optional, Tuple, Union
 
-import cv2
 import numpy as np
 from openvino.model_api.adapters.inference_adapter import InferenceAdapter
 from openvino.model_api.models import ImageModel, SegmentationModel
@@ -71,6 +70,20 @@ def parameters(cls) -> Dict[str, Any]:  # noqa: D102
         parameters.update({"sim_threshold": NumericalValue(value_type=float, default_value=0.5, min=0, max=1)})
         parameters.update({"num_bg_points": NumericalValue(value_type=int, default_value=1, min=0, max=1024)})
         return parameters
+    
+    def _get_inputs(self):
+        """Defines the model inputs for images and additional info."""
+        image_blob_names, image_info_blob_names = [], []
+        for name, metadata in self.inputs.items():
+            if len(metadata.shape) == 4:
+                image_blob_names.append(name)
+            else:
+                image_info_blob_names.append(name)
+        if not image_blob_names:
+            self.raise_error(
+                "Failed to identify the input for the image: no 4D input layer found"
+            )
+        return image_blob_names, image_info_blob_names
 
 
 class Decoder(SegmentationModel):
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 8330b312b22..d85a33b5359 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -4,6 +4,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 from collections import OrderedDict, defaultdict
+import os
 from copy import deepcopy
 from itertools import product
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Union
@@ -54,7 +55,8 @@ def forward(
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates."""
         device = image_embedding.device
-        threshold = threshold.to(device)
+        threshold = threshold.squeeze().to(device)
+        num_bg_points = num_bg_points.squeeze()
 
         total_points_scores: Tensor = torch.zeros(used_indices.max() + 1, 0, 3, device=device)
         total_bg_coords: Tensor = torch.zeros(used_indices.max() + 1, num_bg_points, 2, device=device)
@@ -85,12 +87,12 @@ def get_prompt_candidates(
         image_embedding: Tensor,
         reference_feat: Tensor,
         original_size: Tensor,
-        threshold: Tensor = torch.as_tensor([[0.0]], dtype=torch.float32),
-        num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
+        threshold: Union[Tensor, float] = 0.,
+        num_bg_points: Union[Tensor, int] = 1,
         device: Union[torch.device, str] = torch.device("cpu"),
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates from given reference and target features."""
-        assert original_size.dim() == 2 and threshold.dim() == 2 and num_bg_points.dim() == 2
+        assert original_size.dim() == 2
 
         target_feat = image_embedding.squeeze()
         c_feat, h_feat, w_feat = target_feat.shape
@@ -115,14 +117,14 @@ def _point_selection(
         self,
         mask_sim: Tensor,
         original_size: Tensor,
-        threshold: Tensor,
-        num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
+        threshold: Union[Tensor, float] = 0.,
+        num_bg_points: Union[Tensor, int] = 1,
     ) -> Tuple[Tensor, Tensor]:
         """Select point used as point prompts."""
         _, w_sim = mask_sim.shape
 
         # Top-last point selection
-        bg_indices = mask_sim.flatten().topk(num_bg_points[0, 0], largest=False)[1]
+        bg_indices = mask_sim.flatten().topk(num_bg_points, largest=False)[1]
         bg_x = (bg_indices // w_sim).unsqueeze(0)
         bg_y = bg_indices - bg_x * w_sim
         bg_coords = torch.cat((bg_y, bg_x), dim=0).permute(1, 0)
@@ -483,11 +485,10 @@ def _predict_masks(
                 has_mask_input = self.has_mask_inputs[1].to(self.device)
                 coords = torch.nonzero(masks)
                 y, x = coords[:, 0], coords[:, 1]
-                box_coords = ResizeLongestSide.apply_coords(
+                box_coords = self._preprocess_coords(
                     torch.as_tensor([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device),
                     original_size[0],
-                    self.config.model.image_size,
-                )
+                    self.config.model.image_size)
                 point_coords = torch.cat((point_coords, box_coords), dim=1)
                 point_labels = torch.cat((point_labels, self.point_labels_box.to(self.device)), dim=1)
 
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index aca6039d678..5e8e6945701 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -646,7 +646,7 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "reference_feats": torch.randn(1, 1, 256, dtype=torch.float32),
                     "used_indices": torch.as_tensor([[0]], dtype=torch.int64),
                     "original_size": torch.randint(low=0, high=image_size * 2, size=(1, 2), dtype=torch.int64),
-                    "threshold": torch.tensor([[0.1]], dtype=torch.float32),
+                    "threshold": torch.as_tensor([[0.1]], dtype=torch.float32),
                     "num_bg_points": torch.randint(low=1, high=image_size, size=(1, 1), dtype=torch.int64),
                 }
                 output_names = ["total_points_scores", "total_bg_coords"]

From df28d1a247b0c8d205b71609f0f9cb45c0600568 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 15 Feb 2024 13:19:21 +0900
Subject: [PATCH 08/28] Update data & forward pipeline for multiple prompts

---
 .../model_wrappers/openvino_models.py         |  38 +++---
 .../pytorch_lightning/callbacks/inference.py  |   2 +-
 .../pytorch_lightning/datasets/dataset.py     | 129 +++++++++++-------
 .../datasets/pipelines/sam_transforms.py      |   7 +-
 .../datasets/pipelines/transforms.py          |   6 +-
 .../visual_prompters/segment_anything.py      |  39 +++---
 .../configs/sam_tiny_vit/config.yaml          |   2 +
 .../visual_prompting/tasks/openvino.py        |   8 +-
 .../datasets/pipelines/test_sam_transforms.py |   4 -
 .../datasets/pipelines/test_transforms.py     | 116 +++++++++++-----
 .../datasets/test_dataset.py                  |  20 +--
 .../visual_prompters/test_segment_anything.py |   8 +-
 12 files changed, 229 insertions(+), 150 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 2ed3bcec381..128e92687e4 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -98,6 +98,9 @@ def __init__(
         preload: bool = False,
     ):
         super().__init__(model_adapter, configuration, preload)
+        
+        self.mask_input = np.zeros((1, 1, 256, 256), dtype=np.float32)
+        self.has_mask_input = np.zeros((1, 1), dtype=np.float32)
 
     @classmethod
     def parameters(cls):  # noqa: D102
@@ -112,22 +115,25 @@ def _get_outputs(self):
     def preprocess(self, inputs: Dict[str, Any], meta: Dict[str, Any]) -> List[Dict[str, Any]]:
         """Preprocess prompts."""
         processed_prompts = []
-        # TODO (sungchul): process points
-        for bbox, label in zip(inputs["bboxes"], inputs["labels"]):
-            # TODO (sungchul): add condition to check whether using bbox or point
-            point_coords = self._apply_coords(bbox.reshape(-1, 2, 2), inputs["original_size"])
-            point_labels = np.array([2, 3], dtype=np.float32).reshape((-1, 2))
-            processed_prompts.append(
-                {
-                    "point_coords": point_coords,
-                    "point_labels": point_labels,
-                    # TODO (sungchul): how to generate mask_input and has_mask_input
-                    "mask_input": np.zeros((1, 1, 256, 256), dtype=np.float32),
-                    "has_mask_input": np.zeros((1, 1), dtype=np.float32),
-                    "orig_size": np.array(inputs["original_size"], dtype=np.int64).reshape((-1, 2)),
-                    "label": label,
-                }
-            )
+        for prompt_name in ["bboxes", "points"]:
+            for prompt, label in zip(inputs.get(prompt_name), inputs["labels"].get(prompt_name, [])):
+                if prompt_name == "bboxes":
+                    point_coords = self._apply_coords(prompt.reshape(-1, 2, 2), inputs["original_size"])
+                    point_labels = np.array([2, 3], dtype=np.float32).reshape(-1, 2)
+                else:
+                    point_coords = self._apply_coords(prompt.reshape(-1, 1, 2), inputs["original_size"])
+                    point_labels = np.array([1], dtype=np.float32).reshape(-1, 1)
+
+                processed_prompts.append(
+                    {
+                        "point_coords": point_coords,
+                        "point_labels": point_labels,
+                        "mask_input": self.mask_input,
+                        "has_mask_input": self.has_mask_input,
+                        "orig_size": np.asarray(inputs["original_size"], dtype=np.int64).reshape(-1, 2),
+                        "label": label,
+                    }
+                )
         return processed_prompts
 
     def _apply_coords(self, coords: np.ndarray, original_size: Union[List[int], Tuple[int, int]]) -> np.ndarray:
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py
index 1dc39b7cc3f..16960cb32ad 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py
@@ -57,7 +57,7 @@ def on_predict_epoch_end(self, _trainer: Trainer, _pl_module: LightningModule, o
         for output in outputs[0]:
             pred_masks.append(output["masks"][0])
             iou_predictions.append(output["iou_predictions"][0])
-            pred_labels.append(output["labels"][0])
+            pred_labels.append(output["labels"][0]["bboxes"] + output["labels"][0]["points"])
 
         for dataset_item, pred_mask, iou_prediction, labels in zip(
             self.otx_dataset, pred_masks, iou_predictions, pred_labels
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
index 476a2c09d69..82dead10c87 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
@@ -31,6 +31,7 @@
     ResizeLongestSide,
     collate_fn,
 )
+from collections import defaultdict
 from otx.api.entities.dataset_item import DatasetItemEntity
 from otx.api.entities.datasets import DatasetEntity
 from otx.api.entities.image import Image
@@ -130,13 +131,6 @@ def generate_bbox_from_mask(gt_mask: np.ndarray, width: int, height: int) -> Lis
     return generate_bbox(x_min, y_min, x_max, y_max, width, height)
 
 
-def generate_point_from_mask(gt_mask: np.ndarray) -> np.ndarray:
-    """Randomly generate point from given mask."""
-    candidates = np.where(gt_mask == 1)
-    index = np.random.permutation(len(candidates))[0]
-    return candidates[index]
-
-
 class OTXVisualPromptingDataset(Dataset):
     """Visual Prompting Dataset Adaptor.
 
@@ -149,13 +143,32 @@ class OTXVisualPromptingDataset(Dataset):
     """
 
     def __init__(
-        self, dataset: DatasetEntity, image_size: int, mean: List[float], std: List[float], offset_bbox: int = 0
+        self,
+        mode: Subset,
+        dataset: DatasetEntity,
+        image_size: int,
+        mean: List[float],
+        std: List[float],
+        offset_bbox: int = 0,
+        use_point: bool = False,
+        use_bbox: bool = False,
     ) -> None:
-
+        self.mode = mode
         self.dataset = dataset
         self.transform = get_transform(image_size, mean, std)
         self.offset_bbox = offset_bbox
         self.labels = dataset.get_labels()
+        
+        if not use_bbox and not use_point:
+            # if both are False, use bbox as default
+            use_bbox = True
+        self.prob = 1.0  # if using only bbox prompt
+        if use_bbox and use_point:
+            # if using both prompts, divide prob into both
+            self.prob = 0.5
+        if not use_bbox and use_point:
+            # if using only point prompt
+            self.prob = 0.0
 
     def __len__(self) -> int:
         """Get size of the dataset.
@@ -166,21 +179,23 @@ def __len__(self) -> int:
         return len(self.dataset)
 
     @staticmethod
-    def get_prompts(dataset_item: DatasetItemEntity, dataset_labels: List[LabelEntity]) -> Dict[str, Any]:
+    def get_prompts(dataset_item: DatasetItemEntity, dataset_labels: List[LabelEntity], prob: float = 1., mode: Subset = Subset.TESTING) -> Dict[str, Any]:
         """Get propmts from dataset_item.
 
         Args:
             dataset_item (DatasetItemEntity): Dataset item entity.
             dataset_labels (List[LabelEntity]): Label information.
+            prob (float): Probability of which prompts will be generated.
+            mode (Subset): To check which mode is used between training, validation, and testing.
 
         Returns:
             Dict[str, Any]: Processed prompts with ground truths.
         """
         width, height = dataset_item.width, dataset_item.height
-        bboxes: List[List[int]] = []
-        points: List = []  # TBD
+        bboxes: List[np.ndarray] = []
+        points: List[np.ndarray] = []
         gt_masks: List[np.ndarray] = []
-        labels: List[ScoredLabel] = []
+        labels: defaultdict[str, List[ScoredLabel]] = defaultdict(list)
         for annotation in dataset_item.get_annotations(labels=dataset_labels, include_empty=False, preserve_id=True):
             if isinstance(annotation.shape, Image):
                 # use mask as-is
@@ -192,25 +207,36 @@ def get_prompts(dataset_item: DatasetItemEntity, dataset_labels: List[LabelEntit
                 continue
 
             if gt_mask.sum() == 0:
-                # pass no gt
+                # pass no gt or very small region
                 continue
-            gt_masks.append(gt_mask)
-
-            # generate bbox based on gt_mask
-            bbox = generate_bbox_from_mask(gt_mask, width, height)
-            bboxes.append(bbox)
 
-            # TODO (sungchul): generate random points from gt_mask
-
-            # add labels
-            labels.extend(annotation.get_labels(include_empty=False))
+            gt_masks.append(gt_mask)
 
-        bboxes = np.array(bboxes)
+            mask_points = np.nonzero(gt_mask)
+            if np.random.rand() < prob:
+                # generate bbox based on gt_mask
+                bbox = generate_bbox_from_mask(gt_mask, width, height)
+                bboxes.append(bbox)
+                labels["bboxes"].extend(annotation.get_labels(include_empty=False))
+            else:
+                # generate point based on gt_mask
+                if mode == Subset.TRAINING:
+                    # get random point from the mask
+                    idx_chosen = np.random.permutation(len(mask_points[0]))[0]  # noqa: NPY002
+                    point = np.array([mask_points[1][idx_chosen], mask_points[0][idx_chosen]])
+                else:
+                    # get averaged point
+                    point = np.array([mask_points[1].mean(), mask_points[0].mean()])
+                points.append(point)
+                labels["points"].extend(annotation.get_labels(include_empty=False))
+
+        bboxes = np.array(bboxes, dtype=np.float32) if len(bboxes) > 0 else np.zeros((0, 4), dtype=np.float32)
+        points = np.array(points, dtype=np.float32) if len(points) > 0 else np.zeros((0, 2), dtype=np.float32)
         return dict(
             original_size=np.array((height, width), dtype=np.int64),
             gt_masks=gt_masks,
             bboxes=bboxes,
-            points=points,  # TODO (sungchul): update point information
+            points=points,
             labels=labels,
         )
 
@@ -226,7 +252,7 @@ def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
         dataset_item = self.dataset[index]
         item: Dict[str, Union[int, Tensor]] = {"index": index, "images": dataset_item.numpy}
 
-        prompts = self.get_prompts(dataset_item, self.labels)
+        prompts = self.get_prompts(dataset_item, self.labels, self.prob, self.mode)
         if len(prompts["gt_masks"]) == 0:
             return {
                 "images": [],
@@ -238,7 +264,6 @@ def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
                 "labels": [],
             }
 
-        prompts["bboxes"] = np.array(prompts["bboxes"])
         item.update({**prompts, "path": dataset_item.media.path})
         item = self.transform(item)
         return item
@@ -247,20 +272,6 @@ def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
 class OTXZeroShotVisualPromptingDataset(OTXVisualPromptingDataset):
     """Visual Prompting for Zero-shot learning Dataset Adaptor."""
 
-    def __init__(
-        self,
-        dataset: DatasetEntity,
-        image_size: int,
-        mean: List[float],
-        std: List[float],
-        generate_point: bool = False,
-        generate_bbox: bool = False,
-        **kwargs,
-    ) -> None:
-        super().__init__(dataset, image_size, mean, std, offset_bbox=0)
-        self.generate_point = generate_point
-        self.generate_bbox = generate_bbox
-
     def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
         """Get dataset item.
 
@@ -273,7 +284,7 @@ def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
         dataset_item = self.dataset[index]
         item: Dict[str, Union[int, Tensor]] = {"index": index, "images": dataset_item.numpy}
 
-        prompts = self.get_prompts(dataset_item, self.labels)  # , self.generate_point, self.generate_bbox)
+        prompts = self.get_prompts(dataset_item, self.labels, self.prob)
         item.update({**prompts, "path": dataset_item.media.path})
         item = self.transform(item)
         return item
@@ -314,12 +325,12 @@ def __init__(
                 )
                 self.config["train_batch_size"] = 1
 
-            self.kwargs.update(
-                {
-                    "generate_point": self.config.get("generate_point", False),
-                    "generate_bbox": self.config.get("generate_bbox", False),
-                }
-            )
+        self.kwargs.update(
+            {
+                "use_point": self.config.get("use_point", False),
+                "use_bbox": self.config.get("use_bbox", False),
+            }
+        )
 
         self.train_otx_dataset: DatasetEntity
         self.val_otx_dataset: DatasetEntity
@@ -340,6 +351,7 @@ def setup(self, stage: Optional[str] = None) -> None:
         std = self.config.normalize.std
         if stage == "fit" or stage is None:
             self.train_dataset = self.DATASETS[self.train_type](
+                mode=Subset.TRAINING,
                 dataset=self.dataset.get_subset(Subset.TRAINING),
                 image_size=image_size,
                 mean=mean,
@@ -351,17 +363,32 @@ def setup(self, stage: Optional[str] = None) -> None:
             # self.val_dataset = None
             if self.train_type == TrainType.Incremental:
                 self.val_dataset = self.DATASETS[self.train_type](
-                    dataset=self.dataset.get_subset(Subset.VALIDATION), image_size=image_size, mean=mean, std=std
+                    mode=Subset.VALIDATION,
+                    dataset=self.dataset.get_subset(Subset.VALIDATION),
+                    image_size=image_size,
+                    mean=mean,
+                    std=std,
+                    **self.kwargs,
                 )
 
         if stage == "test":
             self.test_dataset = self.DATASETS[self.train_type](
-                dataset=self.dataset.get_subset(Subset.TESTING), image_size=image_size, mean=mean, std=std
+                mode=Subset.TESTING,
+                dataset=self.dataset.get_subset(Subset.TESTING),
+                image_size=image_size,
+                mean=mean,
+                std=std,
+                **self.kwargs,
             )
 
         if stage == "predict":
             self.predict_dataset = self.DATASETS[self.train_type](
-                dataset=self.dataset, image_size=image_size, mean=mean, std=std, **self.kwargs
+                mode=Subset.TESTING,
+                dataset=self.dataset.get_subset(Subset.TESTING),
+                image_size=image_size,
+                mean=mean,
+                std=std,
+                **self.kwargs
             )
 
     def summary(self):
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
index 06d04ea817d..33c1fbb9b59 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
@@ -38,8 +38,7 @@ def __call__(self, item: Dict[str, Union[List, Tensor]]) -> Dict[str, Union[List
         )
         item["gt_masks"] = [torch.as_tensor(gt_mask) for gt_mask in item["gt_masks"]]
         item["bboxes"] = self.apply_boxes(item["bboxes"], item["original_size"], self.target_length)
-        if item["points"]:
-            item["points"] = self.apply_coords(item["points"], item["original_size"], self.target_length)
+        item["points"] = self.apply_coords(item["points"], item["original_size"], self.target_length)
         return item
 
     @classmethod
@@ -78,9 +77,9 @@ def apply_coords(
         old_h, old_w = original_size
         new_h, new_w = cls.get_preprocess_shape(original_size[0], original_size[1], target_length)
         if isinstance(coords, np.ndarray):
-            coords = coords.astype(float)
+            coords = coords.astype(np.float32)
         else:
-            coords = coords.to(torch.float)
+            coords = coords.to(torch.float32)
         coords[..., 0] = coords[..., 0] * (new_w / old_w)
         coords[..., 1] = coords[..., 1] * (new_h / old_h)
         return coords
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py
index 05b0693a39a..dd1abddf740 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/transforms.py
@@ -34,13 +34,13 @@ def _convert_empty_to_none(x: str, dtype: torch.dtype = torch.float32) -> List:
             List: List of batch data.
         """
         func = torch.stack if x == "gt_masks" else torch.tensor
-        items = [func(item[x]).to(dtype) for item in batch if item[x] is not None]
-        return None if len(items) == 0 else items
+        items = [func(item[x]).to(dtype) if len(item[x]) > 0 else None for item in batch]
+        return items
 
     index = [item["index"] for item in batch]
     images = torch.stack([item["images"] for item in batch])
     bboxes = _convert_empty_to_none("bboxes")
-    points = None  # TBD
+    points = _convert_empty_to_none("points")
     gt_masks = _convert_empty_to_none("gt_masks", torch.int32)
     original_size = _convert_empty_to_none("original_size")
     path = [item["path"] for item in batch]
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index f6b59922baf..9ac772ce4ad 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -386,23 +386,30 @@ def forward_train(
         image_embeddings = self.image_encoder(images)
         pred_masks = []
         ious = []
-        for embedding, bbox in zip(image_embeddings, bboxes):
-            sparse_embeddings, dense_embeddings = self.prompt_encoder(
-                points=points,
-                boxes=bbox,
-                masks=masks,
-            )
-
-            low_res_masks, iou_predictions = self.mask_decoder(
-                image_embeddings=embedding.unsqueeze(0),
-                image_pe=self.prompt_encoder.get_dense_pe(),
-                sparse_prompt_embeddings=sparse_embeddings,
-                dense_prompt_embeddings=dense_embeddings,
-                multimask_output=False,  # when given multiple prompts. if there is single prompt True would be better.
-            )
+        for idx, embedding in enumerate(image_embeddings):
+            low_res_masks, iou_predictions = [], []
+            for idx_prompt, prompt in enumerate([bboxes[idx], points[idx]]):
+                if prompt is None:
+                    continue
+                
+                sparse_embeddings, dense_embeddings = self.prompt_encoder(
+                    points=(prompt.unsqueeze(1), torch.ones(len(prompt), 1, device=prompt.device)) if idx_prompt == 1 else None,
+                    boxes=prompt if idx_prompt == 0 else None,
+                    masks=None,
+                )
+                
+                _low_res_masks, _iou_predictions = self.mask_decoder(
+                    image_embeddings=embedding.unsqueeze(0),
+                    image_pe=self.prompt_encoder.get_dense_pe(),
+                    sparse_prompt_embeddings=sparse_embeddings,
+                    dense_prompt_embeddings=dense_embeddings,
+                    multimask_output=False,  # when given multiple prompts. if there is single prompt True would be better. # noqa: E501
+                )
+                low_res_masks.append(_low_res_masks)
+                iou_predictions.append(_iou_predictions)
 
-            pred_masks.append(low_res_masks)
-            ious.append(iou_predictions)
+            pred_masks.append(torch.cat(low_res_masks, dim=0))
+            ious.append(torch.cat(iou_predictions, dim=0))
 
         return pred_masks, ious
 
diff --git a/src/otx/algorithms/visual_prompting/configs/sam_tiny_vit/config.yaml b/src/otx/algorithms/visual_prompting/configs/sam_tiny_vit/config.yaml
index 36e9748338b..f0dd50ca827 100644
--- a/src/otx/algorithms/visual_prompting/configs/sam_tiny_vit/config.yaml
+++ b/src/otx/algorithms/visual_prompting/configs/sam_tiny_vit/config.yaml
@@ -15,6 +15,8 @@ dataset:
       - 57.12
       - 57.375
   offset_bbox: 20 # randomness for generating bounding box, pixel
+  use_point: false
+  use_bbox: false
 
 model:
   name: SAM
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 7751713aa47..d550f1fb33d 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -149,11 +149,15 @@ def __init__(
         self.transform = get_transform()  # TODO (sungchul): insert args
 
     def pre_process(
-        self, dataset_item: DatasetItemEntity, extra_processing: bool = False
+        self, dataset_item: DatasetItemEntity, extra_processing: bool = False, use_bbox: bool = False, use_point: bool = False,
     ) -> Tuple[Dict[str, Any], Dict[str, Any], List[Dict[str, Any]]]:
         """Pre-process function of OpenVINO Visual Prompting Inferencer for image encoder."""
+        if use_bbox and use_point:
+            logger.warning("If both use_bbox and use_point are set, bboxes and points will be generated randomly.")
+            
+        prob = 1. if not use_point else 0. if not use_bbox and use_point else 0.5
         images, meta = self.model["image_encoder"].preprocess(dataset_item.numpy, extra_processing)
-        prompts = OTXVisualPromptingDataset.get_prompts(dataset_item, self.labels)  # to be replaced
+        prompts = OTXVisualPromptingDataset.get_prompts(dataset_item, self.labels, prob=prob)
         prompts = self.model["decoder"].preprocess(prompts, meta)
         return images, meta, prompts  # type: ignore
 
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_sam_transforms.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_sam_transforms.py
index 2636627d2b9..dca279d5175 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_sam_transforms.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_sam_transforms.py
@@ -20,10 +20,6 @@ class TestResizeLongestSide:
     def setup(self):
         self.resize_longest_side = ResizeLongestSide(8)
 
-    @e2e_pytest_unit
-    def test_call(self):
-        """Test __call__."""
-
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "image,expected",
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
index a703c7512c6..c9e32b6f28e 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
@@ -4,7 +4,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-from typing import Any, Dict, Tuple
+from typing import Any, Dict, Tuple, List
 
 import pytest
 import torch
@@ -21,53 +21,97 @@
 
 
 @e2e_pytest_unit
-def test_collate_fn():
-    """Test collate_fn."""
-    batch = [
+@pytest.mark.parametrize("batch,expected",[
+    (
+        [
+            {
+                "index": 0,
+                "images": Tensor([1, 2, 3]),
+                "bboxes": Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]),
+                "points": torch.zeros((0, 2)),
+                "gt_masks": [Tensor([1, 2, 3])],
+                "original_size": Tensor([1, 3]),
+                "path": [],
+                "labels": [],
+            },
+            {
+                "index": 1,
+                "images": Tensor([4, 5, 6]),
+                "bboxes": Tensor([[9, 10, 11, 12]]),
+                "points": torch.zeros((0, 2)),
+                "gt_masks": [Tensor([4, 5, 6])],
+                "original_size": Tensor([1, 3]),
+                "path": [],
+                "labels": [],
+            },
+        ],
         {
-            "index": 0,
-            "images": Tensor([1, 2, 3]),
-            "bboxes": np.array([[1, 2, 3, 4], [5, 6, 7, 8]]),
-            "points": [],
-            "gt_masks": [Tensor([1, 2, 3])],
-            "original_size": np.array([1, 3]),
-            "path": [],
-            "labels": [],
-        },
+            "index": [0, 1],
+            "images": Tensor([[1, 2, 3], [4, 5, 6]]),
+            "bboxes": [Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]), Tensor([[9, 10, 11, 12]])],
+            "points": [None, None],
+            "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
+            "original_size": [Tensor([1, 3]), Tensor([1, 3])],
+            "path": [[], []],
+            "labels": [[], []],
+        }
+    ),
+    (
+        [
+            {
+                "index": 0,
+                "images": Tensor([1, 2, 3]),
+                "bboxes": torch.zeros((0, 4)),
+                "points": Tensor([[1, 1]]),
+                "gt_masks": [Tensor([1, 2, 3])],
+                "original_size": Tensor([1, 3]),
+                "path": [],
+                "labels": [],
+            },
+            {
+                "index": 1,
+                "images": Tensor([4, 5, 6]),
+                "bboxes": torch.zeros((0, 4)),
+                "points": Tensor([[2, 2]]),
+                "gt_masks": [Tensor([4, 5, 6])],
+                "original_size": Tensor([1, 3]),
+                "path": [],
+                "labels": [],
+            },
+        ],
         {
-            "index": 1,
-            "images": Tensor([4, 5, 6]),
-            "bboxes": np.array([[9, 10, 11, 12]]),
-            "points": [],
-            "gt_masks": [Tensor([4, 5, 6])],
-            "original_size": np.array([1, 3]),
-            "path": [],
-            "labels": [],
-        },
-    ]
-    expected = {
-        "index": [0, 1],
-        "images": Tensor([[1, 2, 3], [4, 5, 6]]),
-        "bboxes": [Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]), Tensor([[9, 10, 11, 12]])],
-        "points": None,
-        "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
-        "original_size": [Tensor([1, 3]), Tensor([1, 3])],
-        "path": [[], []],
-        "labels": [[], []],
-    }
-
+            "index": [0, 1],
+            "images": Tensor([[1, 2, 3], [4, 5, 6]]),
+            "bboxes": [None, None],
+            "points": [Tensor([[1, 1]]), Tensor([[2, 2]])],
+            "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
+            "original_size": [Tensor([1, 3]), Tensor([1, 3])],
+            "path": [[], []],
+            "labels": [[], []],
+        }
+    )
+])
+def test_collate_fn(batch: List[Dict[str, Any]], expected: Dict[str, Any]):
+    """Test collate_fn."""
     results = collate_fn(batch)
 
     assert results["index"] == expected["index"]
     assert torch.all(results["images"] == expected["images"])
     for r, e in zip(results["bboxes"], expected["bboxes"]):
-        assert torch.all(r == e)
-    assert results["points"] == expected["points"]
+        if r is not None and e is not None:
+            assert torch.all(r == e)
+
+    for r, e in zip(results["points"], expected["points"]):
+        if r is not None and e is not None:
+            assert torch.all(r == e)
+
     assert len(results["gt_masks"]) == len(expected["gt_masks"])
     for r, e in zip(results["gt_masks"], expected["gt_masks"]):
         assert torch.all(r == e)
+
     for r, e in zip(results["original_size"], expected["original_size"]):
         assert torch.all(r == e)
+
     assert results["path"] == expected["path"]
     assert results["labels"] == expected["labels"]
 
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/test_dataset.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/test_dataset.py
index 99a76c3b17b..6e5211c1899 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/test_dataset.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/test_dataset.py
@@ -19,7 +19,6 @@
     generate_bbox,
     generate_bbox_from_mask,
     get_transform,
-    generate_point_from_mask,
 )
 from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.pipelines import (
     MultipleInputsCompose,
@@ -146,11 +145,6 @@ def test_generate_bbox_from_mask(mocker) -> None:
     assert bbox[3] >= 0 and bbox[3] <= height
 
 
-@e2e_pytest_unit
-def test_generate_point_from_mask() -> None:
-    """TODO"""
-
-
 class TestOTXVIsualPromptingDataset:
     @e2e_pytest_unit
     def test_len(self, mocker, dataset_polygon, transform, image_size, mean, std) -> None:
@@ -159,7 +153,7 @@ def test_len(self, mocker, dataset_polygon, transform, image_size, mean, std) ->
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.dataset.get_transform",
             return_value=transform,
         )
-        otx_dataset = OTXVisualPromptingDataset(dataset_polygon, image_size, mean, std)
+        otx_dataset = OTXVisualPromptingDataset("testing", dataset_polygon, image_size, mean, std)
         assert len(otx_dataset) == 4
 
     @e2e_pytest_unit
@@ -173,7 +167,7 @@ def test_getitem(
             return_value=transform,
         )
         dataset = dataset_mask if use_mask else dataset_polygon
-        otx_dataset = OTXVisualPromptingDataset(dataset, image_size, mean, std)
+        otx_dataset = OTXVisualPromptingDataset("testing", dataset, image_size, mean, std)
 
         item = otx_dataset[0]
 
@@ -189,7 +183,7 @@ def test_getitem(
         assert isinstance(item["gt_masks"], list)
         assert isinstance(item["gt_masks"][0], np.ndarray)
         assert isinstance(item["bboxes"], np.ndarray)
-        assert item["points"] == []
+        assert len(item["points"]) == 0
 
 
 class TestOTXZeroShotVisualPromptingDataset:
@@ -209,7 +203,7 @@ def test_getitem(
             return_value=transform,
         )
         dataset = dataset_mask if use_mask else dataset_polygon
-        otx_dataset = OTXZeroShotVisualPromptingDataset(dataset, image_size, mean, std)
+        otx_dataset = OTXZeroShotVisualPromptingDataset("testing", dataset, image_size, mean, std)
 
         item = otx_dataset[0]
 
@@ -225,7 +219,7 @@ def test_getitem(
         assert isinstance(item["gt_masks"], list)
         assert isinstance(item["gt_masks"][0], np.ndarray)
         assert isinstance(item["bboxes"], np.ndarray)
-        assert item["points"] == []
+        assert len(item["points"]) == 0
 
 
 class TestOTXVisualPromptingDataModule:
@@ -248,8 +242,8 @@ def test_init_zeroshot(self, set_datamodule):
         datamodule = set_datamodule(train_type=TrainType.Zeroshot)
 
         assert datamodule.config.get("train_batch_size") == 1
-        assert "generate_point" in datamodule.kwargs
-        assert "generate_bbox" in datamodule.kwargs
+        assert "use_point" in datamodule.kwargs
+        assert "use_bbox" in datamodule.kwargs
 
     @e2e_pytest_unit
     def test_setup(self, mocker, set_datamodule) -> None:
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index f5dce1d8396..24ae52aa3cb 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -351,7 +351,7 @@ def test_forward_train(self) -> None:
         images = torch.zeros((1, 3, 4, 4))
         bboxes = torch.zeros((1))
 
-        results = sam.forward_train(images=images, bboxes=bboxes, points=None)
+        results = sam.forward_train(images=images, bboxes=bboxes, points=[None])
         pred_masks, ious = results
 
         assert len(bboxes) == len(pred_masks) == len(ious)
@@ -378,7 +378,7 @@ def test_training_step(self, mocker, loss_type: str, expected: Tensor) -> None:
             images=torch.ones((1, 3, 4, 4)),
             gt_masks=[torch.Tensor([[0, 1, 1, 0] for _ in range(4)]).to(torch.int32)],
             bboxes=torch.Tensor([[0, 0, 1, 1]]),
-            points=[],
+            points=[None],
             padding=[[0, 0, 0, 0]],
             original_size=[[4, 4]],
         )
@@ -420,7 +420,7 @@ def test_validation_step(self, mocker) -> None:
             images=torch.ones((1, 3, 4, 4)),
             gt_masks=[torch.Tensor([[0, 1, 1, 0] for _ in range(4)]).to(torch.int32)],
             bboxes=torch.Tensor([[0, 0, 1, 1]]),
-            points=[],
+            points=[None],
             path=None,
             labels=None,
             padding=[0],
@@ -471,7 +471,7 @@ def test_predict_step(self, mocker, return_logits: bool, expected: Tensor) -> No
         batch = dict(
             images=torch.zeros((1, 3, 4, 4)),
             bboxes=torch.Tensor([[0, 0, 1, 1]]),
-            points=[],
+            points=[None],
             path=None,
             labels=None,
             padding=[0],

From b8d89e42a5b9806da15c598a0bb051bcdece176f Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Fri, 16 Feb 2024 00:23:16 +0900
Subject: [PATCH 09/28] (WIP) Update zsl pipeline

---
 .../pytorch_lightning/callbacks/inference.py  |   2 +-
 .../pytorch_lightning/datasets/dataset.py     |  10 +-
 .../datasets/pipelines/sam_transforms.py      |   6 +-
 .../zero_shot_segment_anything.py             | 354 ++++++++++--------
 .../zero_shot_sam_tiny_vit/config.yaml        |   5 +-
 .../visual_prompting/tasks/inference.py       |   2 +
 .../model_wrappers/test_openvino_models.py    |  37 +-
 .../callbacks/test_inference_callback.py      |   6 +-
 .../test_zero_shot_segment_anything.py        |  78 ++--
 .../visual_prompting/tasks/test_inference.py  |   4 +-
 10 files changed, 289 insertions(+), 215 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py
index 16960cb32ad..df751eeaba5 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/inference.py
@@ -57,7 +57,7 @@ def on_predict_epoch_end(self, _trainer: Trainer, _pl_module: LightningModule, o
         for output in outputs[0]:
             pred_masks.append(output["masks"][0])
             iou_predictions.append(output["iou_predictions"][0])
-            pred_labels.append(output["labels"][0]["bboxes"] + output["labels"][0]["points"])
+            pred_labels.append(output["labels"][0].get("bboxes", []) + output["labels"][0].get("points", []))
 
         for dataset_item, pred_mask, iou_prediction, labels in zip(
             self.otx_dataset, pred_masks, iou_predictions, pred_labels
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
index 82dead10c87..83ada16d9c6 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
@@ -286,7 +286,7 @@ def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
 
         prompts = self.get_prompts(dataset_item, self.labels, self.prob)
         item.update({**prompts, "path": dataset_item.media.path})
-        item = self.transform(item)
+            
         return item
 
 
@@ -413,7 +413,7 @@ def train_dataloader(self) -> Union[DataLoader, List[DataLoader], Dict[str, Data
             shuffle=True,
             batch_size=self.config.train_batch_size,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn,
+            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
         )
 
     def val_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
@@ -427,7 +427,7 @@ def val_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
             shuffle=False,
             batch_size=self.config.val_batch_size,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn,
+            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
         )
 
     def test_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
@@ -441,7 +441,7 @@ def test_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
             shuffle=False,
             batch_size=self.config.test_batch_size,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn,
+            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
         )
 
     def predict_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
@@ -455,5 +455,5 @@ def predict_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
             shuffle=False,
             batch_size=1,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn,
+            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
         )
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
index 33c1fbb9b59..a47efc97540 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
@@ -37,8 +37,10 @@ def __call__(self, item: Dict[str, Union[List, Tensor]]) -> Dict[str, Union[List
             self.apply_image(item["images"], self.target_length).transpose((2, 0, 1)), dtype=torch.get_default_dtype()
         )
         item["gt_masks"] = [torch.as_tensor(gt_mask) for gt_mask in item["gt_masks"]]
-        item["bboxes"] = self.apply_boxes(item["bboxes"], item["original_size"], self.target_length)
-        item["points"] = self.apply_coords(item["points"], item["original_size"], self.target_length)
+        if "bboxes" in item:
+            item["bboxes"] = self.apply_boxes(item["bboxes"], item["original_size"], self.target_length)
+        if "points" in item:
+            item["points"] = self.apply_coords(item["points"], item["original_size"], self.target_length)
         return item
 
     @classmethod
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index d85a33b5359..ed902e2c4be 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -5,23 +5,26 @@
 
 from collections import OrderedDict, defaultdict
 import os
+import json
 from copy import deepcopy
 from itertools import product
+import ast
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Union
 
 import torch
+import numpy as np
+import cv2
 from omegaconf import DictConfig
 from torch import nn, Tensor
 from torch.nn import Parameter, ParameterDict
 from torch.nn import functional as F
 
-from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.pipelines import (
-    ResizeLongestSide,
-)
 from otx.api.entities.scored_label import ScoredLabel
 from otx.utils.logger import get_logger
+from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.dataset import get_transform
 
 from .segment_anything import SegmentAnything
+from datetime import datetime
 
 logger = get_logger()
 
@@ -55,6 +58,7 @@ def forward(
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates."""
         device = image_embedding.device
+        original_size = original_size.squeeze()
         threshold = threshold.squeeze().to(device)
         num_bg_points = num_bg_points.squeeze()
 
@@ -92,8 +96,6 @@ def get_prompt_candidates(
         device: Union[torch.device, str] = torch.device("cpu"),
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates from given reference and target features."""
-        assert original_size.dim() == 2
-
         target_feat = image_embedding.squeeze()
         c_feat, h_feat, w_feat = target_feat.shape
         target_feat = target_feat / target_feat.norm(dim=0, keepdim=True)
@@ -101,12 +103,12 @@ def get_prompt_candidates(
 
         sim = reference_feat.to(device) @ target_feat
         sim = sim.reshape(1, 1, h_feat, w_feat)
-        sim = ZeroShotSegmentAnything.postprocess_masks(sim, self.image_size, original_size[0])
+        sim = ZeroShotSegmentAnything.postprocess_masks(sim, self.image_size, original_size)
 
         threshold = (threshold == 0) * self.default_threshold_target + threshold
         points_scores, bg_coords = self._point_selection(
             mask_sim=sim[0, 0],
-            original_size=original_size[0],
+            original_size=original_size,
             threshold=threshold,
             num_bg_points=num_bg_points,
         )
@@ -167,9 +169,13 @@ def _point_selection(
 class ZeroShotSegmentAnything(SegmentAnything):
     """Zero-shot learning module using Segment Anything."""
 
-    def __init__(self, config: Optional[DictConfig] = None, state_dict: Optional[OrderedDict] = None) -> None:
+    def __init__(self, config: Optional[DictConfig] = None, manual_config_update: Optional[Dict] = None, state_dict: Optional[OrderedDict] = None) -> None:
         if config is None:
             config = self.set_default_config()
+            
+        if manual_config_update is not None and isinstance(manual_config_update, dict) and len(manual_config_update) > 0:
+            for k, v in manual_config_update.items():
+                exec(f"config.{k} = {v}")
 
         # check freeze conditions
         for condition in ["freeze_image_encoder", "freeze_prompt_encoder", "freeze_mask_decoder"]:
@@ -177,10 +183,9 @@ def __init__(self, config: Optional[DictConfig] = None, state_dict: Optional[Ord
                 logger.warning(f"config.model.{condition}(=False) must be set to True, changed.")
                 setattr(config.model, condition, True)
 
-        super().__init__(config, None)
+        super().__init__(config, state_dict)
+
         self.set_empty_reference_info()
-        self._register_load_state_dict_pre_hook(self.load_state_dict_pre_hook)
-        self.load_checkpoint(state_dict)
 
         self.prompt_getter = PromptGetter(image_size=config.model.image_size)
         self.prompt_getter.set_default_thresholds(
@@ -191,6 +196,13 @@ def __init__(self, config: Optional[DictConfig] = None, state_dict: Optional[Ord
         self.point_labels_box = torch.as_tensor([[2, 3]], dtype=torch.float32)
         self.has_mask_inputs = [torch.as_tensor([[0.0]]), torch.as_tensor([[1.0]])]
         
+        self.transforms = get_transform(
+            image_size=config.model.image_size,
+            mean=config.dataset.normalize.mean,
+            std=config.dataset.normalize.std)
+        
+        self.path_reference_info = "vpm_zsl_reference_infos/{}/reference_info.pt"
+        
     def load_state_dict_pre_hook(self, state_dict: dict[str, Any], prefix: str = "", *args, **kwargs) -> None:
         """Load reference info manually."""
         _reference_feats: Tensor = state_dict.get("reference_info.reference_feats", torch.as_tensor([], dtype=torch.float32))
@@ -216,6 +228,12 @@ def set_default_config(self) -> DictConfig:
                     "freeze_prompt_encoder": True,
                     "image_size": 1024,
                     "mask_threshold": 0.0,
+                },
+                "dataset": {
+                    "normalize": {
+                        "mean": [123.675, 116.28, 103.53],
+                        "std": [58.395, 57.12, 57.375],
+                    }
                 }
             }
         )
@@ -232,9 +250,9 @@ def set_empty_reference_info(self) -> None:
         )
         self.is_reference_info_empty = True
         
-    def initialize_reference_info(self, largest_label: int) -> None:
+    def initialize_reference_info(self) -> None:
         """Initialize reference information."""
-        self.reference_info["reference_feats"] = Parameter(torch.zeros(largest_label + 1, 1, 256), requires_grad=False)
+        self.reference_info["reference_feats"] = Parameter(torch.zeros(0, 1, 256), requires_grad=False)
         self.reference_info["used_indices"] = Parameter(torch.as_tensor([[]], dtype=torch.int64), requires_grad=False)
         self.is_reference_info_empty = False
         
@@ -242,20 +260,11 @@ def expand_reference_info(self, new_largest_label: int) -> None:
         """Expand reference info dimensions if newly given processed prompts have more lables."""
         if new_largest_label > (cur_largest_label := len(self.reference_info["reference_feats"]) - 1):
             diff = new_largest_label - cur_largest_label
-            self.reference_info["reference_feats"] = F.pad(
-                self.reference_info["reference_feats"],
-                (0, 0, 0, 0, 0, diff),
-                value=0.0,
-            )
+            padded_reference_feats = F.pad(self.reference_info["reference_feats"], (0, 0, 0, 0, 0, diff), value=0.0)
+            self.reference_info["reference_feats"] = Parameter(padded_reference_feats, requires_grad=False)
 
     @torch.no_grad()
-    def learn(
-        self,
-        images: Tensor,
-        processed_prompts: Dict[ScoredLabel, List[Dict[str, Tensor]]],
-        original_size: Tensor,
-        return_outputs: bool = False,
-    ) -> None:
+    def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[None, Tuple[ParameterDict, Tensor]]:
         """Get reference features.
 
         Using given images, get reference features and save it to PromptGetter.
@@ -263,86 +272,102 @@ def learn(
         Currently, single batch is only supported.
 
         Args:
-            images (Tensor): Given images for reference features.
-            processed_prompts (Dict[ScoredLabel, List[Dict[str, Tensor]]]): The whole class-wise prompts
-                processed at _preprocess_prompts.
-            original_size (Tensor): Original image size.
+            batch (List[Dict[str, Any]]): List of dictionaries containing images, prompts, and metas.
+                `batch` must contain images, prompts with bboxes, points, annotations, and polygons.
+            reset_feat (bool): Whether reset reference_info.
+                For OTX standalone, resetting reference_info will be conducted in on_train_start.
+                For other frameworks, setting it to True is required to reset reference_info. Defaults to False.
+            
+        Returns:
+            (Tuple[ParameterDict, Tensor]): reference_info and ref_masks.
         """
-        assert images.shape[0] == 1, "Only single batch is supported."
+        if reset_feat:
+            self.initialize_reference_info()
+            
+        # preprocess images and prompts
+        transformed_batch = [self.transforms(b.copy()) for b in batch]
+        processed_prompts = [self._preprocess_prompts(tb) for tb in transformed_batch]
         
         # initialize tensors to contain reference features and prompts
-        largest_label = max([int(label.id) for label in processed_prompts.keys()])
-        if self.is_reference_info_empty:
-            self.initialize_reference_info(largest_label)
-        else:
-            self.expand_reference_info(largest_label)
-            # TODO(sungchul): consider who to handle multiple reference features, currently replace it # noqa: TD003
-
-        image_embeddings = self.image_encoder(images)
-        processed_embedding = image_embeddings.squeeze().permute(1, 2, 0)
+        largest_label = max([int(label.id) for pp in processed_prompts for label in pp.keys()])
+        self.expand_reference_info(largest_label)
+        # TODO(sungchul): consider who to handle multiple reference features, currently replace it
+
+        batch_ref_masks: List[Tensor] = []
+        for tb, pp in zip(transformed_batch, processed_prompts):
+            # assign components
+            images = tb["images"].unsqueeze(0).to(self.device)
+            original_size = torch.as_tensor(tb["original_size"])
         
-        ref_masks = torch.zeros(largest_label + 1, *map(int, original_size))
-        for label, input_prompts in processed_prompts.items():
-            if label.name.lower() == "background":
-                # skip background
-                # TODO (sungchul): how to skip background class
-                continue
-
-            # generate reference mask
-            # TODO (sungchul): ensemble multi reference features (current : use merged masks)
-            ref_mask = torch.zeros(*map(int, original_size), dtype=torch.uint8, device=self.device)
-            for input_prompt in input_prompts:
-                if (anno_prompt := input_prompt.get("annotation", None)) is not None:
-                    # directly use annotation information as a mask
-                    ref_mask[anno_prompt == 1] += 1
-                else:
-                    merged_input_prompts: Dict[str, Tensor] = self._merge_prompts(label, input_prompt, processed_prompts)
-                    # TODO (sungchul): they must be processed in `_merge_prompts`
-                    # and it is required to be expanded to other prompts.
-                    point_coords = []
-                    point_labels = []
-                    if (input_prompt := merged_input_prompts.get("box", None)) is not None:
-                        point_coords = input_prompt.reshape(1, 2, 2)
-                        point_labels = torch.as_tensor([[2, 3]], device=self.device)
-
-                    if (input_prompt := merged_input_prompts.get("points", None)) is not None:
-                        point_coords = input_prompt.reshape(1, 1, 2)
-                        point_labels = torch.as_tensor([[1]], device=self.device)
-
-                    if (input_prompt := merged_input_prompts.get("annotations", None)) is not None:
-                        raise NotImplementedError()
-
-                    masks = self._predict_masks(
-                        image_embeddings=image_embeddings,
-                        point_coords=point_coords,
-                        point_labels=point_labels,
-                        original_size=original_size.unsqueeze(0),
-                        is_cascade=False,
-                    )
-                    ref_mask[masks] += 1
-            ref_mask = torch.clip(ref_mask, 0, 1).to(torch.float32)
-
-            ref_feat = None
-            default_threshold_reference = deepcopy(self.prompt_getter.default_threshold_reference)
-            while ref_feat is None:
-                logger.info(f"[*] default_threshold_reference : {default_threshold_reference:.4f}")
-                ref_feat = self._generate_masked_features(processed_embedding, ref_mask, default_threshold_reference)
-                default_threshold_reference -= 0.05
-
-            self.reference_info["reference_feats"][int(label.id)] = ref_feat.detach().cpu()
-            self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([[int(label.id)]])), dim=1), requires_grad=False)
-            ref_masks[int(label.id)] = ref_mask.detach().cpu()
-            
-        if return_outputs:
-            return self.reference_info, ref_masks
+            image_embeddings = self.image_encoder(images)
+            processed_embedding = image_embeddings.squeeze().permute(1, 2, 0)
+        
+            ref_masks = torch.zeros(largest_label + 1, *map(int, original_size))
+            for label, input_prompts in pp.items():
+                if label.name.lower() == "background":
+                    # skip background
+                    # TODO (sungchul): how to skip background class
+                    continue
+
+                # generate reference mask
+                # TODO (sungchul): ensemble multi reference features (current : use merged masks)
+                ref_mask = torch.zeros(*map(int, original_size), dtype=torch.uint8, device=self.device)
+                for input_prompt in input_prompts:
+                    if (prompt := input_prompt.get("annotations", None)) is not None:
+                        # directly use annotation information as a mask
+                        ref_mask[prompt == 1] += 1
+                    elif (prompt := input_prompt.get("polygons", None)) is not None:
+                        for polygon in prompt["polygons"]:
+                            contour = [[int(point[0]), int(point[1])] for point in polygon]
+                            mask_from_polygon = np.zeros(original_size, dtype=np.uint8)
+                            mask_from_polygon = cv2.drawContours(mask_from_polygon, np.asarray([contour]), 0, 1, -1)
+                            ref_mask[mask_from_polygon == 1] += 1
+                    elif (prompt := input_prompt.get("scribble_annotation", None)) is not None:
+                        logger.warning(f"scribble_annotation is not supported yet.")
+                        continue
+                    elif (prompt := input_prompt.get("scribble_polygon", None)) is not None:
+                        logger.warning(f"scribble_polygon is not supported yet.")
+                        continue
+                    else:
+                        point_coords = []
+                        point_labels = []
+                        if (prompt := input_prompt.get("bboxes", None)) is not None:
+                            point_coords = prompt["point_coords"].reshape(1, 2, 2)
+
+                        elif (prompt := input_prompt.get("points", None)) is not None:
+                            point_coords = prompt["point_coords"].reshape(1, 1, 2)
+
+                        point_labels = prompt["point_labels"]
+
+                        masks = self._predict_masks(
+                            image_embeddings=image_embeddings,
+                            point_coords=point_coords,
+                            point_labels=point_labels,
+                            original_size=original_size,
+                            is_cascade=False,
+                        )
+                        ref_mask[masks] += 1
+                ref_mask = torch.clip(ref_mask, 0, 1).to(torch.float32)
+
+                ref_feat = None
+                default_threshold_reference = deepcopy(self.prompt_getter.default_threshold_reference)
+                while ref_feat is None:
+                    logger.info(f"[*] default_threshold_reference : {default_threshold_reference:.4f}")
+                    ref_feat = self._generate_masked_features(processed_embedding, ref_mask, default_threshold_reference)
+                    default_threshold_reference -= 0.05
+
+                self.reference_info["reference_feats"][int(label.id)] = ref_feat.detach().cpu()
+                self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([[int(label.id)]])), dim=1), requires_grad=False)
+                ref_masks[int(label.id)] = ref_mask.detach().cpu()
+            batch_ref_masks.append(ref_masks)
+        return self.reference_info, batch_ref_masks
 
     @torch.no_grad()
     def infer(
-        self,
-        images: Tensor,
+        self, 
+        batch: List[Dict[str, Any]],
         reference_feats: Tensor,
         used_indices: Tensor,
-        original_size: Tensor,
         is_cascade: bool = False,
     ) -> List[List[DefaultDict[int, List[Tensor]]]]:
         """Zero-shot inference with reference features.
@@ -350,10 +375,9 @@ def infer(
         Get target results by using reference features and target images' features.
 
         Args:
-            images (Tensor): Given images for target results.
+            batch (List[Dict[str, Any]]): List of dictionaries containing images and metas.
             reference_feats (Tensor): Reference features for target prediction.
             used_indices (Tensor): To check which indices of reference features are validate.
-            original_size (Tensor): Original image size.
             is_cascade (bool): Whether use cascade inference. Defaults to False.
 
         Returns:
@@ -362,21 +386,21 @@ def infer(
                     1. Target images
                     2. Tuple of predicted masks and used points gotten by point selection
         """
-        assert images.shape[0] == 1, "Only single batch is supported."
-
-        total_results = []
-        for image in images:
-            if image.ndim == 3:
-                image = image.unsqueeze(0)
-
+        # preprocess images and prompts
+        transformed_batch = [self.transforms(b.copy()) for b in batch]
+        
+        total_results: List[List[Tensor]] = []
+        for tb in transformed_batch:
+            # assign components
+            images = tb["images"].unsqueeze(0).to(self.device)
+            original_size = torch.as_tensor(tb["original_size"])
+            
             image_embedding = self.image_encoder(images)
-
             total_points_scores, total_bg_coords = self.prompt_getter(
                 image_embedding=image_embedding,
                 reference_feats=reference_feats,
                 used_indices=used_indices,
-                original_size=original_size
-            )
+                original_size=original_size)
             predicted_masks: defaultdict = defaultdict(list)
             used_points: defaultdict = defaultdict(list)
             for label, (points_scores, bg_coords) in enumerate(zip(total_points_scores, total_bg_coords)):
@@ -394,7 +418,7 @@ def infer(
                         continue
 
                     point_coords = torch.cat((points_score[:2].unsqueeze(0), bg_coords), dim=0).unsqueeze(0)
-                    point_coords = self._preprocess_coords(point_coords, original_size[0], self.config.model.image_size)
+                    point_coords = self._preprocess_coords(point_coords, original_size, self.config.model.image_size)
                     point_labels = torch.as_tensor(
                         [1] + [0] * len(bg_coords), dtype=torch.float32, device=self.device
                     ).unsqueeze(0)
@@ -418,7 +442,7 @@ def _inspect_overlapping_areas(
         predicted_masks: Dict[int, List[Tensor]],
         used_points: Dict[int, List[Tensor]],
         threshold_iou: float = 0.8,
-    ):
+    ) -> None:
         def _calculate_mask_iou(mask1: Tensor, mask2: Tensor):
             assert mask1.ndim == 2 and mask2.ndim == 2
             intersection = torch.logical_and(mask1, mask2).sum().item()
@@ -487,7 +511,7 @@ def _predict_masks(
                 y, x = coords[:, 0], coords[:, 1]
                 box_coords = self._preprocess_coords(
                     torch.as_tensor([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device),
-                    original_size[0],
+                    original_size,
                     self.config.model.image_size)
                 point_coords = torch.cat((point_coords, box_coords), dim=1)
                 point_labels = torch.cat((point_labels, self.point_labels_box.to(self.device)), dim=1)
@@ -498,7 +522,7 @@ def _predict_masks(
                 point_labels=point_labels,
                 mask_input=mask_input,
                 has_mask_input=has_mask_input,
-                orig_size=original_size,
+                orig_size=original_size.unsqueeze(0),
             )
             masks = high_res_masks > self.config.model.mask_threshold
         _, masks = self._postprocess_masks(masks, logits, scores)
@@ -506,48 +530,20 @@ def _predict_masks(
 
     def training_step(self, batch, batch_idx) -> None:
         """Training step for `learn`."""
-        # TODO (sungchul): each prompt will be assigned with each label
-        bboxes = batch["bboxes"]
-        labels = batch["labels"]
-        # TODO (sungchul): support other below prompts
-        # points = batch["points"]
-        # annotations = batch["annotations"]
-
-        # organize prompts based on label
-        processed_prompts = self._preprocess_prompts(bboxes=bboxes[0], labels=labels[0])
-
-        self.learn(
-            images=batch["images"],
-            processed_prompts=processed_prompts,
-            original_size=batch.get("original_size")[0],
-        )
+        self.learn(batch)
 
     def predict_step(self, batch, batch_idx):
         """Predict step for `infer`."""
-        results = self.infer(
-            images=batch["images"],
-            reference_feats=self.reference_info["reference_feats"],
-            used_indices=self.reference_info["used_indices"],
-            original_size=batch.get("original_size")[0].unsqueeze(0))
+        results = self.infer(batch, self.reference_info["reference_feats"], self.reference_info["used_indices"])
         return [result[0] for result in results]  # tmp: only mask
 
-    def _preprocess_prompts(
-        self,
-        bboxes: Optional[Tensor] = None,
-        points: Optional[Tensor] = None,
-        annotations: Optional[Tensor] = None,
-        labels: Optional[Tensor] = None,
-    ) -> Dict[ScoredLabel, List[Dict[str, Tensor]]]:
+    def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[ScoredLabel, List[Dict[str, Tensor]]]:
         """Preprocess prompts.
 
         Currently, preprocessing for bounding boxes is only supported.
 
         Args:
-            bboxes (Tensor, optional): Bounding box prompts to be preprocessed.
-            points (Tensor, optional): Point prompts to be preprocessed, to be supported.
-            annotations (Tensor, optional): annotation prompts to be preprocessed, to be supported.
-            labels (Tensor, optional): Assigned labels according to given prompts.
-                Currently, it is only matched to bboxes, and it will be deprecated.
+            batch (Dict[str, Any]): Dictionary containing data and prompts information.
 
         Returns:
             (defaultdict[ScoredLabel, List[Dict[str, Tensor]]]): Processed and arranged each single prompt
@@ -555,18 +551,37 @@ def _preprocess_prompts(
                 as single annotation.
         """
         processed_prompts = defaultdict(list)
-        # TODO (sungchul): will be updated
-        if bboxes is not None:
-            for bbox, label in zip(bboxes, labels):
-                processed_prompts[label].append({"box": bbox.reshape(-1, 4)})
-
-        if points:
-            pass
-
-        if annotations:
-            pass
-
-        processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x[0].id_))  # type: ignore[assignment]
+        for prompt_name in ["annotations", "polygons", "bboxes", "points"]:
+            prompts = batch.get(prompt_name, None)
+            labels = batch["labels"].get(prompt_name, None)
+            if prompts is None or len(prompts) == 0:
+                continue
+            for prompt, label in zip(prompts, labels):
+                # TODO (sungchul): revisit annotations and polygons
+                if prompt_name == "annotations":
+                    processed_prompts[label].append({prompt_name: torch.as_tensor(prompt, device=self.device)})
+                elif prompt_name == "polygons":
+                    masks = []
+                    for polygon in prompt:
+                        contour = [[int(point[0]), int(point[1])] for point in polygon]
+                        mask_from_polygon = np.zeros(batch["original_size"], dtype=np.uint8)
+                        mask_from_polygon = cv2.drawContours(mask_from_polygon, np.asarray([contour]), 0, 1, -1)
+                        masks.append(mask_from_polygon)
+                    processed_prompts[label].append({prompt_name: torch.tensor(prompt, device=self.device)})
+                elif prompt_name == "bboxes":
+                    processed_prompts[label].append({
+                        prompt_name: {
+                            "point_coords": torch.as_tensor(prompt.reshape(-1, 2, 2), device=self.device),
+                            "point_labels": torch.tensor([[1]], device=self.device),
+                        }})
+                elif prompt_name == "points":
+                    processed_prompts[label].append({
+                        prompt_name: {
+                            "point_coords": torch.as_tensor(prompt.reshape(-1, 2), device=self.device),
+                            "point_labels": torch.tensor([[1]], device=self.device),
+                        }})
+
+        processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x[0].id_))
         return processed_prompts
     
     def _preprocess_coords(
@@ -730,7 +745,30 @@ def set_metrics(self) -> None:
     def configure_optimizers(self) -> None:
         """Skip configure_optimizers unused in zero-shot learning."""
         pass
+    
+    def _find_latest_reference_info(self) -> Union[str, None]:
+        """Find latest reference info to be used."""
+        if len(stamps := sorted(os.listdir("vpm_zsl_reference_infos"), reverse=True)) > 0:
+            return stamps[0]
+        self.initialize_reference_info()
+
+    def on_train_start(self) -> None:
+        """Called at the beginning of training after sanity check."""
+        self.initialize_reference_info()
+    
+    def on_predict_start(self) -> None:
+        """Called at the beginning of predicting."""
+        if (latest_stamp := self._find_latest_reference_info()) is not None:
+            latest_reference_info = self.path_reference_info.format(latest_stamp)
+            self.reference_info = torch.load(latest_reference_info)
+            self.reference_info.to(self.device)
+            logger.info(f"reference info saved at {latest_reference_info} was successfully loaded.")
 
     def training_epoch_end(self, outputs) -> None:
-        """Skip training_epoch_end unused in zero-shot learning."""
-        pass
+        """Called in the training loop at the very end of the epoch."""
+        if self.config.model.save_outputs:
+            path_reference_info = self.path_reference_info.format(datetime.now().strftime("%Y%m%d-%H%M%S"))
+            os.makedirs(os.path.dirname(path_reference_info), exist_ok=True)
+            torch.save(self.reference_info, path_reference_info)
+            json.dump(repr(self.trainer.datamodule.train_dataset.dataset), open(path_reference_info.replace("reference_info.pt", "reference_meta.json"), "w"))
+            logger.info(f"Saved reference info at {path_reference_info}")
diff --git a/src/otx/algorithms/visual_prompting/configs/zero_shot_sam_tiny_vit/config.yaml b/src/otx/algorithms/visual_prompting/configs/zero_shot_sam_tiny_vit/config.yaml
index 097390fba0f..a50ea244fbd 100644
--- a/src/otx/algorithms/visual_prompting/configs/zero_shot_sam_tiny_vit/config.yaml
+++ b/src/otx/algorithms/visual_prompting/configs/zero_shot_sam_tiny_vit/config.yaml
@@ -15,8 +15,8 @@ dataset:
       - 57.12
       - 57.375
   offset_bbox: 0
-  generate_point: false
-  generate_bbox: false
+  use_point: false
+  use_bbox: false
 
 model:
   name: SAM
@@ -36,6 +36,7 @@ model:
   # zero-shot
   default_threshold_reference: 0.3
   default_threshold_target: 0.65
+  save_outputs: True
 
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index 5e8e6945701..f5033bfb99a 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -710,6 +710,8 @@ def save_model(self, output_model: ModelEntity) -> None:
         logger.info("Saving the model weights and reference features.")
 
         model_info = self.model.state_dict()
+        model_info.pop("reference_info.reference_feats")
+        model_info.pop("reference_info.used_indices")
 
         buffer = io.BytesIO()
         torch.save(model_info, buffer)
diff --git a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
index 9b5093b004a..e6ba00166d0 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
@@ -4,7 +4,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-from typing import Tuple
+from typing import Tuple, Dict, Any
 
 import numpy as np
 import pytest
@@ -82,17 +82,42 @@ def test_get_outputs(self):
         assert "upscaled_masks" == results
 
     @e2e_pytest_unit
-    def test_preprocess(self):
+    @pytest.mark.parametrize("prompts,expected",
+    [
+        (
+            {
+                "bboxes": [np.array([[1, 1], [2, 2]])],
+                "points": [],
+                "labels": {"bboxes": [1]},
+                "original_size": (4, 4)
+            },
+            {
+                "point_coords": (1, 2, 2),
+                "point_labels": (1, 2),
+            }
+        ),
+        (
+            {
+                "bboxes": [],
+                "points": [np.array([[1, 1]])],
+                "labels": {"points": [1]},
+                "original_size": (4, 4)
+            },
+            {
+                "point_coords": (1, 1, 2),
+                "point_labels": (1, 1),
+            }
+        )
+    ])
+    def test_preprocess(self, prompts: Dict[str, Any], expected: Dict[str, Any]):
         """Test preprocess"""
-        prompts = {"bboxes": [np.array([[1, 1], [2, 2]])], "labels": [1], "original_size": (4, 4)}
-
         results = self.decoder.preprocess(prompts, {})
 
         assert isinstance(results, list)
         assert "point_coords" in results[0]
-        assert results[0]["point_coords"].shape == (1, 2, 2)
+        assert results[0]["point_coords"].shape == expected["point_coords"]
         assert "point_labels" in results[0]
-        assert results[0]["point_labels"].shape == (1, 2)
+        assert results[0]["point_labels"].shape == expected["point_labels"]
         assert "mask_input" in results[0]
         assert "has_mask_input" in results[0]
         assert "orig_size" in results[0]
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py
index 3cb1710d76e..0bac7032fa1 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py
@@ -77,9 +77,9 @@ def test_on_predict_epoch_end(self, use_mask: bool, expected: Any):
                 {
                     "masks": [torch.Tensor([[[0, 1, 0], [1, 1, 1], [0, 1, 0]]])],
                     "iou_predictions": [torch.Tensor([[0.9]])],
-                    "labels": [
-                        [ScoredLabel(label=LabelEntity("foreground", domain=Domain.VISUAL_PROMPTING), probability=0.0)],
-                    ],
+                    "labels": [{
+                        "bboxes": [ScoredLabel(label=LabelEntity("foreground", domain=Domain.VISUAL_PROMPTING), probability=0.0)],
+                    }],
                 }
             ]
         ]
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 58fa1554ed7..cba6b616aa3 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -9,6 +9,7 @@
 from collections import OrderedDict
 from tests.test_suite.e2e_test_system import e2e_pytest_unit
 import torch
+import numpy as np
 from torch import nn
 from omegaconf import DictConfig
 
@@ -105,7 +106,7 @@ def test_point_selection(self, mask_sim: torch.Tensor, expected: torch.Tensor) -
 class TestZeroShotSegmentAnything:
     @pytest.fixture
     def set_zero_shot_segment_anything(self, monkeypatch):
-        def zero_shot_segment_anything(state_dict: Optional[OrderedDict] = None):
+        def zero_shot_segment_anything(manual_config_update: Optional[Dict] = None, state_dict: Optional[OrderedDict] = None):
             monkeypatch.setattr(
                 "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SAMImageEncoder",
                 MockImageEncoder,
@@ -114,7 +115,7 @@ def zero_shot_segment_anything(state_dict: Optional[OrderedDict] = None):
                 "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SAMMaskDecoder",
                 MockMaskDecoder,
             )
-            return ZeroShotSegmentAnything(state_dict=state_dict)
+            return ZeroShotSegmentAnything(manual_config_update=manual_config_update, state_dict=state_dict)
 
         return zero_shot_segment_anything
 
@@ -123,18 +124,15 @@ def zero_shot_segment_anything(state_dict: Optional[OrderedDict] = None):
         "state_dict",
         [
             None,
-            {
-                "reference_info.reference_feats": torch.zeros(1),
-                "reference_info.used_indices": torch.zeros(1, dtype=torch.int64),
-            },
+            {},
         ],
     )
     def test_init(self, set_zero_shot_segment_anything, state_dict: Optional[Dict[str, Any]]) -> None:
         """Test __init__."""
         if state_dict is not None:
-            zero_shot_segment_anything_for_init_weights = set_zero_shot_segment_anything().state_dict()
-            zero_shot_segment_anything_for_init_weights.update(state_dict)
-            state_dict = zero_shot_segment_anything_for_init_weights
+            state_dict = set_zero_shot_segment_anything().state_dict()
+            state_dict.pop("reference_info.reference_feats")
+            state_dict.pop("reference_info.used_indices")
         
         zero_shot_segment_anything = set_zero_shot_segment_anything(state_dict=state_dict)
 
@@ -143,8 +141,8 @@ def test_init(self, set_zero_shot_segment_anything, state_dict: Optional[Dict[st
         assert zero_shot_segment_anything.config.model.freeze_mask_decoder
 
         if state_dict:
-            assert zero_shot_segment_anything.reference_info.reference_feats == torch.zeros(1)
-            assert zero_shot_segment_anything.reference_info.used_indices == torch.zeros(1, dtype=torch.int64)
+            assert zero_shot_segment_anything.reference_info.reference_feats is not None
+            assert zero_shot_segment_anything.reference_info.used_indices is not None
 
         assert zero_shot_segment_anything.reference_info.reference_feats.dtype == torch.float32
         assert zero_shot_segment_anything.reference_info.used_indices.dtype == torch.int64
@@ -171,7 +169,7 @@ def test_set_default_config(self, set_zero_shot_segment_anything) -> None:
     @e2e_pytest_unit
     def test_learn(self, mocker, set_zero_shot_segment_anything) -> None:
         """Test learn."""
-        zero_shot_segment_anything = set_zero_shot_segment_anything()
+        zero_shot_segment_anything = set_zero_shot_segment_anything(manual_config_update={"model.image_size": 4})
         mocker.patch.object(
             zero_shot_segment_anything,
             "_predict_masks",
@@ -179,12 +177,15 @@ def test_learn(self, mocker, set_zero_shot_segment_anything) -> None:
         )
         mocker.patch.object(zero_shot_segment_anything, "_generate_masked_features", return_value=torch.ones(1, 256))
 
-        processed_prompts = {MockScoredLabel(label=0, name="label"): [{"box": torch.tensor([[0, 0, 1, 1]])}]}
-        zero_shot_segment_anything.learn(
-            images=torch.ones((1, 3, 4, 4)),
-            processed_prompts=processed_prompts,
-            original_size=torch.tensor((4, 4)),
-        )
+        batch = [{
+            "images": np.ones((4, 4, 3), dtype=np.uint8),
+            "gt_masks": np.ones((4, 4), dtype=np.uint8),
+            "bboxes": np.array([[0, 0, 1, 1]], dtype=np.float32),
+            "points": np.zeros((0, 2), dtype=np.float32),
+            "labels": {"bboxes": [MockScoredLabel(label=0, name="label")]},
+            "original_size": np.array([4, 4], dtype=np.int64)
+        }]
+        zero_shot_segment_anything.learn(batch=batch, reset_feat=True)
 
         assert zero_shot_segment_anything.reference_info.reference_feats.shape == (1, 1, 256)
         assert zero_shot_segment_anything.reference_info.used_indices == torch.as_tensor([0])
@@ -198,18 +199,22 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
             MockPromptGetter,
         )
 
-        zero_shot_segment_anything = set_zero_shot_segment_anything()
+        zero_shot_segment_anything = set_zero_shot_segment_anything(manual_config_update={"model.image_size": 4})
         reference_feats = nn.Parameter(torch.rand(1, 1, 256), requires_grad=False)
         used_indices = nn.Parameter(torch.as_tensor([[0]], dtype=torch.int64), requires_grad=False)
         mocker.patch.object(
             SegmentAnything, "forward", return_value=(torch.ones(1, 4, 4, 4), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
         )
 
+        batch = [{
+            "images": np.ones((4, 4, 3), dtype=np.uint8),
+            "gt_masks": np.ones((4, 4), dtype=np.uint8),
+            "original_size": np.array([4, 4], dtype=np.int64)
+        }]
         total_results = zero_shot_segment_anything.infer(
-            images=torch.ones((1, 3, 4, 4)), 
+            batch=batch,
             reference_feats=reference_feats,
             used_indices=used_indices,
-            original_size=torch.tensor([[4, 4]], dtype=torch.int64)
         )
 
         for i, results in enumerate(total_results[0]):
@@ -332,28 +337,27 @@ def test_predict_masks(self, mocker, set_zero_shot_segment_anything) -> None:
             image_embeddings=torch.rand(1),
             point_coords=torch.rand(1, 2, 2),
             point_labels=torch.randint(low=0, high=2, size=(1, 2)),
-            original_size=torch.tensor([[8, 8]], dtype=torch.int64),
+            original_size=torch.tensor([8, 8], dtype=torch.int64),
         )
         assert mask.shape == (8, 8)
 
     @e2e_pytest_unit
     def test_preprocess_prompts(self, set_zero_shot_segment_anything) -> None:
-        """Test _preprocess_prompts.
-
-        TODO (sungchul)
-        - get inputs grouped as label and prompts
-        - use points and annotations.
-        """
+        """Test _preprocess_prompts."""
         zero_shot_segment_anything = set_zero_shot_segment_anything()
-        bboxes = [torch.tensor([0, 0, 1, 1])]
-        labels = [MockScoredLabel(label=1)]
-        processed_prompts = zero_shot_segment_anything._preprocess_prompts(
-            bboxes=bboxes,
-            labels=labels,
-        )
-
-        # processed_prompts = {labels[0]: [{"box": torch.tensor([[0, 0, 1, 1]])}]}
-        assert torch.equal(processed_prompts[labels[0]][0].get("box")[0], bboxes[0])
+        transformed_batch = {
+            "bboxes": torch.tensor([[0, 0, 1, 1]]),
+            "points": torch.tensor([[2, 2]]),
+            "labels": {"bboxes": [MockScoredLabel(label=1)], "points": [MockScoredLabel(label=1)]}
+        }
+        processed_prompts = zero_shot_segment_anything._preprocess_prompts(transformed_batch)
+        
+        for prompts in processed_prompts.values():
+            for prompt in prompts:
+                if "bboxes" in prompt:
+                    prompt["bboxes"]["point_coords"].shape == (1, 2, 2)
+                elif "points" in prompt:
+                    prompt["points"]["point_coords"].shape == (1, 1, 2)
 
     @e2e_pytest_unit
     def test_generate_masked_features(self, set_zero_shot_segment_anything) -> None:
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index d2a5852e0af..fd45f13f1fb 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -343,6 +343,8 @@ def test_export_to_onnx(self):
             },
             "visual_prompting_prompt_getter": {
                 "image_embeddings": np.random.randn(1, embed_dim, *embed_size).astype(dtype=np.float32),
+                "reference_feats": np.random.randn(1, 1, 256).astype(dtype=np.float32),
+                "used_indices": np.array([[0]], dtype=np.int64),
                 "original_size": np.random.randint(low=0, high=image_size * 2, size=(1, 2), dtype=np.int64),
                 "threshold": np.array([[0.1]], dtype=np.float32),
                 "num_bg_points": np.random.randint(low=1, high=image_size, size=(1, 1), dtype=np.int64),
@@ -379,7 +381,7 @@ def test_save_model(self, mocker):
         mocker_otx_model = mocker.patch("otx.api.entities.model.ModelEntity")
         mocker_io_bytes_io = mocker.patch("io.BytesIO")
         mocker_torch_save = mocker.patch("torch.save")
-        mocker.patch.object(self.zero_shot_task.model, "state_dict", return_value={"reference_info.reference_feats": None})
+        mocker.patch.object(self.zero_shot_task.model, "state_dict", return_value={"reference_info.reference_feats": None, "reference_info.used_indices": None})
 
         self.zero_shot_task.model.reference_info = "reference_info"
 

From bdd7fc5a90b4edb1c4f5a548eff84891a666ecdd Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Fri, 16 Feb 2024 14:24:37 +0900
Subject: [PATCH 10/28] Update OV inference

---
 .../model_wrappers/openvino_models.py         |   5 +-
 .../zero_shot_segment_anything.py             |  19 +--
 .../visual_prompting/tasks/inference.py       |   6 +-
 .../visual_prompting/tasks/openvino.py        |  92 +++++++----
 .../model_wrappers/test_openvino_models.py    |  22 ++-
 .../test_zero_shot_segment_anything.py        |   8 +-
 .../visual_prompting/tasks/test_inference.py  |   4 +-
 .../visual_prompting/tasks/test_openvino.py   | 146 +++++++++++++++---
 8 files changed, 227 insertions(+), 75 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 128e92687e4..4861f6183be 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -62,6 +62,9 @@ class PromptGetter(ImageModel):
     """PromptGetter class for zero-shot visual prompting of openvino model wrapper."""
 
     __model__ = "prompt_getter"
+    
+    def __init__(self, inference_adapter, configuration=None, preload=False):
+        super().__init__(inference_adapter, configuration, preload)
 
     @classmethod
     def parameters(cls) -> Dict[str, Any]:  # noqa: D102
@@ -187,7 +190,7 @@ def sigmoid(x):
             blur_strength=self.blur_strength,
         )
 
-        probability = max(min(float(outputs["iou_predictions"]), 1.0), 0.0)
+        probability = max(min(float(outputs["scores"]), 1.0), 0.0)
         meta["label"].probability = probability
 
         return hard_prediction, soft_prediction
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index ed902e2c4be..f33b8170132 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -8,7 +8,7 @@
 import json
 from copy import deepcopy
 from itertools import product
-import ast
+import pickle
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Union
 
 import torch
@@ -49,7 +49,7 @@ def set_default_thresholds(self, default_threshold_reference: float, default_thr
 
     def forward(
         self,
-        image_embedding: Tensor,
+        image_embeddings: Tensor,
         reference_feats: Tensor,
         used_indices: Tensor,
         original_size: Tensor,
@@ -57,7 +57,7 @@ def forward(
         num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates."""
-        device = image_embedding.device
+        device = image_embeddings.device
         original_size = original_size.squeeze()
         threshold = threshold.squeeze().to(device)
         num_bg_points = num_bg_points.squeeze()
@@ -66,7 +66,7 @@ def forward(
         total_bg_coords: Tensor = torch.zeros(used_indices.max() + 1, num_bg_points, 2, device=device)
         for label in used_indices[0]:
             points_scores, bg_coords = self.get_prompt_candidates(
-                image_embedding=image_embedding,
+                image_embeddings=image_embeddings,
                 reference_feat=reference_feats[label],
                 original_size=original_size,
                 threshold=threshold,
@@ -88,7 +88,7 @@ def forward(
 
     def get_prompt_candidates(
         self,
-        image_embedding: Tensor,
+        image_embeddings: Tensor,
         reference_feat: Tensor,
         original_size: Tensor,
         threshold: Union[Tensor, float] = 0.,
@@ -96,7 +96,7 @@ def get_prompt_candidates(
         device: Union[torch.device, str] = torch.device("cpu"),
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates from given reference and target features."""
-        target_feat = image_embedding.squeeze()
+        target_feat = image_embeddings.squeeze()
         c_feat, h_feat, w_feat = target_feat.shape
         target_feat = target_feat / target_feat.norm(dim=0, keepdim=True)
         target_feat = target_feat.reshape(c_feat, h_feat * w_feat)
@@ -395,9 +395,9 @@ def infer(
             images = tb["images"].unsqueeze(0).to(self.device)
             original_size = torch.as_tensor(tb["original_size"])
             
-            image_embedding = self.image_encoder(images)
+            image_embeddings = self.image_encoder(images)
             total_points_scores, total_bg_coords = self.prompt_getter(
-                image_embedding=image_embedding,
+                image_embeddings=image_embeddings,
                 reference_feats=reference_feats,
                 used_indices=used_indices,
                 original_size=original_size)
@@ -423,7 +423,7 @@ def infer(
                         [1] + [0] * len(bg_coords), dtype=torch.float32, device=self.device
                     ).unsqueeze(0)
                     mask = self._predict_masks(
-                        image_embeddings=image_embedding,
+                        image_embeddings=image_embeddings,
                         point_coords=point_coords,
                         point_labels=point_labels,
                         original_size=original_size,
@@ -770,5 +770,6 @@ def training_epoch_end(self, outputs) -> None:
             path_reference_info = self.path_reference_info.format(datetime.now().strftime("%Y%m%d-%H%M%S"))
             os.makedirs(os.path.dirname(path_reference_info), exist_ok=True)
             torch.save(self.reference_info, path_reference_info)
+            pickle.dump({k: v.numpy() for k, v in self.reference_info.items()}, open(path_reference_info.replace(".pt", ".pickle"), "wb"))
             json.dump(repr(self.trainer.datamodule.train_dataset.dataset), open(path_reference_info.replace("reference_info.pt", "reference_meta.json"), "w"))
             logger.info(f"Saved reference info at {path_reference_info}")
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index f5033bfb99a..f9c705a2489 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -643,10 +643,10 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
             elif module == "visual_prompting_prompt_getter":
                 dummy_inputs = {
                     "image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float32),
-                    "reference_feats": torch.randn(1, 1, 256, dtype=torch.float32),
-                    "used_indices": torch.as_tensor([[0]], dtype=torch.int64),
+                    "reference_feats": torch.randn(2, 1, 256, dtype=torch.float32),
+                    "used_indices": torch.tensor([[0, 1]], dtype=torch.int64),
                     "original_size": torch.randint(low=0, high=image_size * 2, size=(1, 2), dtype=torch.int64),
-                    "threshold": torch.as_tensor([[0.1]], dtype=torch.float32),
+                    "threshold": torch.tensor([[0.1]], dtype=torch.float32),
                     "num_bg_points": torch.randint(low=1, high=image_size, size=(1, 1), dtype=torch.int64),
                 }
                 output_names = ["total_points_scores", "total_bg_coords"]
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index d550f1fb33d..f55e6f6bdbd 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -15,6 +15,7 @@
 # and limitations under the License.
 
 import io
+import pickle
 import json
 import os
 import random
@@ -206,10 +207,6 @@ def await_all(self) -> None:
         self.model["image_encoder"].await_all()
         self.model["decoder"].await_all()
 
-    def pre_process_prompt_getter(self, *args, **kwargs) -> Any:
-        """Pre-process function of OpenVINO Zero-shot VIsual Prompting Inferencer for prompt getter."""
-        pass
-
 
 class OpenVINOZeroShotVisualPromptingInferencer(OpenVINOVisualPromptingInferencer):
     """Inferencer implementation for Zero-shot Visual Prompting using OpenVINO backend.
@@ -291,34 +288,45 @@ def __init__(
 
         self.point_labels_box = np.array([[2, 3]], dtype=np.float32)
         self.has_mask_inputs = [np.array([[0.0]]), np.array([[1.0]])]
+        
+        self.reference_feats = None
+        self.used_indices = None
+        self.path_reference_info = "vpm_zsl_reference_infos/{}/reference_info.pickle"
 
-    def pre_process(  # type: ignore
-        self, dataset_item: DatasetItemEntity, extra_processing: bool = False
-    ) -> Tuple[Dict[str, np.ndarray], Dict[str, Any]]:
+    def pre_process_image_encoder(self, inputs: np.ndarray, extra_processing: bool = False) -> Tuple[Dict[str, np.ndarray], Dict[str, Any]]:
         """Pre-process function of OpenVINO Zero-shot Visual Prompting Inferencer for image encoder."""
-        return self.model["image_encoder"].preprocess(dataset_item.numpy, extra_processing)
+        return self.model["image_encoder"].preprocess(inputs, extra_processing)
 
     def pre_process_prompt_getter(
-        self, image_embeddings: Dict[str, np.ndarray], original_size: np.ndarray
+        self,
+        image_embeddings: Dict[str, np.ndarray],
+        reference_feats: np.ndarray,
+        used_indices: np.ndarray,
+        original_size: np.ndarray,
     ) -> Dict[str, np.ndarray]:
         """Pre-process function of OpenVINO Zero-shot VIsual Prompting Inferencer for prompt getter."""
         inputs_prompt_getter = {
             "original_size": original_size[None],
-            "threshold": np.array([[self.model["prompt_getter"].sim_threshold]], dtype=np.float32),
-            "num_bg_points": np.array([[self.model["prompt_getter"].num_bg_points]], dtype=np.int64),
+            "reference_feats": reference_feats,
+            "used_indices": used_indices,
+            "threshold": np.asarray([[self.model["prompt_getter"].sim_threshold]], dtype=np.float32),
+            "num_bg_points": np.asarray([[self.model["prompt_getter"].num_bg_points]], dtype=np.int64),
         }
         inputs_prompt_getter.update(image_embeddings)
         return inputs_prompt_getter
-
-    def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
+    
+    def learn(self, images: np.ndarray):
+        """Learn"""
+    
+    def infer(self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray) -> List:
         """Perform a prediction for a given input image."""
         # forward image encoder
-        images, meta = self.pre_process(dataset_item)
-        original_size = np.array(meta["original_shape"][:2], dtype=np.int64)
+        images, meta = self.pre_process_image_encoder(images)
+        original_size = np.asarray(meta["original_shape"][:2], dtype=np.int64)
         image_embeddings = self.forward_image_encoder(images)
 
         # get point candidates
-        inputs_prompt_getter = self.pre_process_prompt_getter(image_embeddings, original_size)
+        inputs_prompt_getter = self.pre_process_prompt_getter(image_embeddings, reference_feats, used_indices, original_size)
         total_prompts = self.forward_prompt_getter(inputs_prompt_getter)
 
         annotations: DefaultDict = defaultdict(list)
@@ -343,10 +351,14 @@ def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type:
                 point_coords = np.concatenate((np.array([[x, y]]), bg_coords), axis=0, dtype=np.float32)
                 point_coords = self.model["decoder"]._apply_coords(point_coords, original_size)
                 point_labels = np.array([1] + [0] * len(bg_coords), dtype=np.float32)
-                inputs_decoder = {"point_coords": point_coords[None], "point_labels": point_labels[None]}
+                inputs_decoder = {
+                    "point_coords": point_coords[None],
+                    "point_labels": point_labels[None],
+                    "orig_size": original_size[None]}
                 inputs_decoder.update(image_embeddings)
 
                 prediction = self.forward_decoder(inputs_decoder, original_size)
+                prediction.update({"scores": points_score[-1]})
                 metadata = {
                     "label": [_label for _label in self.labels if int(_label.id_) == label][0],
                     "original_size": original_size[None],
@@ -357,8 +369,14 @@ def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type:
                 annotations[label].extend(annotation)
                 predicted_masks[label].append(hard_prediction)
                 used_points[label].append(points_score)
-        self.__inspect_overlapping_areas(predicted_masks, used_points, annotations)
-        return sum(annotations.values(), [])
+        self._inspect_overlapping_areas(predicted_masks, used_points, annotations)
+        return sum(annotations.values(), []), predicted_masks, used_points
+
+    def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
+        """Perform a prediction for a given input image."""            
+        self._get_reference_info()
+        results = self.infer(dataset_item.numpy, self.reference_feats, self.used_indices)
+        return results[0]
 
     def forward_prompt_getter(self, inputs: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
@@ -370,7 +388,6 @@ def forward_decoder(  # type: ignore
         """Forward function of OpenVINO Visual Prompting Inferencer."""
         logits: np.ndarray
         scores: np.ndarray
-        mask_slice = slice(0, 1)
         for i in range(3):
             if i == 0:
                 # First-step prediction
@@ -383,7 +400,7 @@ def forward_decoder(  # type: ignore
                 # Cascaded Post-refinement-1
                 mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)
                 if masks.sum() == 0:
-                    return {"masks": masks}
+                    return {"upscaled_masks": masks}
 
                 has_mask_input = self.has_mask_inputs[1]
 
@@ -391,7 +408,7 @@ def forward_decoder(  # type: ignore
                 # Cascaded Post-refinement-2
                 mask_input, masks = self._postprocess_masks(masks, logits, scores)
                 if masks.sum() == 0:
-                    return {"masks": masks}
+                    return {"upscaled_masks": masks}
 
                 has_mask_input = self.has_mask_inputs[1]
                 y, x = np.nonzero(masks)
@@ -409,14 +426,12 @@ def forward_decoder(  # type: ignore
             masks = upscaled_masks > self.model["decoder"].mask_threshold
 
         _, masks = self._postprocess_masks(masks, logits, scores)
-        return {"masks": masks}
+        return {"upscaled_masks": masks}
 
     def _postprocess_masks(
         self, masks: np.ndarray, logits: np.ndarray, scores: np.ndarray, is_single: bool = False
     ) -> Tuple[np.ndarray, ...]:
         """Post-process logits for resized masks according to best index based on scores."""
-        masks = masks.transpose(2, 0, 1)[None]
-
         if is_single:
             best_idx = 0
         else:
@@ -436,14 +451,14 @@ def _postprocess_masks(
             best_idx = np.argmax(scores[0])
         return logits[:, [best_idx]], masks[0, best_idx]
 
-    def __inspect_overlapping_areas(
+    def _inspect_overlapping_areas(
         self,
         predicted_masks: Dict[int, List[np.ndarray]],
         used_points: Dict[int, List[np.ndarray]],
         annotations: Dict[int, List[np.ndarray]],
         threshold_iou: float = 0.8,
     ):
-        def __calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
+        def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
             assert mask1.ndim == 2 and mask2.ndim == 2
             intersection = np.logical_and(mask1, mask2).sum().item()
             union = np.logical_or(mask1, mask2).sum().item()
@@ -461,21 +476,36 @@ def __calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
             overlapped_label = []
             overlapped_other_label = []
             for (im, mask), (jm, other_mask) in product(enumerate(masks), enumerate(other_masks)):
-                if __calculate_mask_iou(mask, other_mask) > threshold_iou:
+                if _calculate_mask_iou(mask, other_mask) > threshold_iou:
                     if used_points[label][im][2] > used_points[other_label][jm][2]:
                         overlapped_other_label.append(jm)
                     else:
                         overlapped_label.append(im)
 
-            for im in overlapped_label[::-1]:
+            for im in sorted(list(set(overlapped_label)), reverse=True):
                 masks.pop(im)
                 used_points[label].pop(im)
                 annotations[label].pop(im)
 
-            for jm in overlapped_other_label[::-1]:
+            for jm in sorted(list(set(overlapped_other_label)), reverse=True):
                 other_masks.pop(jm)
                 used_points[other_label].pop(jm)
                 annotations[other_label].pop(jm)
+                
+    def _find_latest_reference_info(self) -> Union[str, None]:
+        """Find latest reference info to be used."""
+        if len(stamps := sorted(os.listdir("vpm_zsl_reference_infos"), reverse=True)) > 0:
+            return stamps[0]
+        return None
+    
+    def _get_reference_info(self) -> None:
+        """Get reference info through loading previously saved one or running `learn`."""
+        if self.reference_feats is None and self.used_indices is None and (latest_stamp := self._find_latest_reference_info()) is not None:
+            # load previously saved reference info
+            latest_reference_info = self.path_reference_info.format(latest_stamp)
+            reference_info = pickle.load(open(latest_reference_info, "rb"))
+            self.reference_feats = reference_info["reference_feats"]
+            self.used_indices = reference_info["used_indices"]
 
 
 class OTXOpenVinoDataLoader:
@@ -562,7 +592,7 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
 
         items = self.dataset[index]
         images, meta = self.inferencer.pre_process(items, extra_processing=True)  # type: ignore
-        original_size = np.array(meta["original_shape"][:2])
+        original_size = np.asarray(meta["original_shape"][:2])
         _, _, h, w = images["images"].shape
         pad_width = ((0, 0), (0, 0), (0, self.target_length - h), (0, self.target_length - w))
         images["images"] = np.pad(images["images"], pad_width, mode="constant", constant_values=0)
diff --git a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
index e6ba00166d0..1518dbd6131 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
@@ -56,6 +56,26 @@ def test_parameters(self):
 
         assert params.get("sim_threshold").default_value == 0.5
         assert params.get("num_bg_points").default_value == 1
+        
+    @e2e_pytest_unit
+    def test_get_inputs(self, mocker):
+        """Test _get_inputs."""
+        mocker.patch.object(ImageModel, "__init__")
+        prompt_getter = PromptGetter("adapter")
+        
+        prompt_getter.inputs = {
+            "image_embeddings": np.ones((1, 4, 4, 3)),
+            "reference_feats": np.ones((2, 1, 256)),
+            "used_indices": np.array([[0, 1]], dtype=np.int64),
+            "original_size": np.array([[4, 4]], dtype=np.int64),
+            "threshold": np.array([[0.1]]),
+            "num_bg_points": np.array([[1]], dtype=np.int64),
+        }
+
+        returned_value = prompt_getter._get_inputs()
+
+        assert returned_value[0] == ["image_embeddings"]
+        assert returned_value[1] == ["reference_feats", "used_indices", "original_size", "threshold", "num_bg_points"]
 
 
 class TestDecoder:
@@ -162,7 +182,7 @@ def test_postprocess(self, mocker):
         self.decoder.output_blob_name = "masks"
         self.decoder.soft_threshold = 0.5
         self.decoder.blur_strength = 2
-        fake_output = {"masks": np.ones((4, 4)), "iou_predictions": 0.1}
+        fake_output = {"masks": np.ones((4, 4)), "scores": 0.1}
         fake_metadata = {"original_size": np.array([[6, 6]]), "label": mocker.Mock(spec=LabelEntity)}
         returned_value = self.decoder.postprocess(outputs=fake_output, meta=fake_metadata)
 
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index cba6b616aa3..61a8b3cd7de 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -52,13 +52,13 @@ def test_forward(self, mocker, result_point_selection: torch.Tensor) -> None:
             self.prompt_getter,
             "get_prompt_candidates",
             return_value=(result_point_selection, torch.zeros(1, 2)))
-        image_embedding = torch.ones(1, 4, 4, 4)
+        image_embeddings = torch.ones(1, 4, 4, 4)
         reference_feats = torch.rand(1, 1, 4)
         used_indices = torch.as_tensor([[0]])
         original_size = torch.tensor((self.prompt_getter.image_size, self.prompt_getter.image_size), dtype=torch.int64)
 
         total_points_scores, total_bg_coords = self.prompt_getter(
-            image_embedding=image_embedding, reference_feats=reference_feats, used_indices=used_indices, original_size=original_size
+            image_embeddings=image_embeddings, reference_feats=reference_feats, used_indices=used_indices, original_size=original_size
         )
         
         assert total_points_scores.shape[0] == 1
@@ -72,14 +72,14 @@ def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tenso
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
         )
         mocker.patch.object(self.prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2)))
-        image_embedding = torch.ones(1, 4, 4, 4)
+        image_embeddings = torch.ones(1, 4, 4, 4)
         reference_feat = torch.rand(1, 4)
         original_size = torch.tensor(
             [[self.prompt_getter.image_size, self.prompt_getter.image_size]], dtype=torch.int64
         )
 
         points_scores, bg_coords = self.prompt_getter.get_prompt_candidates(
-            image_embedding=image_embedding, reference_feat=reference_feat, original_size=original_size
+            image_embeddings=image_embeddings, reference_feat=reference_feat, original_size=original_size
         )
 
         assert torch.all(points_scores == result_point_selection)
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index fd45f13f1fb..59c5fa37215 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -343,8 +343,8 @@ def test_export_to_onnx(self):
             },
             "visual_prompting_prompt_getter": {
                 "image_embeddings": np.random.randn(1, embed_dim, *embed_size).astype(dtype=np.float32),
-                "reference_feats": np.random.randn(1, 1, 256).astype(dtype=np.float32),
-                "used_indices": np.array([[0]], dtype=np.int64),
+                "reference_feats": np.random.randn(2, 1, 256).astype(dtype=np.float32),
+                "used_indices": np.array([[0, 1]], dtype=np.int64),
                 "original_size": np.random.randint(low=0, high=image_size * 2, size=(1, 2), dtype=np.int64),
                 "threshold": np.array([[0.1]], dtype=np.float32),
                 "num_bg_points": np.random.randint(low=1, high=image_size, size=(1, 1), dtype=np.int64),
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 459aba889f5..bf791571efd 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -193,24 +193,24 @@ def setup(self, mocker):
         visual_prompting_hparams = self.task_environment.get_hyper_parameters(VisualPromptingBaseConfig)
         label_schema = self.task_environment.label_schema
 
-        self.visual_prompting_ov_inferencer = OpenVINOZeroShotVisualPromptingInferencer(
+        self.zero_shot_visual_prompting_ov_inferencer = OpenVINOZeroShotVisualPromptingInferencer(
             visual_prompting_hparams,
             label_schema,
             {"image_encoder": "", "prompt_getter": "", "decoder": ""},
             {"image_encoder": "", "prompt_getter": "", "decoder": ""},
         )
-        self.visual_prompting_ov_inferencer.model["decoder"] = mocker.patch(
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"] = mocker.patch(
             "otx.algorithms.visual_prompting.tasks.openvino.model_wrappers.Decoder", autospec=True
         )
-        self.visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.3
-        self.visual_prompting_ov_inferencer.model["decoder"]._apply_coords.return_value = np.array([[1, 1]])
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.3
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"]._apply_coords.return_value = np.array([[1, 1]])
 
     @e2e_pytest_unit
     def test_predict(self, mocker):
         """Test predict."""
         mocker_pre_process = mocker.patch.object(
             OpenVINOZeroShotVisualPromptingInferencer,
-            "pre_process",
+            "pre_process_image_encoder",
             return_value=(torch.zeros((1, 3, 2, 2)), {"original_shape": (4, 4, 1)}),
         )
         mocker_forward = mocker.patch.object(
@@ -223,15 +223,13 @@ def test_predict(self, mocker):
             "forward_prompt_getter",
             return_value={"total_points_scores": np.array([[[1, 1, 1]]]), "total_bg_coords": np.array([[[2, 2]]])},
         )
-        mocker_forward_decoder = mocker.patch.object(
-            OpenVINOZeroShotVisualPromptingInferencer, "forward_decoder", return_value=None
-        )
+        mocker_forward_decoder = mocker.patch.object(OpenVINOZeroShotVisualPromptingInferencer, "forward_decoder", return_value={})
         mocker_post_process = mocker.patch.object(
             OpenVINOZeroShotVisualPromptingInferencer, "post_process", return_value=(self.fake_annotation, None, None)
         )
         fake_input = mocker.Mock(spec=DatasetItemEntity)
 
-        returned_value = self.visual_prompting_ov_inferencer.predict(fake_input)
+        returned_value = self.zero_shot_visual_prompting_ov_inferencer.predict(fake_input)
 
         mocker_pre_process.assert_called_once()
         mocker_forward.assert_called_once()
@@ -246,12 +244,12 @@ def test_predict(self, mocker):
             (
                 (np.ones((1, 1)), np.ones((3, 3))),
                 {"upscaled_masks": np.ones((3, 3)), "iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
-                {"masks": np.ones((3, 3))},
+                {"upscaled_masks": np.ones((3, 3))},
             ),
             (
                 (np.zeros((2, 2)), np.zeros((3, 3))),
                 {"upscaled_masks": np.zeros((3, 3)), "iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
-                {"masks": np.zeros((3, 3))},
+                {"upscaled_masks": np.zeros((3, 3))},
             ),
         ],
     )
@@ -264,16 +262,16 @@ def test_forward_decoder(
     ):
         """Test forward_decoder."""
         mocker.patch.object(
-            self.visual_prompting_ov_inferencer.model["decoder"], "infer_sync", return_value=infer_sync_output
+            self.zero_shot_visual_prompting_ov_inferencer.model["decoder"], "infer_sync", return_value=infer_sync_output
         )
         mocker.patch.object(
-            self.visual_prompting_ov_inferencer.model["decoder"],
+            self.zero_shot_visual_prompting_ov_inferencer.model["decoder"],
             "_apply_coords",
             return_value=np.array([[[1, 1]]], dtype=np.float32),
         )
-        mocker.patch.object(self.visual_prompting_ov_inferencer, "_postprocess_masks", return_value=postprocess_output)
+        mocker.patch.object(self.zero_shot_visual_prompting_ov_inferencer, "_postprocess_masks", return_value=postprocess_output)
 
-        result = self.visual_prompting_ov_inferencer.forward_decoder(
+        result = self.zero_shot_visual_prompting_ov_inferencer.forward_decoder(
             inputs={
                 "image_embeddings": np.empty((1, 4, 2, 2)),
                 "point_coords": np.array([[[1, 1]]], dtype=np.float32),
@@ -282,39 +280,139 @@ def test_forward_decoder(
             original_size=np.array([3, 3]),
         )
 
-        assert np.all(result["masks"] == expected["masks"])
+        assert np.all(result["upscaled_masks"] == expected["upscaled_masks"])
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "masks,expected_masks",
         [
             (
-                np.repeat(np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])[..., None], 4, axis=-1),
+                np.repeat(np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])[None], 4, axis=0)[None],
                 np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_)
             ),
             (
                 np.concatenate(
                     (
-                        np.repeat(np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])[..., None], 3, axis=-1),
-                        np.zeros((3, 3, 1)),
+                        np.repeat(np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])[None], 3, axis=0)[None],
+                        np.zeros((1, 1, 3, 3)),
                     ),
-                    axis=-1,
+                    axis=1,
                 ),
                 np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_)
             ),
-            (np.zeros((3, 3, 4)), np.zeros((3, 3))),
+            (np.zeros((1, 4, 3, 3)), np.zeros((3, 3))),
         ],
     )
     def test_postprocess_masks(self, masks: np.ndarray, expected_masks: np.ndarray):
         """Test _postprocess_masks."""
-        self.visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.0
-        self.visual_prompting_ov_inferencer.model["decoder"].image_size = 3
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.0
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].image_size = 3
 
-        _, result_masks = self.visual_prompting_ov_inferencer._postprocess_masks(
+        _, result_masks = self.zero_shot_visual_prompting_ov_inferencer._postprocess_masks(
             masks=masks, logits=np.empty((1, 4, 2, 2)), scores=np.array([[0.5, 0.7, 0.8, 0.9]]))
 
         assert result_masks.shape == (3, 3)
         assert np.all(result_masks == expected_masks)
+        
+    @e2e_pytest_unit
+    def test_inspect_overlapping_areas(self) -> None:
+        """Test _inspect_overlapping_areas."""
+        predicted_masks = {
+            0: [
+                np.array(
+                    [
+                        [1, 1, 0, 0, 0, 0],
+                        [1, 1, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+                np.array(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 1, 1, 1, 0],
+                        [0, 0, 1, 1, 1, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+                np.array(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 1, 1, 1, 0, 0],
+                        [0, 1, 1, 1, 0, 0],
+                    ],
+                ),
+            ],
+            1: [
+                np.array(
+                    [
+                        [0, 0, 0, 1, 1, 0],
+                        [0, 0, 0, 1, 1, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+                np.array(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 1, 1],
+                        [0, 0, 0, 0, 1, 1],
+                    ],
+                ),
+                np.array(
+                    [
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 1, 1, 0, 0, 0],
+                        [0, 1, 1, 0, 0, 0],
+                    ],
+                ),
+                np.array(
+                    [
+                        [1, 1, 0, 0, 0, 0],
+                        [1, 1, 0, 0, 0, 0],
+                        [1, 1, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0],
+                    ],
+                ),
+            ],
+        }
+        used_points = {
+            0: [
+                np.array([0, 0, 0.5]),  # to be removed
+                np.array([2, 2, 0.5]),
+                np.array([1, 4, 0.5]),
+            ],
+            1: [
+                np.array([3, 0, 0.5]),
+                np.array([4, 4, 0.5]),
+                np.array([1, 4, 0.3]),  # to be removed
+                np.array([0, 0, 0.7]),
+            ],
+        }
+
+        self.zero_shot_visual_prompting_ov_inferencer._inspect_overlapping_areas(predicted_masks, used_points, predicted_masks.copy(), threshold_iou=0.5)
+
+        assert len(predicted_masks[0]) == 1
+        assert len(predicted_masks[1]) == 2
+        assert all(np.array([2, 2, 0.5]) == used_points[0][0])
+        assert all(np.array([0, 0, 0.7]) == used_points[1][2])
 
 
 class TestOTXOpenVinoDataLoader:

From 358d97b88acaf723da36661946da81c2779ee8ce Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Fri, 16 Feb 2024 22:32:41 +0900
Subject: [PATCH 11/28] Fix for unit & intg tests

---
 .../datasets/pipelines/sam_transforms.py      |  3 +-
 .../zero_shot_segment_anything.py             | 98 +++++--------------
 .../visual_prompting/tasks/openvino.py        |  1 +
 .../test_zero_shot_segment_anything.py        | 26 -----
 .../visual_prompting/tasks/test_openvino.py   |  2 +-
 .../visual_prompting/test_helpers.py          |  7 +-
 6 files changed, 36 insertions(+), 101 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
index a47efc97540..63a58b9229e 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/sam_transforms.py
@@ -36,7 +36,8 @@ def __call__(self, item: Dict[str, Union[List, Tensor]]) -> Dict[str, Union[List
         item["images"] = torch.as_tensor(
             self.apply_image(item["images"], self.target_length).transpose((2, 0, 1)), dtype=torch.get_default_dtype()
         )
-        item["gt_masks"] = [torch.as_tensor(gt_mask) for gt_mask in item["gt_masks"]]
+        if "gt_masks" in item:
+            item["gt_masks"] = [torch.as_tensor(gt_mask) for gt_mask in item["gt_masks"]]
         if "bboxes" in item:
             item["bboxes"] = self.apply_boxes(item["bboxes"], item["original_size"], self.target_length)
         if "points" in item:
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index f33b8170132..ebdd7ae3783 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -64,7 +64,8 @@ def forward(
 
         total_points_scores: Tensor = torch.zeros(used_indices.max() + 1, 0, 3, device=device)
         total_bg_coords: Tensor = torch.zeros(used_indices.max() + 1, num_bg_points, 2, device=device)
-        for label in used_indices[0]:
+        for i in range(len(used_indices[0])):
+            label = used_indices[0][i]
             points_scores, bg_coords = self.get_prompt_candidates(
                 image_embeddings=image_embeddings,
                 reference_feat=reference_feats[label],
@@ -158,7 +159,7 @@ def _point_selection(
         matched_grid = fg_coords_scores.unsqueeze(1) * matched_matrix.unsqueeze(-1)
 
         # sample the highest score one of the samples that are in the same grid
-        points_scores = matched_grid[matched_grid[..., -1].argsort(dim=0, descending=True)[0]].diagonal().T
+        points_scores = matched_grid[matched_grid[..., -1].topk(k=1, dim=0, largest=True)[1][0]].diagonal().T
 
         # sort by the highest score
         points_scores = points_scores[torch.argsort(points_scores[:, -1], descending=True)]
@@ -228,6 +229,10 @@ def set_default_config(self) -> DictConfig:
                     "freeze_prompt_encoder": True,
                     "image_size": 1024,
                     "mask_threshold": 0.0,
+                    "return_single_mask": False,
+                    "use_stability_score": False,
+                    "stability_score_offset": 1.,
+                    "return_extra_metrics": False,
                 },
                 "dataset": {
                     "normalize": {
@@ -289,7 +294,7 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
         processed_prompts = [self._preprocess_prompts(tb) for tb in transformed_batch]
         
         # initialize tensors to contain reference features and prompts
-        largest_label = max([int(label.id) for pp in processed_prompts for label in pp.keys()])
+        largest_label = max([label for pp in processed_prompts for label in pp.keys()])
         self.expand_reference_info(largest_label)
         # TODO(sungchul): consider who to handle multiple reference features, currently replace it
 
@@ -304,10 +309,7 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
         
             ref_masks = torch.zeros(largest_label + 1, *map(int, original_size))
             for label, input_prompts in pp.items():
-                if label.name.lower() == "background":
-                    # skip background
-                    # TODO (sungchul): how to skip background class
-                    continue
+                # TODO (sungchul): how to skip background class
 
                 # generate reference mask
                 # TODO (sungchul): ensemble multi reference features (current : use merged masks)
@@ -356,9 +358,9 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
                     ref_feat = self._generate_masked_features(processed_embedding, ref_mask, default_threshold_reference)
                     default_threshold_reference -= 0.05
 
-                self.reference_info["reference_feats"][int(label.id)] = ref_feat.detach().cpu()
-                self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([[int(label.id)]])), dim=1), requires_grad=False)
-                ref_masks[int(label.id)] = ref_mask.detach().cpu()
+                self.reference_info["reference_feats"][label] = ref_feat.detach().cpu()
+                self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([[label]])), dim=1), requires_grad=False)
+                ref_masks[label] = ref_mask.detach().cpu()
             batch_ref_masks.append(ref_masks)
         return self.reference_info, batch_ref_masks
 
@@ -366,8 +368,8 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
     def infer(
         self, 
         batch: List[Dict[str, Any]],
-        reference_feats: Tensor,
-        used_indices: Tensor,
+        reference_feats: Union[np.ndarray, Tensor],
+        used_indices: Union[np.ndarray, Tensor],
         is_cascade: bool = False,
     ) -> List[List[DefaultDict[int, List[Tensor]]]]:
         """Zero-shot inference with reference features.
@@ -376,8 +378,8 @@ def infer(
 
         Args:
             batch (List[Dict[str, Any]]): List of dictionaries containing images and metas.
-            reference_feats (Tensor): Reference features for target prediction.
-            used_indices (Tensor): To check which indices of reference features are validate.
+            reference_feats (Union[np.ndarray, Tensor]): Reference features for target prediction. If it is np.ndarray, it will be converted to torch tensor.
+            used_indices (Union[np.ndarray, Tensor]): To check which indices of reference features are validate. If it is np.ndarray, it will be converted to torch tensor.
             is_cascade (bool): Whether use cascade inference. Defaults to False.
 
         Returns:
@@ -386,6 +388,11 @@ def infer(
                     1. Target images
                     2. Tuple of predicted masks and used points gotten by point selection
         """
+        if isinstance(reference_feats, np.ndarray):
+            reference_feats = torch.as_tensor(reference_feats, device=self.device)
+        if isinstance(used_indices, np.ndarray):
+            used_indices = torch.as_tensor(used_indices, device=self.device)
+
         # preprocess images and prompts
         transformed_batch = [self.transforms(b.copy()) for b in batch]
         
@@ -537,7 +544,7 @@ def predict_step(self, batch, batch_idx):
         results = self.infer(batch, self.reference_info["reference_feats"], self.reference_info["used_indices"])
         return [result[0] for result in results]  # tmp: only mask
 
-    def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[ScoredLabel, List[Dict[str, Tensor]]]:
+    def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[int, List[Dict[str, Tensor]]]:
         """Preprocess prompts.
 
         Currently, preprocessing for bounding boxes is only supported.
@@ -546,7 +553,7 @@ def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[ScoredLabel, List[D
             batch (Dict[str, Any]): Dictionary containing data and prompts information.
 
         Returns:
-            (defaultdict[ScoredLabel, List[Dict[str, Tensor]]]): Processed and arranged each single prompt
+            (defaultdict[int, List[Dict[str, Tensor]]]): Processed and arranged each single prompt
                 using label information as keys. Unlike other prompts, `annotation` prompts will be aggregated
                 as single annotation.
         """
@@ -557,6 +564,8 @@ def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[ScoredLabel, List[D
             if prompts is None or len(prompts) == 0:
                 continue
             for prompt, label in zip(prompts, labels):
+                if isinstance(label, ScoredLabel):
+                    label = int(label.id_)
                 # TODO (sungchul): revisit annotations and polygons
                 if prompt_name == "annotations":
                     processed_prompts[label].append({prompt_name: torch.as_tensor(prompt, device=self.device)})
@@ -581,7 +590,7 @@ def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[ScoredLabel, List[D
                             "point_labels": torch.tensor([[1]], device=self.device),
                         }})
 
-        processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x[0].id_))
+        processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x))
         return processed_prompts
     
     def _preprocess_coords(
@@ -684,60 +693,6 @@ def _postprocess_masks(
             best_idx = torch.argmax(scores[0])
         return logits[:, best_idx], masks[0, best_idx]
 
-    def _update_value(self, target: Dict[str, Any], key: str, value: Tensor) -> None:
-        """Update tensor to target dictionary.
-
-        Args:
-            target (Dict[str, Any]): Target dictionary to be updated.
-            key (str): Key to be used for update.
-            value (Tensor): Value to be used for update.
-        """
-        if key in target:
-            target[key] = torch.cat((target[key], value))
-        else:
-            target[key] = value
-
-    def _merge_prompts(
-        self,
-        label: ScoredLabel,
-        input_prompts: Dict[str, Tensor],
-        processed_prompts: Dict[ScoredLabel, List[Dict[str, Tensor]]],
-        use_only_background: bool = True,
-    ) -> Dict[str, Tensor]:
-        """Merge target prompt and other prompts.
-
-        Merge a foreground prompt and other prompts (background or prompts with other classes).
-
-        Args:
-            label (ScoredLabel): Label information. Background is 0 and other foregrounds are >= 0.
-            input_prompts (Dict[str, Tensor]): A foreground prompt to be merged with other prompts.
-            processed_prompts (Dict[ScoredLabel, List[Dict[str, Tensor]]]): The whole class-wise prompts
-                processed at _preprocess_prompts.
-            use_only_background (bool): Whether merging only background prompt, defaults to True.
-                It is applied to only point_coords.
-
-        Returns:
-            (Dict[str, Tensor]): Merged prompts.
-        """
-        merged_input_prompts = deepcopy(input_prompts)
-        for other_label, other_input_prompts in processed_prompts.items():
-            if other_label.id_ == label.id_:
-                continue
-            if (use_only_background and other_label.id_ == 0) or (not use_only_background):
-                # only add point (and scribble) prompts
-                # use_only_background=True -> background prompts are only added as background
-                # use_only_background=False -> other prompts are added as background
-                for other_input_prompt in other_input_prompts:
-                    if "point_coords" in other_input_prompt:
-                        # point, scribble
-                        self._update_value(merged_input_prompts, "point_coords", other_input_prompt.get("point_coords"))
-                        self._update_value(
-                            merged_input_prompts,
-                            "point_labels",
-                            torch.zeros_like(other_input_prompt.get("point_labels")),
-                        )
-        return merged_input_prompts
-
     def set_metrics(self) -> None:
         """Skip set_metrics unused in zero-shot learning."""
         pass
@@ -766,6 +721,7 @@ def on_predict_start(self) -> None:
 
     def training_epoch_end(self, outputs) -> None:
         """Called in the training loop at the very end of the epoch."""
+        self.reference_info["used_indices"] = Parameter(self.reference_info["used_indices"].unique().unsqueeze(0), requires_grad=False)
         if self.config.model.save_outputs:
             path_reference_info = self.path_reference_info.format(datetime.now().strftime("%Y%m%d-%H%M%S"))
             os.makedirs(os.path.dirname(path_reference_info), exist_ok=True)
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index f55e6f6bdbd..1a293313ced 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -185,6 +185,7 @@ def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type:
 
             # forward decoder to get predicted mask
             prediction = self.forward_decoder(prompt)
+            prediction["scores"] = prediction["iou_predictions"]
             metadata = {"label": label}
 
             # set annotation for eval
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 61a8b3cd7de..e48c8b88e03 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -404,29 +404,3 @@ def test_postprocess_masks(
         _, result = zero_shot_segment_anything._postprocess_masks(masks, logits, scores)
 
         assert torch.equal(result, expected)
-
-    @e2e_pytest_unit
-    @pytest.mark.parametrize("use_only_background", [True, False])
-    def test_merge_prompts(self, set_zero_shot_segment_anything, use_only_background: bool) -> None:
-        """Test _merge_prompts."""
-        zero_shot_segment_anything = set_zero_shot_segment_anything()
-
-        input_prompts = {"point_coords": torch.tensor([1]), "point_labels": torch.tensor([1])}
-        processed_prompts = {
-            MockScoredLabel(label=0): [{"point_coords": torch.tensor([0]), "point_labels": torch.tensor([0])}],
-            MockScoredLabel(label=2): [{"point_coords": torch.tensor([2]), "point_labels": torch.tensor([1])}],
-        }
-
-        merged_input_prompts = zero_shot_segment_anything._merge_prompts(
-            label=MockScoredLabel(label=1),
-            input_prompts=input_prompts,
-            processed_prompts=processed_prompts,
-            use_only_background=use_only_background,
-        )
-
-        if use_only_background:
-            assert torch.equal(merged_input_prompts.get("point_coords"), torch.tensor([1, 0]))
-            assert torch.equal(merged_input_prompts.get("point_labels"), torch.tensor([1, 0]))
-        else:
-            assert torch.equal(merged_input_prompts.get("point_coords"), torch.tensor([1, 0, 2]))
-            assert torch.equal(merged_input_prompts.get("point_labels"), torch.tensor([1, 0, 0]))
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index bf791571efd..9196204dbaa 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -138,7 +138,7 @@ def test_predict(self, mocker):
             return_value={"image_embeddings": np.empty((4, 2, 2))},
         )
         mocker_forward_decoder = mocker.patch.object(
-            OpenVINOVisualPromptingInferencer, "forward_decoder", return_value=None
+            OpenVINOVisualPromptingInferencer, "forward_decoder", return_value={"iou_predictions": 0.1}
         )
         mocker_post_process = mocker.patch.object(
             OpenVINOVisualPromptingInferencer, "post_process", return_value=(self.fake_annotation, None, None)
diff --git a/tests/unit/algorithms/visual_prompting/test_helpers.py b/tests/unit/algorithms/visual_prompting/test_helpers.py
index ac16b4cea4e..5775c7712de 100644
--- a/tests/unit/algorithms/visual_prompting/test_helpers.py
+++ b/tests/unit/algorithms/visual_prompting/test_helpers.py
@@ -17,6 +17,8 @@
     AnnotationSceneEntity,
     AnnotationSceneKind,
 )
+from unittest.mock import Mock
+from otx.api.entities.scored_label import ScoredLabel
 from otx.api.entities.color import Color
 from otx.api.entities.dataset_item import DatasetItemEntity
 from otx.api.entities.datasets import DatasetEntity
@@ -184,8 +186,9 @@ def predict_mask(self, *args, **kwargs):
 class MockScoredLabel:
     def __init__(self, label: int, name: str = "background"):
         self.name = name
-        self.id_ = label
-        self.id = label
+        self.label = Mock()
+        self.label.id_ = label
+        self.__class__ = ScoredLabel
 
 
 class MockPromptGetter(nn.Module):

From cc60bb7fa432166ce66ac7d84b4519916e3975e7 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Fri, 16 Feb 2024 23:06:20 +0900
Subject: [PATCH 12/28] precommit

---
 .../model_wrappers/openvino_models.py         |  10 +-
 .../pytorch_lightning/datasets/dataset.py     |  51 ++++--
 .../models/encoders/sam_image_encoder.py      |   3 +-
 .../visual_prompters/segment_anything.py      |  28 ++-
 .../zero_shot_segment_anything.py             | 173 +++++++++++-------
 .../visual_prompting/tasks/inference.py       |   6 +-
 .../visual_prompting/tasks/openvino.py        | 109 +++++++----
 .../model_wrappers/test_openvino_models.py    |  55 +++---
 .../callbacks/test_inference_callback.py      |  12 +-
 .../datasets/pipelines/test_transforms.py     | 135 +++++++-------
 .../models/encoders/test_sam_image_encoder.py |   5 +-
 .../visual_prompters/test_segment_anything.py |   6 +-
 .../test_zero_shot_segment_anything.py        |  90 +++++----
 .../visual_prompting/tasks/test_inference.py  |   6 +-
 .../visual_prompting/tasks/test_openvino.py   |  33 +++-
 15 files changed, 428 insertions(+), 294 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 4861f6183be..816e7db118a 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -62,7 +62,7 @@ class PromptGetter(ImageModel):
     """PromptGetter class for zero-shot visual prompting of openvino model wrapper."""
 
     __model__ = "prompt_getter"
-    
+
     def __init__(self, inference_adapter, configuration=None, preload=False):
         super().__init__(inference_adapter, configuration, preload)
 
@@ -73,7 +73,7 @@ def parameters(cls) -> Dict[str, Any]:  # noqa: D102
         parameters.update({"sim_threshold": NumericalValue(value_type=float, default_value=0.5, min=0, max=1)})
         parameters.update({"num_bg_points": NumericalValue(value_type=int, default_value=1, min=0, max=1024)})
         return parameters
-    
+
     def _get_inputs(self):
         """Defines the model inputs for images and additional info."""
         image_blob_names, image_info_blob_names = [], []
@@ -83,9 +83,7 @@ def _get_inputs(self):
             else:
                 image_info_blob_names.append(name)
         if not image_blob_names:
-            self.raise_error(
-                "Failed to identify the input for the image: no 4D input layer found"
-            )
+            self.raise_error("Failed to identify the input for the image: no 4D input layer found")
         return image_blob_names, image_info_blob_names
 
 
@@ -101,7 +99,7 @@ def __init__(
         preload: bool = False,
     ):
         super().__init__(model_adapter, configuration, preload)
-        
+
         self.mask_input = np.zeros((1, 1, 256, 256), dtype=np.float32)
         self.has_mask_input = np.zeros((1, 1), dtype=np.float32)
 
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
index 83ada16d9c6..dcf336776b1 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/dataset.py
@@ -14,7 +14,8 @@
 # See the License for the specific language governing permissions
 # and limitations under the License.
 
-from typing import Any, Dict, List, Optional, Union
+from collections import defaultdict
+from typing import Any, DefaultDict, Dict, List, Optional, Union
 
 import cv2
 import numpy as np
@@ -31,7 +32,6 @@
     ResizeLongestSide,
     collate_fn,
 )
-from collections import defaultdict
 from otx.api.entities.dataset_item import DatasetItemEntity
 from otx.api.entities.datasets import DatasetEntity
 from otx.api.entities.image import Image
@@ -158,7 +158,7 @@ def __init__(
         self.transform = get_transform(image_size, mean, std)
         self.offset_bbox = offset_bbox
         self.labels = dataset.get_labels()
-        
+
         if not use_bbox and not use_point:
             # if both are False, use bbox as default
             use_bbox = True
@@ -179,7 +179,12 @@ def __len__(self) -> int:
         return len(self.dataset)
 
     @staticmethod
-    def get_prompts(dataset_item: DatasetItemEntity, dataset_labels: List[LabelEntity], prob: float = 1., mode: Subset = Subset.TESTING) -> Dict[str, Any]:
+    def get_prompts(
+        dataset_item: DatasetItemEntity,
+        dataset_labels: List[LabelEntity],
+        prob: float = 1.0,
+        mode: Subset = Subset.TESTING,
+    ) -> Dict[str, Any]:
         """Get propmts from dataset_item.
 
         Args:
@@ -195,7 +200,7 @@ def get_prompts(dataset_item: DatasetItemEntity, dataset_labels: List[LabelEntit
         bboxes: List[np.ndarray] = []
         points: List[np.ndarray] = []
         gt_masks: List[np.ndarray] = []
-        labels: defaultdict[str, List[ScoredLabel]] = defaultdict(list)
+        labels: DefaultDict[str, List[ScoredLabel]] = defaultdict(list)
         for annotation in dataset_item.get_annotations(labels=dataset_labels, include_empty=False, preserve_id=True):
             if isinstance(annotation.shape, Image):
                 # use mask as-is
@@ -286,7 +291,7 @@ def __getitem__(self, index: int) -> Dict[str, Union[int, List, Tensor]]:
 
         prompts = self.get_prompts(dataset_item, self.labels, self.prob)
         item.update({**prompts, "path": dataset_item.media.path})
-            
+
         return item
 
 
@@ -388,7 +393,7 @@ def setup(self, stage: Optional[str] = None) -> None:
                 image_size=image_size,
                 mean=mean,
                 std=std,
-                **self.kwargs
+                **self.kwargs,
             )
 
     def summary(self):
@@ -402,58 +407,66 @@ def summary(self):
                 num_items,
             )
 
-    def train_dataloader(self) -> Union[DataLoader, List[DataLoader], Dict[str, DataLoader]]:
+    def train_dataloader(self) -> DataLoader:
         """Train Dataloader.
 
         Returns:
-            Union[DataLoader, List[DataLoader], Dict[str, DataLoader]]: Train dataloader.
+            DataLoader: Train dataloader.
         """
         return DataLoader(
             self.train_dataset,
             shuffle=True,
             batch_size=self.config.train_batch_size,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
+            collate_fn=collate_fn
+            if self.train_type != TrainType.Zeroshot
+            else lambda x: x,  # type: ignore[return-value]
         )
 
-    def val_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
+    def val_dataloader(self) -> DataLoader:
         """Validation Dataloader.
 
         Returns:
-            Union[DataLoader, List[DataLoader]]: Validation Dataloader.
+            DataLoader: Validation Dataloader.
         """
         return DataLoader(
             self.val_dataset,
             shuffle=False,
             batch_size=self.config.val_batch_size,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
+            collate_fn=collate_fn
+            if self.train_type != TrainType.Zeroshot
+            else lambda x: x,  # type: ignore[return-value]
         )
 
-    def test_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
+    def test_dataloader(self) -> DataLoader:
         """Test Dataloader.
 
         Returns:
-            Union[DataLoader, List[DataLoader]]: Test Dataloader.
+            DataLoader: Test Dataloader.
         """
         return DataLoader(
             self.test_dataset,
             shuffle=False,
             batch_size=self.config.test_batch_size,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
+            collate_fn=collate_fn
+            if self.train_type != TrainType.Zeroshot
+            else lambda x: x,  # type: ignore[return-value]
         )
 
-    def predict_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
+    def predict_dataloader(self) -> DataLoader:
         """Predict Dataloader.
 
         Returns:
-            Union[DataLoader, List[DataLoader]]: Predict Dataloader.
+            DataLoader: Predict Dataloader.
         """
         return DataLoader(
             self.predict_dataset,
             shuffle=False,
             batch_size=1,
             num_workers=self.config.num_workers,
-            collate_fn=collate_fn if self.train_type != TrainType.Zeroshot else lambda x: x,
+            collate_fn=collate_fn
+            if self.train_type != TrainType.Zeroshot
+            else lambda x: x,  # type: ignore[return-value]
         )
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py
index d56df51fa5f..6944754c660 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/sam_image_encoder.py
@@ -19,8 +19,9 @@ class SAMImageEncoder(nn.Module):
     Args:
         config (DictConfig): Config for image encoder.
     """
-    
+
     def __new__(cls, config: DictConfig):
+        """Initialize SAM image encoder to the target backbone."""
         if "tiny_vit" == config.backbone:
             return build_tiny_vit(config.image_size)
         elif "vit" in config.backbone:
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index 9ac772ce4ad..f50afe06c8f 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -6,11 +6,10 @@
 
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
-#
 
-import re
+
 from collections import OrderedDict
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Dict, List, Optional, Tuple
 
 import torch
 from omegaconf import DictConfig
@@ -139,6 +138,7 @@ def load_checkpoint(self, state_dict: Optional[OrderedDict] = None) -> None:
         Args:
             state_dict (Optional[OrderedDict], optional): State dict of SAM. Defaults to None.
         """
+
         def skip_unused_parameters(state_dict):
             if self.config.model.backbone == "tiny_vit":
                 for key in [
@@ -149,7 +149,7 @@ def skip_unused_parameters(state_dict):
                 ]:
                     if key in state_dict:
                         state_dict.pop(key)
-        
+
         if state_dict:
             # state_dict from args.load_from
             skip_unused_parameters(state_dict)
@@ -326,7 +326,7 @@ def select_masks(self, masks: Tensor, iou_preds: Tensor, num_points: int) -> Tup
         iou_preds = iou_preds[torch.arange(masks.shape[0]), best_idx].unsqueeze(1)
 
         return masks, iou_preds
-    
+
     @classmethod
     def postprocess_masks(cls, masks: Tensor, input_size: int, orig_size: Tensor) -> Tensor:
         """Postprocess the predicted masks.
@@ -342,7 +342,7 @@ def postprocess_masks(cls, masks: Tensor, input_size: int, orig_size: Tensor) ->
         """
         masks = F.interpolate(masks, size=(input_size, input_size), mode="bilinear", align_corners=False)
 
-        prepadded_size = cls.get_prepadded_size(cls, orig_size, input_size)
+        prepadded_size = cls.get_prepadded_size(cls, orig_size, input_size)  # type: ignore[arg-type]
         masks = masks[..., : prepadded_size[0], : prepadded_size[1]]
 
         orig_size = orig_size.to(torch.int64)
@@ -391,13 +391,15 @@ def forward_train(
             for idx_prompt, prompt in enumerate([bboxes[idx], points[idx]]):
                 if prompt is None:
                     continue
-                
+
                 sparse_embeddings, dense_embeddings = self.prompt_encoder(
-                    points=(prompt.unsqueeze(1), torch.ones(len(prompt), 1, device=prompt.device)) if idx_prompt == 1 else None,
+                    points=(prompt.unsqueeze(1), torch.ones(len(prompt), 1, device=prompt.device))
+                    if idx_prompt == 1
+                    else None,
                     boxes=prompt if idx_prompt == 0 else None,
                     masks=None,
                 )
-                
+
                 _low_res_masks, _iou_predictions = self.mask_decoder(
                     image_embeddings=embedding.unsqueeze(0),
                     image_pe=self.prompt_encoder.get_dense_pe(),
@@ -439,9 +441,7 @@ def training_step(self, batch, batch_idx) -> Tensor:
 
         num_masks = sum(len(pred_mask) for pred_mask in pred_masks)
         for i, (pred_mask, gt_mask, iou_prediction) in enumerate(zip(pred_masks, gt_masks, iou_predictions)):
-            pred_mask = self.postprocess_masks(
-                pred_mask, self.config.model.image_size, batch["original_size"][i]
-            )
+            pred_mask = self.postprocess_masks(pred_mask, self.config.model.image_size, batch["original_size"][i])
             pred_mask = pred_mask.sigmoid().squeeze(1)
             self.train_metrics["train_IoU"].update(pred_mask, gt_mask)
             self.train_metrics["train_F1"].update(pred_mask, gt_mask)
@@ -497,9 +497,7 @@ def validation_step(self, batch, batch_idx) -> MetricCollection:
 
         pred_masks, _ = self.forward_train(images, bboxes, points)
         for i, (pred_mask, gt_mask) in enumerate(zip(pred_masks, gt_masks)):
-            pred_mask = self.postprocess_masks(
-                pred_mask, self.config.model.image_size, batch["original_size"][i]
-            )
+            pred_mask = self.postprocess_masks(pred_mask, self.config.model.image_size, batch["original_size"][i])
             pred_mask = pred_mask.sigmoid().squeeze(1)
             for k, v in self.val_metrics.items():
                 v.update(pred_mask, gt_mask)
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index ebdd7ae3783..a7c3181e6b6 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -3,28 +3,28 @@
 # Copyright (C) 2023 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
-from collections import OrderedDict, defaultdict
-import os
 import json
+import os
+import pickle
+from collections import OrderedDict, defaultdict
 from copy import deepcopy
+from datetime import datetime
 from itertools import product
-import pickle
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Union
 
-import torch
-import numpy as np
 import cv2
+import numpy as np
+import torch
 from omegaconf import DictConfig
-from torch import nn, Tensor
+from torch import Tensor, nn
 from torch.nn import Parameter, ParameterDict
 from torch.nn import functional as F
 
+from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.dataset import get_transform
 from otx.api.entities.scored_label import ScoredLabel
 from otx.utils.logger import get_logger
-from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.dataset import get_transform
 
 from .segment_anything import SegmentAnything
-from datetime import datetime
 
 logger = get_logger()
 
@@ -74,7 +74,7 @@ def forward(
                 num_bg_points=num_bg_points,
                 device=device,
             )
-                
+
             pad_size = torch.as_tensor(points_scores.shape[0] - total_points_scores.shape[1])
             pad_tot = torch.max(self.zero_tensor, pad_size)
             pad_cur = torch.max(self.zero_tensor, -pad_size)
@@ -92,7 +92,7 @@ def get_prompt_candidates(
         image_embeddings: Tensor,
         reference_feat: Tensor,
         original_size: Tensor,
-        threshold: Union[Tensor, float] = 0.,
+        threshold: Union[Tensor, float] = 0.0,
         num_bg_points: Union[Tensor, int] = 1,
         device: Union[torch.device, str] = torch.device("cpu"),
     ) -> Tuple[Tensor, Tensor]:
@@ -120,7 +120,7 @@ def _point_selection(
         self,
         mask_sim: Tensor,
         original_size: Tensor,
-        threshold: Union[Tensor, float] = 0.,
+        threshold: Union[Tensor, float] = 0.0,
         num_bg_points: Union[Tensor, int] = 1,
     ) -> Tuple[Tensor, Tensor]:
         """Select point used as point prompts."""
@@ -135,7 +135,7 @@ def _point_selection(
 
         point_coords = torch.where(mask_sim > threshold)
         fg_coords_scores = torch.stack(point_coords[::-1] + (mask_sim[point_coords],), dim=0).T
-        
+
         # to handle empty tensor
         len_fg_coords_scores = len(fg_coords_scores)
         fg_coords_scores = F.pad(fg_coords_scores, (0, 0, 0, max(0, 1 - len_fg_coords_scores)), value=-1)
@@ -170,11 +170,20 @@ def _point_selection(
 class ZeroShotSegmentAnything(SegmentAnything):
     """Zero-shot learning module using Segment Anything."""
 
-    def __init__(self, config: Optional[DictConfig] = None, manual_config_update: Optional[Dict] = None, state_dict: Optional[OrderedDict] = None) -> None:
+    def __init__(
+        self,
+        config: Optional[DictConfig] = None,
+        manual_config_update: Optional[Dict] = None,
+        state_dict: Optional[OrderedDict] = None,
+    ) -> None:
         if config is None:
             config = self.set_default_config()
-            
-        if manual_config_update is not None and isinstance(manual_config_update, dict) and len(manual_config_update) > 0:
+
+        if (
+            manual_config_update is not None
+            and isinstance(manual_config_update, dict)
+            and len(manual_config_update) > 0
+        ):
             for k, v in manual_config_update.items():
                 exec(f"config.{k} = {v}")
 
@@ -196,17 +205,18 @@ def __init__(self, config: Optional[DictConfig] = None, manual_config_update: Op
 
         self.point_labels_box = torch.as_tensor([[2, 3]], dtype=torch.float32)
         self.has_mask_inputs = [torch.as_tensor([[0.0]]), torch.as_tensor([[1.0]])]
-        
+
         self.transforms = get_transform(
-            image_size=config.model.image_size,
-            mean=config.dataset.normalize.mean,
-            std=config.dataset.normalize.std)
-        
+            image_size=config.model.image_size, mean=config.dataset.normalize.mean, std=config.dataset.normalize.std
+        )
+
         self.path_reference_info = "vpm_zsl_reference_infos/{}/reference_info.pt"
-        
-    def load_state_dict_pre_hook(self, state_dict: dict[str, Any], prefix: str = "", *args, **kwargs) -> None:
+
+    def load_state_dict_pre_hook(self, state_dict: Dict[str, Any], prefix: str = "", *args, **kwargs) -> None:
         """Load reference info manually."""
-        _reference_feats: Tensor = state_dict.get("reference_info.reference_feats", torch.as_tensor([], dtype=torch.float32))
+        _reference_feats: Tensor = state_dict.get(
+            "reference_info.reference_feats", torch.as_tensor([], dtype=torch.float32)
+        )
         _used_indices: Tensor = state_dict.get("reference_info.used_indices", torch.as_tensor([], dtype=torch.int64))
         self.reference_info = ParameterDict(
             {
@@ -231,7 +241,7 @@ def set_default_config(self) -> DictConfig:
                     "mask_threshold": 0.0,
                     "return_single_mask": False,
                     "use_stability_score": False,
-                    "stability_score_offset": 1.,
+                    "stability_score_offset": 1.0,
                     "return_extra_metrics": False,
                 },
                 "dataset": {
@@ -239,10 +249,10 @@ def set_default_config(self) -> DictConfig:
                         "mean": [123.675, 116.28, 103.53],
                         "std": [58.395, 57.12, 57.375],
                     }
-                }
+                },
             }
         )
-        
+
     def set_empty_reference_info(self) -> None:
         """Set empty reference information."""
         reference_feats: Parameter = Parameter(torch.as_tensor([], dtype=torch.float32), requires_grad=False)
@@ -254,13 +264,13 @@ def set_empty_reference_info(self) -> None:
             },
         )
         self.is_reference_info_empty = True
-        
+
     def initialize_reference_info(self) -> None:
         """Initialize reference information."""
         self.reference_info["reference_feats"] = Parameter(torch.zeros(0, 1, 256), requires_grad=False)
         self.reference_info["used_indices"] = Parameter(torch.as_tensor([[]], dtype=torch.int64), requires_grad=False)
         self.is_reference_info_empty = False
-        
+
     def expand_reference_info(self, new_largest_label: int) -> None:
         """Expand reference info dimensions if newly given processed prompts have more lables."""
         if new_largest_label > (cur_largest_label := len(self.reference_info["reference_feats"]) - 1):
@@ -282,17 +292,17 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
             reset_feat (bool): Whether reset reference_info.
                 For OTX standalone, resetting reference_info will be conducted in on_train_start.
                 For other frameworks, setting it to True is required to reset reference_info. Defaults to False.
-            
+
         Returns:
             (Tuple[ParameterDict, Tensor]): reference_info and ref_masks.
         """
         if reset_feat:
             self.initialize_reference_info()
-            
+
         # preprocess images and prompts
         transformed_batch = [self.transforms(b.copy()) for b in batch]
         processed_prompts = [self._preprocess_prompts(tb) for tb in transformed_batch]
-        
+
         # initialize tensors to contain reference features and prompts
         largest_label = max([label for pp in processed_prompts for label in pp.keys()])
         self.expand_reference_info(largest_label)
@@ -301,12 +311,12 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
         batch_ref_masks: List[Tensor] = []
         for tb, pp in zip(transformed_batch, processed_prompts):
             # assign components
-            images = tb["images"].unsqueeze(0).to(self.device)
+            images = tb["images"].unsqueeze(0).to(self.device)  # type: ignore[union-attr]
             original_size = torch.as_tensor(tb["original_size"])
-        
+
             image_embeddings = self.image_encoder(images)
             processed_embedding = image_embeddings.squeeze().permute(1, 2, 0)
-        
+
             ref_masks = torch.zeros(largest_label + 1, *map(int, original_size))
             for label, input_prompts in pp.items():
                 # TODO (sungchul): how to skip background class
@@ -325,10 +335,10 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
                             mask_from_polygon = cv2.drawContours(mask_from_polygon, np.asarray([contour]), 0, 1, -1)
                             ref_mask[mask_from_polygon == 1] += 1
                     elif (prompt := input_prompt.get("scribble_annotation", None)) is not None:
-                        logger.warning(f"scribble_annotation is not supported yet.")
+                        logger.warning("scribble_annotation is not supported yet.")
                         continue
                     elif (prompt := input_prompt.get("scribble_polygon", None)) is not None:
-                        logger.warning(f"scribble_polygon is not supported yet.")
+                        logger.warning("scribble_polygon is not supported yet.")
                         continue
                     else:
                         point_coords = []
@@ -355,18 +365,23 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
                 default_threshold_reference = deepcopy(self.prompt_getter.default_threshold_reference)
                 while ref_feat is None:
                     logger.info(f"[*] default_threshold_reference : {default_threshold_reference:.4f}")
-                    ref_feat = self._generate_masked_features(processed_embedding, ref_mask, default_threshold_reference)
+                    ref_feat = self._generate_masked_features(
+                        processed_embedding, ref_mask, default_threshold_reference
+                    )
                     default_threshold_reference -= 0.05
 
                 self.reference_info["reference_feats"][label] = ref_feat.detach().cpu()
-                self.reference_info["used_indices"] = Parameter(torch.cat((self.reference_info["used_indices"], torch.as_tensor([[label]])), dim=1), requires_grad=False)
+                self.reference_info["used_indices"] = Parameter(
+                    torch.cat((self.reference_info["used_indices"], torch.as_tensor([[label]])), dim=1),
+                    requires_grad=False,
+                )
                 ref_masks[label] = ref_mask.detach().cpu()
             batch_ref_masks.append(ref_masks)
         return self.reference_info, batch_ref_masks
 
     @torch.no_grad()
     def infer(
-        self, 
+        self,
         batch: List[Dict[str, Any]],
         reference_feats: Union[np.ndarray, Tensor],
         used_indices: Union[np.ndarray, Tensor],
@@ -378,8 +393,10 @@ def infer(
 
         Args:
             batch (List[Dict[str, Any]]): List of dictionaries containing images and metas.
-            reference_feats (Union[np.ndarray, Tensor]): Reference features for target prediction. If it is np.ndarray, it will be converted to torch tensor.
-            used_indices (Union[np.ndarray, Tensor]): To check which indices of reference features are validate. If it is np.ndarray, it will be converted to torch tensor.
+            reference_feats (Union[np.ndarray, Tensor]): Reference features for target prediction.
+                If it is np.ndarray, it will be converted to torch tensor.
+            used_indices (Union[np.ndarray, Tensor]): To check which indices of reference features are validate.
+                If it is np.ndarray, it will be converted to torch tensor.
             is_cascade (bool): Whether use cascade inference. Defaults to False.
 
         Returns:
@@ -395,19 +412,20 @@ def infer(
 
         # preprocess images and prompts
         transformed_batch = [self.transforms(b.copy()) for b in batch]
-        
+
         total_results: List[List[Tensor]] = []
         for tb in transformed_batch:
             # assign components
-            images = tb["images"].unsqueeze(0).to(self.device)
+            images = tb["images"].unsqueeze(0).to(self.device)  # type: ignore[union-attr]
             original_size = torch.as_tensor(tb["original_size"])
-            
+
             image_embeddings = self.image_encoder(images)
             total_points_scores, total_bg_coords = self.prompt_getter(
                 image_embeddings=image_embeddings,
                 reference_feats=reference_feats,
                 used_indices=used_indices,
-                original_size=original_size)
+                original_size=original_size,
+            )
             predicted_masks: defaultdict = defaultdict(list)
             used_points: defaultdict = defaultdict(list)
             for label, (points_scores, bg_coords) in enumerate(zip(total_points_scores, total_bg_coords)):
@@ -491,6 +509,7 @@ def _predict_masks(
         is_cascade: bool = True,
     ) -> Tensor:
         """Predict target masks."""
+        masks: Tensor
         logits: Tensor
         scores: Tensor
         for i in range(3):
@@ -517,9 +536,12 @@ def _predict_masks(
                 coords = torch.nonzero(masks)
                 y, x = coords[:, 0], coords[:, 1]
                 box_coords = self._preprocess_coords(
-                    torch.as_tensor([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device),
+                    torch.as_tensor(
+                        [[[x.min(), y.min()], [x.max(), y.max()]]], dtype=torch.float32, device=self.device
+                    ),
                     original_size,
-                    self.config.model.image_size)
+                    self.config.model.image_size,
+                )
                 point_coords = torch.cat((point_coords, box_coords), dim=1)
                 point_labels = torch.cat((point_labels, self.point_labels_box.to(self.device)), dim=1)
 
@@ -544,7 +566,7 @@ def predict_step(self, batch, batch_idx):
         results = self.infer(batch, self.reference_info["reference_feats"], self.reference_info["used_indices"])
         return [result[0] for result in results]  # tmp: only mask
 
-    def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[int, List[Dict[str, Tensor]]]:
+    def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[Any, Any]:
         """Preprocess prompts.
 
         Currently, preprocessing for bounding boxes is only supported.
@@ -553,7 +575,7 @@ def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[int, List[Dict[str,
             batch (Dict[str, Any]): Dictionary containing data and prompts information.
 
         Returns:
-            (defaultdict[int, List[Dict[str, Tensor]]]): Processed and arranged each single prompt
+            (Dict[Any, Any]): Processed and arranged each single prompt
                 using label information as keys. Unlike other prompts, `annotation` prompts will be aggregated
                 as single annotation.
         """
@@ -578,25 +600,31 @@ def _preprocess_prompts(self, batch: Dict[str, Any]) -> Dict[int, List[Dict[str,
                         masks.append(mask_from_polygon)
                     processed_prompts[label].append({prompt_name: torch.tensor(prompt, device=self.device)})
                 elif prompt_name == "bboxes":
-                    processed_prompts[label].append({
-                        prompt_name: {
-                            "point_coords": torch.as_tensor(prompt.reshape(-1, 2, 2), device=self.device),
-                            "point_labels": torch.tensor([[1]], device=self.device),
-                        }})
+                    processed_prompts[label].append(
+                        {
+                            prompt_name: {
+                                "point_coords": torch.as_tensor(prompt.reshape(-1, 2, 2), device=self.device),
+                                "point_labels": torch.tensor([[1]], device=self.device),
+                            }
+                        }
+                    )
                 elif prompt_name == "points":
-                    processed_prompts[label].append({
-                        prompt_name: {
-                            "point_coords": torch.as_tensor(prompt.reshape(-1, 2), device=self.device),
-                            "point_labels": torch.tensor([[1]], device=self.device),
-                        }})
+                    processed_prompts[label].append(
+                        {
+                            prompt_name: {
+                                "point_coords": torch.as_tensor(prompt.reshape(-1, 2), device=self.device),
+                                "point_labels": torch.tensor([[1]], device=self.device),
+                            }
+                        }
+                    )
 
-        processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x))
+        processed_prompts = dict(sorted(processed_prompts.items(), key=lambda x: x))  # type: ignore[assignment]
         return processed_prompts
-    
+
     def _preprocess_coords(
         self,
         coords: Tensor,
-        ori_shape: Union[list[int], tuple[int, int], Tensor],
+        ori_shape: Union[List[int], Tuple[int, int], Tensor],
         target_length: int,
     ) -> Tensor:
         """Expects a torch tensor of length 2 in the final dimension.
@@ -605,7 +633,7 @@ def _preprocess_coords(
 
         Args:
             coords (Tensor): Coordinates tensor.
-            ori_shape (Union[list[int], tuple[int, int], Tensor]): Original size of image.
+            ori_shape (Union[List[int], Tuple[int, int], Tensor]): Original size of image.
             target_length (int): The length of the longest side of the image.
 
         Returns:
@@ -700,17 +728,18 @@ def set_metrics(self) -> None:
     def configure_optimizers(self) -> None:
         """Skip configure_optimizers unused in zero-shot learning."""
         pass
-    
+
     def _find_latest_reference_info(self) -> Union[str, None]:
         """Find latest reference info to be used."""
         if len(stamps := sorted(os.listdir("vpm_zsl_reference_infos"), reverse=True)) > 0:
             return stamps[0]
         self.initialize_reference_info()
+        return None
 
     def on_train_start(self) -> None:
         """Called at the beginning of training after sanity check."""
         self.initialize_reference_info()
-    
+
     def on_predict_start(self) -> None:
         """Called at the beginning of predicting."""
         if (latest_stamp := self._find_latest_reference_info()) is not None:
@@ -721,11 +750,19 @@ def on_predict_start(self) -> None:
 
     def training_epoch_end(self, outputs) -> None:
         """Called in the training loop at the very end of the epoch."""
-        self.reference_info["used_indices"] = Parameter(self.reference_info["used_indices"].unique().unsqueeze(0), requires_grad=False)
+        self.reference_info["used_indices"] = Parameter(
+            self.reference_info["used_indices"].unique().unsqueeze(0), requires_grad=False
+        )
         if self.config.model.save_outputs:
             path_reference_info = self.path_reference_info.format(datetime.now().strftime("%Y%m%d-%H%M%S"))
             os.makedirs(os.path.dirname(path_reference_info), exist_ok=True)
             torch.save(self.reference_info, path_reference_info)
-            pickle.dump({k: v.numpy() for k, v in self.reference_info.items()}, open(path_reference_info.replace(".pt", ".pickle"), "wb"))
-            json.dump(repr(self.trainer.datamodule.train_dataset.dataset), open(path_reference_info.replace("reference_info.pt", "reference_meta.json"), "w"))
+            pickle.dump(
+                {k: v.numpy() for k, v in self.reference_info.items()},
+                open(path_reference_info.replace(".pt", ".pickle"), "wb"),
+            )
+            json.dump(
+                repr(self.trainer.datamodule.train_dataset.dataset),
+                open(path_reference_info.replace("reference_info.pt", "reference_meta.json"), "w"),
+            )
             logger.info(f"Saved reference info at {path_reference_info}")
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index f9c705a2489..a9b6105bb12 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -179,7 +179,7 @@ def get_model(config: DictConfig, train_type: TrainType, state_dict: Optional[Or
                         SegmentAnything as VisualPrompter,
                     )
                 elif train_type == TrainType.Zeroshot:
-                    from otx.algorithms.visual_prompting.adapters.pytorch_lightning.models import (
+                    from otx.algorithms.visual_prompting.adapters.pytorch_lightning.models import (  # type: ignore[assignment] # noqa: E501
                         ZeroShotSegmentAnything as VisualPrompter,
                     )
 
@@ -305,7 +305,7 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "point_labels": torch.randint(low=0, high=4, size=(1, 2), dtype=torch.float32),
                     "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float32),
                     "has_mask_input": torch.tensor([[1]], dtype=torch.float32),
-                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.int64)
+                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.int64),
                 }
                 output_names = ["upscaled_masks", "iou_predictions", "low_res_masks"]
                 model_to_export = self.model
@@ -671,7 +671,7 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                     "point_labels": torch.randint(low=0, high=4, size=(1, 2), dtype=torch.float32),
                     "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float32),
                     "has_mask_input": torch.tensor([[1]], dtype=torch.float32),
-                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.int64)
+                    "orig_size": torch.randint(low=256, high=2048, size=(1, 2), dtype=torch.int64),
                 }
                 output_names = ["upscaled_masks", "iou_predictions", "low_res_masks"]
                 model_to_export = self.model
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 1a293313ced..9db584f6af2 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -15,9 +15,9 @@
 # and limitations under the License.
 
 import io
-import pickle
 import json
 import os
+import pickle
 import random
 import tempfile
 import time
@@ -150,13 +150,17 @@ def __init__(
         self.transform = get_transform()  # TODO (sungchul): insert args
 
     def pre_process(
-        self, dataset_item: DatasetItemEntity, extra_processing: bool = False, use_bbox: bool = False, use_point: bool = False,
+        self,
+        dataset_item: DatasetItemEntity,
+        extra_processing: bool = False,
+        use_bbox: bool = False,
+        use_point: bool = False,
     ) -> Tuple[Dict[str, Any], Dict[str, Any], List[Dict[str, Any]]]:
         """Pre-process function of OpenVINO Visual Prompting Inferencer for image encoder."""
         if use_bbox and use_point:
             logger.warning("If both use_bbox and use_point are set, bboxes and points will be generated randomly.")
-            
-        prob = 1. if not use_point else 0. if not use_bbox and use_point else 0.5
+
+        prob = 1.0 if not use_point else 0.0 if not use_bbox and use_point else 0.5
         images, meta = self.model["image_encoder"].preprocess(dataset_item.numpy, extra_processing)
         prompts = OTXVisualPromptingDataset.get_prompts(dataset_item, self.labels, prob=prob)
         prompts = self.model["decoder"].preprocess(prompts, meta)
@@ -289,12 +293,13 @@ def __init__(
 
         self.point_labels_box = np.array([[2, 3]], dtype=np.float32)
         self.has_mask_inputs = [np.array([[0.0]]), np.array([[1.0]])]
-        
+
         self.reference_feats = None
         self.used_indices = None
-        self.path_reference_info = "vpm_zsl_reference_infos/{}/reference_info.pickle"
 
-    def pre_process_image_encoder(self, inputs: np.ndarray, extra_processing: bool = False) -> Tuple[Dict[str, np.ndarray], Dict[str, Any]]:
+    def pre_process_image_encoder(
+        self, inputs: np.ndarray, extra_processing: bool = False
+    ) -> Tuple[Dict[str, np.ndarray], Dict[str, Any]]:
         """Pre-process function of OpenVINO Zero-shot Visual Prompting Inferencer for image encoder."""
         return self.model["image_encoder"].preprocess(inputs, extra_processing)
 
@@ -315,11 +320,13 @@ def pre_process_prompt_getter(
         }
         inputs_prompt_getter.update(image_embeddings)
         return inputs_prompt_getter
-    
+
     def learn(self, images: np.ndarray):
-        """Learn"""
-    
-    def infer(self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray) -> List:
+        """Learn."""
+
+    def infer(
+        self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray
+    ) -> Tuple[List[Any], DefaultDict[Any, Any], DefaultDict[Any, Any]]:
         """Perform a prediction for a given input image."""
         # forward image encoder
         images, meta = self.pre_process_image_encoder(images)
@@ -327,7 +334,9 @@ def infer(self, images: np.ndarray, reference_feats: np.ndarray, used_indices: n
         image_embeddings = self.forward_image_encoder(images)
 
         # get point candidates
-        inputs_prompt_getter = self.pre_process_prompt_getter(image_embeddings, reference_feats, used_indices, original_size)
+        inputs_prompt_getter = self.pre_process_prompt_getter(
+            image_embeddings, reference_feats, used_indices, original_size
+        )
         total_prompts = self.forward_prompt_getter(inputs_prompt_getter)
 
         annotations: DefaultDict = defaultdict(list)
@@ -355,7 +364,8 @@ def infer(self, images: np.ndarray, reference_feats: np.ndarray, used_indices: n
                 inputs_decoder = {
                     "point_coords": point_coords[None],
                     "point_labels": point_labels[None],
-                    "orig_size": original_size[None]}
+                    "orig_size": original_size[None],
+                }
                 inputs_decoder.update(image_embeddings)
 
                 prediction = self.forward_decoder(inputs_decoder, original_size)
@@ -373,12 +383,6 @@ def infer(self, images: np.ndarray, reference_feats: np.ndarray, used_indices: n
         self._inspect_overlapping_areas(predicted_masks, used_points, annotations)
         return sum(annotations.values(), []), predicted_masks, used_points
 
-    def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
-        """Perform a prediction for a given input image."""            
-        self._get_reference_info()
-        results = self.infer(dataset_item.numpy, self.reference_feats, self.used_indices)
-        return results[0]
-
     def forward_prompt_getter(self, inputs: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
         return self.model["prompt_getter"].infer_sync(inputs)
@@ -387,6 +391,7 @@ def forward_decoder(  # type: ignore
         self, inputs: Dict[str, np.ndarray], original_size: np.ndarray
     ) -> Dict[str, np.ndarray]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
+        masks: np.ndarray
         logits: np.ndarray
         scores: np.ndarray
         for i in range(3):
@@ -399,7 +404,7 @@ def forward_decoder(  # type: ignore
 
             elif i == 1:
                 # Cascaded Post-refinement-1
-                mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)
+                mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)  # noqa: F821
                 if masks.sum() == 0:
                     return {"upscaled_masks": masks}
 
@@ -407,7 +412,7 @@ def forward_decoder(  # type: ignore
 
             elif i == 2:
                 # Cascaded Post-refinement-2
-                mask_input, masks = self._postprocess_masks(masks, logits, scores)
+                mask_input, masks = self._postprocess_masks(masks, logits, scores)  # noqa: F821
                 if masks.sum() == 0:
                     return {"upscaled_masks": masks}
 
@@ -416,14 +421,20 @@ def forward_decoder(  # type: ignore
                 box_coords = self.model["decoder"]._apply_coords(
                     np.array([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=np.float32), original_size
                 )
-                inputs.update({
-                    "point_coords": np.concatenate((inputs["point_coords"], box_coords), axis=1),
-                    "point_labels": np.concatenate((inputs["point_labels"], self.point_labels_box), axis=1),
-                })
+                inputs.update(
+                    {
+                        "point_coords": np.concatenate((inputs["point_coords"], box_coords), axis=1),
+                        "point_labels": np.concatenate((inputs["point_labels"], self.point_labels_box), axis=1),
+                    }
+                )
 
             inputs.update({"mask_input": mask_input, "has_mask_input": has_mask_input})
             prediction = self.model["decoder"].infer_sync(inputs)
-            upscaled_masks, scores, logits = prediction["upscaled_masks"], prediction["iou_predictions"], prediction["low_res_masks"]
+            upscaled_masks, scores, logits = (
+                prediction["upscaled_masks"],
+                prediction["iou_predictions"],
+                prediction["low_res_masks"],
+            )
             masks = upscaled_masks > self.model["decoder"].mask_threshold
 
         _, masks = self._postprocess_masks(masks, logits, scores)
@@ -492,21 +503,30 @@ def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
                 other_masks.pop(jm)
                 used_points[other_label].pop(jm)
                 annotations[other_label].pop(jm)
-                
+
+    def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
+        """Perform a prediction for a given input image."""
+        if self.reference_feats is None and self.used_indices is None:
+            self.reference_feats, self.used_indices = self._get_reference_info()
+        results = self.infer(dataset_item.numpy, self.reference_feats, self.used_indices)
+        return results[0]
+
     def _find_latest_reference_info(self) -> Union[str, None]:
         """Find latest reference info to be used."""
         if len(stamps := sorted(os.listdir("vpm_zsl_reference_infos"), reverse=True)) > 0:
             return stamps[0]
         return None
-    
-    def _get_reference_info(self) -> None:
+
+    def _get_reference_info(
+        self, path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle"
+    ) -> Union[Tuple[np.ndarray, np.ndarray], None]:
         """Get reference info through loading previously saved one or running `learn`."""
-        if self.reference_feats is None and self.used_indices is None and (latest_stamp := self._find_latest_reference_info()) is not None:
+        if (latest_stamp := self._find_latest_reference_info()) is not None:
             # load previously saved reference info
-            latest_reference_info = self.path_reference_info.format(latest_stamp)
+            latest_reference_info = path_reference_info.format(latest_stamp)
             reference_info = pickle.load(open(latest_reference_info, "rb"))
-            self.reference_feats = reference_info["reference_feats"]
-            self.used_indices = reference_info["used_indices"]
+            return reference_info["reference_feats"], reference_info["used_indices"]
+        return None, None
 
 
 class OTXOpenVinoDataLoader:
@@ -519,6 +539,7 @@ def __init__(
         module_name: str,
         shuffle: bool = True,
         output_model: Optional[ModelEntity] = None,
+        **kwargs,
     ):
         self.dataset = dataset
         self.inferencer = inferencer
@@ -577,6 +598,8 @@ def __init__(
         module_name: str,
         shuffle: bool = True,
         output_model: Optional[ModelEntity] = None,
+        reference_feats: Optional[np.ndarray] = None,
+        used_indices: Optional[np.ndarray] = None,
     ):
         super().__init__(
             dataset=dataset, inferencer=inferencer, module_name=module_name, shuffle=shuffle, output_model=output_model
@@ -584,6 +607,10 @@ def __init__(
         if self.module_name == "decoder":
             self.prompt_getter = self._load_module("prompt_getter", output_model)
 
+        self.inferencer: OpenVINOZeroShotVisualPromptingInferencer
+        self.reference_feats = reference_feats
+        self.used_indices = used_indices
+
     def __getitem__(self, index: int) -> Dict[str, Any]:
         """Get item from dataset."""
         images: Dict[str, np.ndarray]
@@ -592,7 +619,7 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
             index = self.shuffler[index]
 
         items = self.dataset[index]
-        images, meta = self.inferencer.pre_process(items, extra_processing=True)  # type: ignore
+        images, meta = self.inferencer.pre_process_image_encoder(items.numpy, extra_processing=True)  # type: ignore
         original_size = np.asarray(meta["original_shape"][:2])
         _, _, h, w = images["images"].shape
         pad_width = ((0, 0), (0, 0), (0, self.target_length - h), (0, self.target_length - w))
@@ -601,7 +628,9 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
             return images
         else:
             image_embeddings = self.image_encoder(images["images"])
-            inputs_prompt_getter = self.inferencer.pre_process_prompt_getter(image_embeddings, original_size)
+            inputs_prompt_getter = self.inferencer.pre_process_prompt_getter(
+                image_embeddings, self.reference_feats, self.used_indices, original_size
+            )
             if self.module_name == "prompt_getter":
                 return inputs_prompt_getter
 
@@ -803,6 +832,7 @@ def optimize(
         optimization_parameters: Optional[OptimizationParameters] = None,
         module_names: List[str] = ["image_encoder", "decoder"],
         ov_dataloader: Type[OTXOpenVinoDataLoader] = OTXOpenVinoDataLoader,
+        **kwargs,
     ):
         """Optimize function of OpenVINOVisualPromptingTask."""
         logger.info("Start PTQ optimization")
@@ -815,7 +845,9 @@ def optimize(
         dataset = dataset.get_subset(Subset.TRAINING)
 
         for i, module_name in enumerate(module_names, 1):
-            data_loader = ov_dataloader(dataset, self.inferencer, module_name=module_name, output_model=output_model)
+            data_loader = ov_dataloader(
+                dataset, self.inferencer, module_name=module_name, output_model=output_model, **kwargs
+            )
             quantization_dataset = nncf.Dataset(data_loader, lambda data: data)
 
             with tempfile.TemporaryDirectory() as tempdir:
@@ -903,8 +935,11 @@ def optimize(
         optimization_parameters: Optional[OptimizationParameters] = None,
         module_names: List[str] = ["image_encoder", "prompt_getter", "decoder"],
         ov_dataloader: Type[OTXOpenVinoDataLoader] = OTXZeroShotOpenVinoDataLoader,
+        **kwargs,
     ):
         """Optimize function of OpenVINOZeroShotVisualPromptingTask."""
+        self.inferencer: OpenVINOZeroShotVisualPromptingInferencer
+        reference_feats, used_indices = self.inferencer._get_reference_info()
         return super().optimize(
             optimization_type=optimization_type,
             dataset=dataset,
@@ -912,4 +947,6 @@ def optimize(
             optimization_parameters=optimization_parameters,
             module_names=module_names,
             ov_dataloader=ov_dataloader,
+            reference_feats=reference_feats,
+            used_indices=used_indices,
         )
diff --git a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
index 1518dbd6131..d000c818c9f 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
@@ -56,13 +56,13 @@ def test_parameters(self):
 
         assert params.get("sim_threshold").default_value == 0.5
         assert params.get("num_bg_points").default_value == 1
-        
+
     @e2e_pytest_unit
     def test_get_inputs(self, mocker):
         """Test _get_inputs."""
         mocker.patch.object(ImageModel, "__init__")
         prompt_getter = PromptGetter("adapter")
-        
+
         prompt_getter.inputs = {
             "image_embeddings": np.ones((1, 4, 4, 3)),
             "reference_feats": np.ones((2, 1, 256)),
@@ -102,33 +102,30 @@ def test_get_outputs(self):
         assert "upscaled_masks" == results
 
     @e2e_pytest_unit
-    @pytest.mark.parametrize("prompts,expected",
-    [
-        (
-            {
-                "bboxes": [np.array([[1, 1], [2, 2]])],
-                "points": [],
-                "labels": {"bboxes": [1]},
-                "original_size": (4, 4)
-            },
-            {
-                "point_coords": (1, 2, 2),
-                "point_labels": (1, 2),
-            }
-        ),
-        (
-            {
-                "bboxes": [],
-                "points": [np.array([[1, 1]])],
-                "labels": {"points": [1]},
-                "original_size": (4, 4)
-            },
-            {
-                "point_coords": (1, 1, 2),
-                "point_labels": (1, 1),
-            }
-        )
-    ])
+    @pytest.mark.parametrize(
+        "prompts,expected",
+        [
+            (
+                {
+                    "bboxes": [np.array([[1, 1], [2, 2]])],
+                    "points": [],
+                    "labels": {"bboxes": [1]},
+                    "original_size": (4, 4),
+                },
+                {
+                    "point_coords": (1, 2, 2),
+                    "point_labels": (1, 2),
+                },
+            ),
+            (
+                {"bboxes": [], "points": [np.array([[1, 1]])], "labels": {"points": [1]}, "original_size": (4, 4)},
+                {
+                    "point_coords": (1, 1, 2),
+                    "point_labels": (1, 1),
+                },
+            ),
+        ],
+    )
     def test_preprocess(self, prompts: Dict[str, Any], expected: Dict[str, Any]):
         """Test preprocess"""
         results = self.decoder.preprocess(prompts, {})
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py
index 0bac7032fa1..8053f3dcc6c 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/callbacks/test_inference_callback.py
@@ -77,9 +77,15 @@ def test_on_predict_epoch_end(self, use_mask: bool, expected: Any):
                 {
                     "masks": [torch.Tensor([[[0, 1, 0], [1, 1, 1], [0, 1, 0]]])],
                     "iou_predictions": [torch.Tensor([[0.9]])],
-                    "labels": [{
-                        "bboxes": [ScoredLabel(label=LabelEntity("foreground", domain=Domain.VISUAL_PROMPTING), probability=0.0)],
-                    }],
+                    "labels": [
+                        {
+                            "bboxes": [
+                                ScoredLabel(
+                                    label=LabelEntity("foreground", domain=Domain.VISUAL_PROMPTING), probability=0.0
+                                )
+                            ],
+                        }
+                    ],
                 }
             ]
         ]
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
index c9e32b6f28e..80890c5a155 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/datasets/pipelines/test_transforms.py
@@ -21,76 +21,79 @@
 
 
 @e2e_pytest_unit
-@pytest.mark.parametrize("batch,expected",[
-    (
-        [
-            {
-                "index": 0,
-                "images": Tensor([1, 2, 3]),
-                "bboxes": Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]),
-                "points": torch.zeros((0, 2)),
-                "gt_masks": [Tensor([1, 2, 3])],
-                "original_size": Tensor([1, 3]),
-                "path": [],
-                "labels": [],
-            },
-            {
-                "index": 1,
-                "images": Tensor([4, 5, 6]),
-                "bboxes": Tensor([[9, 10, 11, 12]]),
-                "points": torch.zeros((0, 2)),
-                "gt_masks": [Tensor([4, 5, 6])],
-                "original_size": Tensor([1, 3]),
-                "path": [],
-                "labels": [],
-            },
-        ],
-        {
-            "index": [0, 1],
-            "images": Tensor([[1, 2, 3], [4, 5, 6]]),
-            "bboxes": [Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]), Tensor([[9, 10, 11, 12]])],
-            "points": [None, None],
-            "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
-            "original_size": [Tensor([1, 3]), Tensor([1, 3])],
-            "path": [[], []],
-            "labels": [[], []],
-        }
-    ),
-    (
-        [
+@pytest.mark.parametrize(
+    "batch,expected",
+    [
+        (
+            [
+                {
+                    "index": 0,
+                    "images": Tensor([1, 2, 3]),
+                    "bboxes": Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]),
+                    "points": torch.zeros((0, 2)),
+                    "gt_masks": [Tensor([1, 2, 3])],
+                    "original_size": Tensor([1, 3]),
+                    "path": [],
+                    "labels": [],
+                },
+                {
+                    "index": 1,
+                    "images": Tensor([4, 5, 6]),
+                    "bboxes": Tensor([[9, 10, 11, 12]]),
+                    "points": torch.zeros((0, 2)),
+                    "gt_masks": [Tensor([4, 5, 6])],
+                    "original_size": Tensor([1, 3]),
+                    "path": [],
+                    "labels": [],
+                },
+            ],
             {
-                "index": 0,
-                "images": Tensor([1, 2, 3]),
-                "bboxes": torch.zeros((0, 4)),
-                "points": Tensor([[1, 1]]),
-                "gt_masks": [Tensor([1, 2, 3])],
-                "original_size": Tensor([1, 3]),
-                "path": [],
-                "labels": [],
+                "index": [0, 1],
+                "images": Tensor([[1, 2, 3], [4, 5, 6]]),
+                "bboxes": [Tensor([[1, 2, 3, 4], [5, 6, 7, 8]]), Tensor([[9, 10, 11, 12]])],
+                "points": [None, None],
+                "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
+                "original_size": [Tensor([1, 3]), Tensor([1, 3])],
+                "path": [[], []],
+                "labels": [[], []],
             },
+        ),
+        (
+            [
+                {
+                    "index": 0,
+                    "images": Tensor([1, 2, 3]),
+                    "bboxes": torch.zeros((0, 4)),
+                    "points": Tensor([[1, 1]]),
+                    "gt_masks": [Tensor([1, 2, 3])],
+                    "original_size": Tensor([1, 3]),
+                    "path": [],
+                    "labels": [],
+                },
+                {
+                    "index": 1,
+                    "images": Tensor([4, 5, 6]),
+                    "bboxes": torch.zeros((0, 4)),
+                    "points": Tensor([[2, 2]]),
+                    "gt_masks": [Tensor([4, 5, 6])],
+                    "original_size": Tensor([1, 3]),
+                    "path": [],
+                    "labels": [],
+                },
+            ],
             {
-                "index": 1,
-                "images": Tensor([4, 5, 6]),
-                "bboxes": torch.zeros((0, 4)),
-                "points": Tensor([[2, 2]]),
-                "gt_masks": [Tensor([4, 5, 6])],
-                "original_size": Tensor([1, 3]),
-                "path": [],
-                "labels": [],
+                "index": [0, 1],
+                "images": Tensor([[1, 2, 3], [4, 5, 6]]),
+                "bboxes": [None, None],
+                "points": [Tensor([[1, 1]]), Tensor([[2, 2]])],
+                "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
+                "original_size": [Tensor([1, 3]), Tensor([1, 3])],
+                "path": [[], []],
+                "labels": [[], []],
             },
-        ],
-        {
-            "index": [0, 1],
-            "images": Tensor([[1, 2, 3], [4, 5, 6]]),
-            "bboxes": [None, None],
-            "points": [Tensor([[1, 1]]), Tensor([[2, 2]])],
-            "gt_masks": [Tensor([[1, 2, 3]]), Tensor([[4, 5, 6]])],
-            "original_size": [Tensor([1, 3]), Tensor([1, 3])],
-            "path": [[], []],
-            "labels": [[], []],
-        }
-    )
-])
+        ),
+    ],
+)
 def test_collate_fn(batch: List[Dict[str, Any]], expected: Dict[str, Any]):
     """Test collate_fn."""
     results = collate_fn(batch)
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py
index 5c2c652ed2e..66ae6958f0b 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/encoders/test_sam_image_encoder.py
@@ -25,9 +25,10 @@ class TestSAMImageEncoder:
     @pytest.fixture()
     def config(self, mocker) -> DictConfig:
         return DictConfig(dict(image_size=1024))
-        
+
     @e2e_pytest_unit
-    @pytest.mark.parametrize("backbone,expected",
+    @pytest.mark.parametrize(
+        "backbone,expected",
         [
             ("tiny_vit", "TinyViT"),
             ("vit_b", "ViT"),
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index 24ae52aa3cb..8f3f727e533 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -490,9 +490,7 @@ def test_predict_step(self, mocker, return_logits: bool, expected: Tensor) -> No
             (6, torch.tensor((8, 8)), (1, 8, 8)),
         ],
     )
-    def test_postprocess_masks(
-        self, input_size: int, original_size: Tuple[int], expected: Tuple[int]
-    ) -> None:
+    def test_postprocess_masks(self, input_size: int, original_size: Tuple[int], expected: Tuple[int]) -> None:
         """Test postprocess_masks."""
         sam = SegmentAnything(config=self.base_config)
         masks = torch.zeros((1, 1, 4, 4))
@@ -500,7 +498,7 @@ def test_postprocess_masks(
         results = sam.postprocess_masks(masks, input_size, original_size)
 
         assert results.shape[1:] == expected
-        
+
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "input_image_size,expected",
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index e48c8b88e03..047dcb4f8bf 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -45,33 +45,43 @@ def test_set_default_thresholds(self) -> None:
         assert self.prompt_getter.default_threshold_target == 0.7
 
     @e2e_pytest_unit
-    @pytest.mark.parametrize("result_point_selection", [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])])
+    @pytest.mark.parametrize(
+        "result_point_selection",
+        [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])],
+    )
     def test_forward(self, mocker, result_point_selection: torch.Tensor) -> None:
         """Test forward."""
         mocker.patch.object(
-            self.prompt_getter,
-            "get_prompt_candidates",
-            return_value=(result_point_selection, torch.zeros(1, 2)))
+            self.prompt_getter, "get_prompt_candidates", return_value=(result_point_selection, torch.zeros(1, 2))
+        )
         image_embeddings = torch.ones(1, 4, 4, 4)
         reference_feats = torch.rand(1, 1, 4)
         used_indices = torch.as_tensor([[0]])
         original_size = torch.tensor((self.prompt_getter.image_size, self.prompt_getter.image_size), dtype=torch.int64)
 
         total_points_scores, total_bg_coords = self.prompt_getter(
-            image_embeddings=image_embeddings, reference_feats=reference_feats, used_indices=used_indices, original_size=original_size
+            image_embeddings=image_embeddings,
+            reference_feats=reference_feats,
+            used_indices=used_indices,
+            original_size=original_size,
         )
-        
+
         assert total_points_scores.shape[0] == 1
         assert total_bg_coords.shape[0] == 1
 
     @e2e_pytest_unit
-    @pytest.mark.parametrize("result_point_selection", [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])])
+    @pytest.mark.parametrize(
+        "result_point_selection",
+        [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])],
+    )
     def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tensor) -> None:
         """Test get_prompt_candidates."""
         mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
         )
-        mocker.patch.object(self.prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2)))
+        mocker.patch.object(
+            self.prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2))
+        )
         image_embeddings = torch.ones(1, 4, 4, 4)
         reference_feat = torch.rand(1, 4)
         original_size = torch.tensor(
@@ -89,9 +99,13 @@ def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tenso
     @pytest.mark.parametrize(
         "mask_sim,expected",
         [
-            (torch.arange(0.1, 1.0, 0.1).reshape(3, 3), torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]])),
-            (torch.zeros(3, 3), torch.tensor([[-1, -1, -1]]))
-        ])
+            (
+                torch.arange(0.1, 1.0, 0.1).reshape(3, 3),
+                torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]),
+            ),
+            (torch.zeros(3, 3), torch.tensor([[-1, -1, -1]])),
+        ],
+    )
     def test_point_selection(self, mask_sim: torch.Tensor, expected: torch.Tensor) -> None:
         """Test _point_selection."""
         points_scores, bg_coords = self.prompt_getter._point_selection(
@@ -106,7 +120,9 @@ def test_point_selection(self, mask_sim: torch.Tensor, expected: torch.Tensor) -
 class TestZeroShotSegmentAnything:
     @pytest.fixture
     def set_zero_shot_segment_anything(self, monkeypatch):
-        def zero_shot_segment_anything(manual_config_update: Optional[Dict] = None, state_dict: Optional[OrderedDict] = None):
+        def zero_shot_segment_anything(
+            manual_config_update: Optional[Dict] = None, state_dict: Optional[OrderedDict] = None
+        ):
             monkeypatch.setattr(
                 "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SAMImageEncoder",
                 MockImageEncoder,
@@ -133,7 +149,7 @@ def test_init(self, set_zero_shot_segment_anything, state_dict: Optional[Dict[st
             state_dict = set_zero_shot_segment_anything().state_dict()
             state_dict.pop("reference_info.reference_feats")
             state_dict.pop("reference_info.used_indices")
-        
+
         zero_shot_segment_anything = set_zero_shot_segment_anything(state_dict=state_dict)
 
         assert zero_shot_segment_anything.config.model.freeze_image_encoder
@@ -177,14 +193,16 @@ def test_learn(self, mocker, set_zero_shot_segment_anything) -> None:
         )
         mocker.patch.object(zero_shot_segment_anything, "_generate_masked_features", return_value=torch.ones(1, 256))
 
-        batch = [{
-            "images": np.ones((4, 4, 3), dtype=np.uint8),
-            "gt_masks": np.ones((4, 4), dtype=np.uint8),
-            "bboxes": np.array([[0, 0, 1, 1]], dtype=np.float32),
-            "points": np.zeros((0, 2), dtype=np.float32),
-            "labels": {"bboxes": [MockScoredLabel(label=0, name="label")]},
-            "original_size": np.array([4, 4], dtype=np.int64)
-        }]
+        batch = [
+            {
+                "images": np.ones((4, 4, 3), dtype=np.uint8),
+                "gt_masks": np.ones((4, 4), dtype=np.uint8),
+                "bboxes": np.array([[0, 0, 1, 1]], dtype=np.float32),
+                "points": np.zeros((0, 2), dtype=np.float32),
+                "labels": {"bboxes": [MockScoredLabel(label=0, name="label")]},
+                "original_size": np.array([4, 4], dtype=np.int64),
+            }
+        ]
         zero_shot_segment_anything.learn(batch=batch, reset_feat=True)
 
         assert zero_shot_segment_anything.reference_info.reference_feats.shape == (1, 1, 256)
@@ -203,14 +221,18 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
         reference_feats = nn.Parameter(torch.rand(1, 1, 256), requires_grad=False)
         used_indices = nn.Parameter(torch.as_tensor([[0]], dtype=torch.int64), requires_grad=False)
         mocker.patch.object(
-            SegmentAnything, "forward", return_value=(torch.ones(1, 4, 4, 4), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
+            SegmentAnything,
+            "forward",
+            return_value=(torch.ones(1, 4, 4, 4), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4)),
         )
 
-        batch = [{
-            "images": np.ones((4, 4, 3), dtype=np.uint8),
-            "gt_masks": np.ones((4, 4), dtype=np.uint8),
-            "original_size": np.array([4, 4], dtype=np.int64)
-        }]
+        batch = [
+            {
+                "images": np.ones((4, 4, 3), dtype=np.uint8),
+                "gt_masks": np.ones((4, 4), dtype=np.uint8),
+                "original_size": np.array([4, 4], dtype=np.int64),
+            }
+        ]
         total_results = zero_shot_segment_anything.infer(
             batch=batch,
             reference_feats=reference_feats,
@@ -220,11 +242,13 @@ def test_infer(self, monkeypatch, mocker, set_zero_shot_segment_anything, expect
         for i, results in enumerate(total_results[0]):
             for _, result in results.items():
                 assert torch.equal(result[0], expected[i])
-                
+
     @e2e_pytest_unit
     def test_inspect_overlapping_areas(self, mocker, set_zero_shot_segment_anything) -> None:
         """Test _inspect_overlapping_areas."""
-        mocker.patch("otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_checkpoint")
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_checkpoint"
+        )
         zero_shot_segment_anything = set_zero_shot_segment_anything()
         predicted_masks = {
             0: [
@@ -327,7 +351,9 @@ def test_inspect_overlapping_areas(self, mocker, set_zero_shot_segment_anything)
     def test_predict_masks(self, mocker, set_zero_shot_segment_anything) -> None:
         """Test _predict_masks."""
         mocker.patch.object(
-            SegmentAnything, "forward", return_value=(torch.ones(1, 4, 8, 8), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4))
+            SegmentAnything,
+            "forward",
+            return_value=(torch.ones(1, 4, 8, 8), torch.tensor([[0.1, 0.2, 0.5, 0.7]]), torch.ones(1, 4, 4, 4)),
         )
 
         zero_shot_segment_anything = set_zero_shot_segment_anything()
@@ -348,10 +374,10 @@ def test_preprocess_prompts(self, set_zero_shot_segment_anything) -> None:
         transformed_batch = {
             "bboxes": torch.tensor([[0, 0, 1, 1]]),
             "points": torch.tensor([[2, 2]]),
-            "labels": {"bboxes": [MockScoredLabel(label=1)], "points": [MockScoredLabel(label=1)]}
+            "labels": {"bboxes": [MockScoredLabel(label=1)], "points": [MockScoredLabel(label=1)]},
         }
         processed_prompts = zero_shot_segment_anything._preprocess_prompts(transformed_batch)
-        
+
         for prompts in processed_prompts.values():
             for prompt in prompts:
                 if "bboxes" in prompt:
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index 59c5fa37215..d595227fa7c 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -381,7 +381,11 @@ def test_save_model(self, mocker):
         mocker_otx_model = mocker.patch("otx.api.entities.model.ModelEntity")
         mocker_io_bytes_io = mocker.patch("io.BytesIO")
         mocker_torch_save = mocker.patch("torch.save")
-        mocker.patch.object(self.zero_shot_task.model, "state_dict", return_value={"reference_info.reference_feats": None, "reference_info.used_indices": None})
+        mocker.patch.object(
+            self.zero_shot_task.model,
+            "state_dict",
+            return_value={"reference_info.reference_feats": None, "reference_info.used_indices": None},
+        )
 
         self.zero_shot_task.model.reference_info = "reference_info"
 
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 9196204dbaa..cdfdab44739 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -223,7 +223,9 @@ def test_predict(self, mocker):
             "forward_prompt_getter",
             return_value={"total_points_scores": np.array([[[1, 1, 1]]]), "total_bg_coords": np.array([[[2, 2]]])},
         )
-        mocker_forward_decoder = mocker.patch.object(OpenVINOZeroShotVisualPromptingInferencer, "forward_decoder", return_value={})
+        mocker_forward_decoder = mocker.patch.object(
+            OpenVINOZeroShotVisualPromptingInferencer, "forward_decoder", return_value={}
+        )
         mocker_post_process = mocker.patch.object(
             OpenVINOZeroShotVisualPromptingInferencer, "post_process", return_value=(self.fake_annotation, None, None)
         )
@@ -243,12 +245,20 @@ def test_predict(self, mocker):
         [
             (
                 (np.ones((1, 1)), np.ones((3, 3))),
-                {"upscaled_masks": np.ones((3, 3)), "iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
+                {
+                    "upscaled_masks": np.ones((3, 3)),
+                    "iou_predictions": np.array([[0.9]]),
+                    "low_res_masks": np.ones((1, 1, 2, 2)),
+                },
                 {"upscaled_masks": np.ones((3, 3))},
             ),
             (
                 (np.zeros((2, 2)), np.zeros((3, 3))),
-                {"upscaled_masks": np.zeros((3, 3)), "iou_predictions": np.array([[0.9]]), "low_res_masks": np.ones((1, 1, 2, 2))},
+                {
+                    "upscaled_masks": np.zeros((3, 3)),
+                    "iou_predictions": np.array([[0.9]]),
+                    "low_res_masks": np.ones((1, 1, 2, 2)),
+                },
                 {"upscaled_masks": np.zeros((3, 3))},
             ),
         ],
@@ -269,7 +279,9 @@ def test_forward_decoder(
             "_apply_coords",
             return_value=np.array([[[1, 1]]], dtype=np.float32),
         )
-        mocker.patch.object(self.zero_shot_visual_prompting_ov_inferencer, "_postprocess_masks", return_value=postprocess_output)
+        mocker.patch.object(
+            self.zero_shot_visual_prompting_ov_inferencer, "_postprocess_masks", return_value=postprocess_output
+        )
 
         result = self.zero_shot_visual_prompting_ov_inferencer.forward_decoder(
             inputs={
@@ -288,7 +300,7 @@ def test_forward_decoder(
         [
             (
                 np.repeat(np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]])[None], 4, axis=0)[None],
-                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_)
+                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_),
             ),
             (
                 np.concatenate(
@@ -298,7 +310,7 @@ def test_forward_decoder(
                     ),
                     axis=1,
                 ),
-                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_)
+                np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.bool_),
             ),
             (np.zeros((1, 4, 3, 3)), np.zeros((3, 3))),
         ],
@@ -309,11 +321,12 @@ def test_postprocess_masks(self, masks: np.ndarray, expected_masks: np.ndarray):
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].image_size = 3
 
         _, result_masks = self.zero_shot_visual_prompting_ov_inferencer._postprocess_masks(
-            masks=masks, logits=np.empty((1, 4, 2, 2)), scores=np.array([[0.5, 0.7, 0.8, 0.9]]))
+            masks=masks, logits=np.empty((1, 4, 2, 2)), scores=np.array([[0.5, 0.7, 0.8, 0.9]])
+        )
 
         assert result_masks.shape == (3, 3)
         assert np.all(result_masks == expected_masks)
-        
+
     @e2e_pytest_unit
     def test_inspect_overlapping_areas(self) -> None:
         """Test _inspect_overlapping_areas."""
@@ -407,7 +420,9 @@ def test_inspect_overlapping_areas(self) -> None:
             ],
         }
 
-        self.zero_shot_visual_prompting_ov_inferencer._inspect_overlapping_areas(predicted_masks, used_points, predicted_masks.copy(), threshold_iou=0.5)
+        self.zero_shot_visual_prompting_ov_inferencer._inspect_overlapping_areas(
+            predicted_masks, used_points, predicted_masks.copy(), threshold_iou=0.5
+        )
 
         assert len(predicted_masks[0]) == 1
         assert len(predicted_masks[1]) == 2

From 4eb93129c2ab2d27a4be30d50c83ef20add1ac96 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Mon, 19 Feb 2024 17:22:54 +0900
Subject: [PATCH 13/28] Fix e2e

---
 .../model_wrappers/openvino_models.py         |  18 +-
 .../visual_prompters/segment_anything.py      |   2 +
 .../zero_shot_segment_anything.py             |  81 ++---
 .../visual_prompting/tasks/inference.py       |  15 +-
 .../visual_prompting/tasks/openvino.py        | 125 ++++---
 .../annotations/instances_train.json          |  66 ++++
 .../annotations/instances_val.json            | 316 ++++++++++++++++++
 .../images/train/Slide4.PNG                   | Bin 0 -> 26804 bytes
 .../images/val/Slide3.PNG                     | Bin 0 -> 21103 bytes
 .../images/val/Slide4.PNG                     | Bin 0 -> 6123 bytes
 .../images/val/Slide5.PNG                     | Bin 0 -> 31177 bytes
 .../images/val/Slide6.PNG                     | Bin 0 -> 21277 bytes
 .../images/val/Slide7.PNG                     | Bin 0 -> 32317 bytes
 .../images/val/Slide8.PNG                     | Bin 0 -> 22874 bytes
 .../images/val/Slide9.PNG                     | Bin 0 -> 26796 bytes
 .../compressed_prompt_getter.yml              |   2 +-
 .../cli/visual_prompting/test_zero_shot.py    |  10 +-
 17 files changed, 509 insertions(+), 126 deletions(-)
 create mode 100644 tests/assets/car_tree_bug_zero_shot/annotations/instances_train.json
 create mode 100644 tests/assets/car_tree_bug_zero_shot/annotations/instances_val.json
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/train/Slide4.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide3.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide4.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide5.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide6.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide7.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide8.PNG
 create mode 100644 tests/assets/car_tree_bug_zero_shot/images/val/Slide9.PNG

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 816e7db118a..7c0e8da2b8e 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -172,23 +172,13 @@ def postprocess(self, outputs: Dict[str, np.ndarray], meta: Dict[str, Any]) -> T
 
         Returns:
             hard_prediction (np.ndarray): The hard prediction.
-            soft_prediction (np.ndarray): Resized, cropped, and normalized soft prediction.
+            soft_prediction (np.ndarray): The soft prediction.
         """
+        probability = max(min(float(outputs["scores"]), 1.0), 0.0)
+        hard_prediction = outputs[self.output_blob_name].squeeze() > self.mask_threshold
+        soft_prediction = hard_prediction * probability
 
-        def sigmoid(x):
-            return np.tanh(x * 0.5) * 0.5 + 0.5  # to avoid overflow
-
-        soft_prediction = outputs[self.output_blob_name].squeeze()
-        soft_prediction = sigmoid(soft_prediction)
         meta["soft_prediction"] = soft_prediction
-
-        hard_prediction = create_hard_prediction_from_soft_prediction(
-            soft_prediction=soft_prediction,
-            soft_threshold=self.soft_threshold,
-            blur_strength=self.blur_strength,
-        )
-
-        probability = max(min(float(outputs["scores"]), 1.0), 0.0)
         meta["label"].probability = probability
 
         return hard_prediction, soft_prediction
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index f50afe06c8f..14535b04cb2 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -351,6 +351,8 @@ def postprocess_masks(cls, masks: Tensor, input_size: int, orig_size: Tensor) ->
 
     def get_prepadded_size(self, input_image_size: Tensor, longest_side: int) -> Tensor:
         """Get pre-padded size."""
+        input_image_size = input_image_size.to(torch.float32)
+        longest_side = torch.tensor(longest_side).to(torch.float32)
         scale = longest_side / torch.max(input_image_size)
         transformed_size = scale * input_image_size
         return torch.floor(transformed_size + 0.5).to(torch.int64)
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index a7c3181e6b6..aa3206eec89 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -40,69 +40,59 @@ def __init__(self, image_size: int, downsizing: int = 64) -> None:
         self.image_size = image_size
         self.downsizing = downsizing
 
-        self.zero_tensor = torch.as_tensor(0)
-
     def set_default_thresholds(self, default_threshold_reference: float, default_threshold_target: float) -> None:
         """Set default thresholds."""
         self.default_threshold_reference = default_threshold_reference
         self.default_threshold_target = default_threshold_target
 
-    def forward(
+    def get_prompt_candidates(
         self,
         image_embeddings: Tensor,
         reference_feats: Tensor,
         used_indices: Tensor,
         original_size: Tensor,
-        threshold: Tensor = torch.as_tensor([[0.0]], dtype=torch.float32),
-        num_bg_points: Tensor = torch.as_tensor([[1]], dtype=torch.int64),
-    ) -> Tuple[Tensor, Tensor]:
+        threshold: Tensor = torch.tensor([[0.0]], dtype=torch.float32),
+        num_bg_points: Tensor = torch.tensor([[1]], dtype=torch.int64),
+        device: Union[torch.device, str] = torch.device("cpu"),
+    ) -> Tuple[Dict[int, Tensor], Dict[int, Tensor]]:
         """Get prompt candidates."""
-        device = image_embeddings.device
-        original_size = original_size.squeeze()
-        threshold = threshold.squeeze().to(device)
-        num_bg_points = num_bg_points.squeeze()
+        threshold = threshold.to(device)
 
-        total_points_scores: Tensor = torch.zeros(used_indices.max() + 1, 0, 3, device=device)
-        total_bg_coords: Tensor = torch.zeros(used_indices.max() + 1, num_bg_points, 2, device=device)
-        for i in range(len(used_indices[0])):
-            label = used_indices[0][i]
-            points_scores, bg_coords = self.get_prompt_candidates(
+        total_points_scores: Dict[int, Tensor] = {}
+        total_bg_coords: Dict[int, Tensor] = {}
+        for label in map(int, used_indices[0]):
+            points_scores, bg_coords = self(
                 image_embeddings=image_embeddings,
                 reference_feat=reference_feats[label],
                 original_size=original_size,
                 threshold=threshold,
                 num_bg_points=num_bg_points,
-                device=device,
             )
 
-            pad_size = torch.as_tensor(points_scores.shape[0] - total_points_scores.shape[1])
-            pad_tot = torch.max(self.zero_tensor, pad_size)
-            pad_cur = torch.max(self.zero_tensor, -pad_size)
-
-            total_points_scores = F.pad(total_points_scores, (0, 0, 0, pad_tot, 0, 0), value=-1)
-            points_scores = F.pad(points_scores, (0, 0, 0, pad_cur), value=-1)
-
             total_points_scores[label] = points_scores
             total_bg_coords[label] = bg_coords
 
         return total_points_scores, total_bg_coords
 
-    def get_prompt_candidates(
+    def forward(
         self,
         image_embeddings: Tensor,
         reference_feat: Tensor,
         original_size: Tensor,
-        threshold: Union[Tensor, float] = 0.0,
-        num_bg_points: Union[Tensor, int] = 1,
-        device: Union[torch.device, str] = torch.device("cpu"),
+        threshold: Tensor = torch.tensor([[0.0]], dtype=torch.float32),
+        num_bg_points: Tensor = torch.tensor([[1]], dtype=torch.int64),
     ) -> Tuple[Tensor, Tensor]:
         """Get prompt candidates from given reference and target features."""
+        original_size = original_size.squeeze()
+        threshold = threshold.squeeze()
+        num_bg_points = num_bg_points.squeeze()
+
         target_feat = image_embeddings.squeeze()
         c_feat, h_feat, w_feat = target_feat.shape
         target_feat = target_feat / target_feat.norm(dim=0, keepdim=True)
         target_feat = target_feat.reshape(c_feat, h_feat * w_feat)
 
-        sim = reference_feat.to(device) @ target_feat
+        sim = reference_feat @ target_feat
         sim = sim.reshape(1, 1, h_feat, w_feat)
         sim = ZeroShotSegmentAnything.postprocess_masks(sim, self.image_size, original_size)
 
@@ -138,7 +128,7 @@ def _point_selection(
 
         # to handle empty tensor
         len_fg_coords_scores = len(fg_coords_scores)
-        fg_coords_scores = F.pad(fg_coords_scores, (0, 0, 0, max(0, 1 - len_fg_coords_scores)), value=-1)
+        fg_coords_scores = F.pad(fg_coords_scores, (0, 0, 0, max(0, 10 - len_fg_coords_scores)), value=-1)
 
         ratio = self.image_size / original_size.max()
         width = (original_size[1] * ratio).to(torch.int64)
@@ -159,10 +149,12 @@ def _point_selection(
         matched_grid = fg_coords_scores.unsqueeze(1) * matched_matrix.unsqueeze(-1)
 
         # sample the highest score one of the samples that are in the same grid
-        points_scores = matched_grid[matched_grid[..., -1].topk(k=1, dim=0, largest=True)[1][0]].diagonal().T
+        matched_indices = matched_grid[..., -1].topk(k=1, dim=0, largest=True)[1][0].to(torch.int64)
+        points_scores = matched_grid[matched_indices].diagonal().T
 
         # sort by the highest score
-        points_scores = points_scores[torch.argsort(points_scores[:, -1], descending=True)]
+        sorted_points_scores_indices = torch.argsort(points_scores[:, -1], descending=True).to(torch.int64)
+        points_scores = points_scores[sorted_points_scores_indices]
 
         return points_scores, bg_coords
 
@@ -203,8 +195,8 @@ def __init__(
             default_threshold_target=config.model.default_threshold_target,
         )
 
-        self.point_labels_box = torch.as_tensor([[2, 3]], dtype=torch.float32)
-        self.has_mask_inputs = [torch.as_tensor([[0.0]]), torch.as_tensor([[1.0]])]
+        self.point_labels_box = torch.tensor([[2, 3]], dtype=torch.float32)
+        self.has_mask_inputs = [torch.tensor([[0.0]]), torch.tensor([[1.0]])]
 
         self.transforms = get_transform(
             image_size=config.model.image_size, mean=config.dataset.normalize.mean, std=config.dataset.normalize.std
@@ -215,9 +207,9 @@ def __init__(
     def load_state_dict_pre_hook(self, state_dict: Dict[str, Any], prefix: str = "", *args, **kwargs) -> None:
         """Load reference info manually."""
         _reference_feats: Tensor = state_dict.get(
-            "reference_info.reference_feats", torch.as_tensor([], dtype=torch.float32)
+            "reference_info.reference_feats", torch.tensor([], dtype=torch.float32)
         )
-        _used_indices: Tensor = state_dict.get("reference_info.used_indices", torch.as_tensor([], dtype=torch.int64))
+        _used_indices: Tensor = state_dict.get("reference_info.used_indices", torch.tensor([], dtype=torch.int64))
         self.reference_info = ParameterDict(
             {
                 "reference_feats": Parameter(_reference_feats, requires_grad=False),
@@ -255,8 +247,8 @@ def set_default_config(self) -> DictConfig:
 
     def set_empty_reference_info(self) -> None:
         """Set empty reference information."""
-        reference_feats: Parameter = Parameter(torch.as_tensor([], dtype=torch.float32), requires_grad=False)
-        used_indices: Parameter = Parameter(torch.as_tensor([[]], dtype=torch.int64), requires_grad=False)
+        reference_feats: Parameter = Parameter(torch.tensor([], dtype=torch.float32), requires_grad=False)
+        used_indices: Parameter = Parameter(torch.tensor([[]], dtype=torch.int64), requires_grad=False)
         self.reference_info = ParameterDict(
             {
                 "reference_feats": reference_feats,
@@ -268,7 +260,7 @@ def set_empty_reference_info(self) -> None:
     def initialize_reference_info(self) -> None:
         """Initialize reference information."""
         self.reference_info["reference_feats"] = Parameter(torch.zeros(0, 1, 256), requires_grad=False)
-        self.reference_info["used_indices"] = Parameter(torch.as_tensor([[]], dtype=torch.int64), requires_grad=False)
+        self.reference_info["used_indices"] = Parameter(torch.tensor([[]], dtype=torch.int64), requires_grad=False)
         self.is_reference_info_empty = False
 
     def expand_reference_info(self, new_largest_label: int) -> None:
@@ -372,7 +364,7 @@ def learn(self, batch: List[Dict[str, Any]], reset_feat: bool = False) -> Union[
 
                 self.reference_info["reference_feats"][label] = ref_feat.detach().cpu()
                 self.reference_info["used_indices"] = Parameter(
-                    torch.cat((self.reference_info["used_indices"], torch.as_tensor([[label]])), dim=1),
+                    torch.cat((self.reference_info["used_indices"], torch.tensor([[label]])), dim=1),
                     requires_grad=False,
                 )
                 ref_masks[label] = ref_mask.detach().cpu()
@@ -420,18 +412,19 @@ def infer(
             original_size = torch.as_tensor(tb["original_size"])
 
             image_embeddings = self.image_encoder(images)
-            total_points_scores, total_bg_coords = self.prompt_getter(
+            total_points_scores, total_bg_coords = self.prompt_getter.get_prompt_candidates(
                 image_embeddings=image_embeddings,
                 reference_feats=reference_feats,
                 used_indices=used_indices,
                 original_size=original_size,
+                device=self.device,
             )
             predicted_masks: defaultdict = defaultdict(list)
             used_points: defaultdict = defaultdict(list)
-            for label, (points_scores, bg_coords) in enumerate(zip(total_points_scores, total_bg_coords)):
+            for label in total_points_scores.keys():
+                points_scores = total_points_scores[label]
+                bg_coords = total_bg_coords[label]
                 for points_score in points_scores:
-                    if points_score[-1] == -1:
-                        continue
                     x, y = points_score[:2]
                     is_done = False
                     for pm in predicted_masks.get(label, []):
@@ -444,7 +437,7 @@ def infer(
 
                     point_coords = torch.cat((points_score[:2].unsqueeze(0), bg_coords), dim=0).unsqueeze(0)
                     point_coords = self._preprocess_coords(point_coords, original_size, self.config.model.image_size)
-                    point_labels = torch.as_tensor(
+                    point_labels = torch.tensor(
                         [1] + [0] * len(bg_coords), dtype=torch.float32, device=self.device
                     ).unsqueeze(0)
                     mask = self._predict_masks(
diff --git a/src/otx/algorithms/visual_prompting/tasks/inference.py b/src/otx/algorithms/visual_prompting/tasks/inference.py
index a9b6105bb12..86a45131b52 100644
--- a/src/otx/algorithms/visual_prompting/tasks/inference.py
+++ b/src/otx/algorithms/visual_prompting/tasks/inference.py
@@ -641,20 +641,19 @@ def _export_to_onnx(self, onnx_path: Dict[str, str]):
                 model_to_export = self.model.image_encoder
 
             elif module == "visual_prompting_prompt_getter":
+                reference_feat = torch.randn(1, 256, dtype=torch.float32)
+                reference_feat /= reference_feat.norm(dim=-1, keepdim=True)
                 dummy_inputs = {
                     "image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float32),
-                    "reference_feats": torch.randn(2, 1, 256, dtype=torch.float32),
-                    "used_indices": torch.tensor([[0, 1]], dtype=torch.int64),
+                    "reference_feat": reference_feat,
                     "original_size": torch.randint(low=0, high=image_size * 2, size=(1, 2), dtype=torch.int64),
-                    "threshold": torch.tensor([[0.1]], dtype=torch.float32),
+                    "threshold": torch.tensor([[0.0]], dtype=torch.float32),
                     "num_bg_points": torch.randint(low=1, high=image_size, size=(1, 1), dtype=torch.int64),
                 }
-                output_names = ["total_points_scores", "total_bg_coords"]
+                output_names = ["points_scores", "bg_coords"]
                 dynamic_axes = {
-                    "reference_feats": {0: "num_labels"},
-                    "used_indices": {1: "num_labels"},
-                    "total_points_scores": {0: "num_labels", 1: "num_points"},
-                    "total_bg_coords": {0: "num_labels", 1: "num_points"},
+                    "points_scores": {0: "num_points"},
+                    "bg_coords": {0: "num_points"},
                 }
                 model_to_export = self.model.prompt_getter
 
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 9db584f6af2..b7aaa85bc51 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -303,24 +303,6 @@ def pre_process_image_encoder(
         """Pre-process function of OpenVINO Zero-shot Visual Prompting Inferencer for image encoder."""
         return self.model["image_encoder"].preprocess(inputs, extra_processing)
 
-    def pre_process_prompt_getter(
-        self,
-        image_embeddings: Dict[str, np.ndarray],
-        reference_feats: np.ndarray,
-        used_indices: np.ndarray,
-        original_size: np.ndarray,
-    ) -> Dict[str, np.ndarray]:
-        """Pre-process function of OpenVINO Zero-shot VIsual Prompting Inferencer for prompt getter."""
-        inputs_prompt_getter = {
-            "original_size": original_size[None],
-            "reference_feats": reference_feats,
-            "used_indices": used_indices,
-            "threshold": np.asarray([[self.model["prompt_getter"].sim_threshold]], dtype=np.float32),
-            "num_bg_points": np.asarray([[self.model["prompt_getter"].num_bg_points]], dtype=np.int64),
-        }
-        inputs_prompt_getter.update(image_embeddings)
-        return inputs_prompt_getter
-
     def learn(self, images: np.ndarray):
         """Learn."""
 
@@ -330,23 +312,20 @@ def infer(
         """Perform a prediction for a given input image."""
         # forward image encoder
         images, meta = self.pre_process_image_encoder(images)
-        original_size = np.asarray(meta["original_shape"][:2], dtype=np.int64)
+        original_size = np.asarray([meta["original_shape"][:2]], dtype=np.int64)
         image_embeddings = self.forward_image_encoder(images)
 
         # get point candidates
-        inputs_prompt_getter = self.pre_process_prompt_getter(
-            image_embeddings, reference_feats, used_indices, original_size
-        )
-        total_prompts = self.forward_prompt_getter(inputs_prompt_getter)
+        total_points_scores, total_bg_coords = self.forward_prompt_getter(image_embeddings, reference_feats, used_indices, original_size)
 
         annotations: DefaultDict = defaultdict(list)
         predicted_masks: DefaultDict = defaultdict(list)
         used_points: DefaultDict = defaultdict(list)
-        for label, (points_scores, bg_coords) in enumerate(
-            zip(total_prompts["total_points_scores"], total_prompts["total_bg_coords"])
-        ):
+        for label in total_points_scores.keys():
+            points_scores = total_points_scores[label]
+            bg_coords = total_bg_coords[label]
             for points_score in points_scores:
-                if points_score[-1] == -1:
+                if points_score[-1] in [-1., 0.]:
                     continue
                 x, y = points_score[:2]
                 is_done = False
@@ -359,33 +338,66 @@ def infer(
                     continue
 
                 point_coords = np.concatenate((np.array([[x, y]]), bg_coords), axis=0, dtype=np.float32)
-                point_coords = self.model["decoder"]._apply_coords(point_coords, original_size)
+                point_coords = self.model["decoder"]._apply_coords(point_coords, original_size[0])
                 point_labels = np.array([1] + [0] * len(bg_coords), dtype=np.float32)
                 inputs_decoder = {
                     "point_coords": point_coords[None],
                     "point_labels": point_labels[None],
-                    "orig_size": original_size[None],
+                    "orig_size": original_size,
                 }
                 inputs_decoder.update(image_embeddings)
 
                 prediction = self.forward_decoder(inputs_decoder, original_size)
                 prediction.update({"scores": points_score[-1]})
-                metadata = {
-                    "label": [_label for _label in self.labels if int(_label.id_) == label][0],
-                    "original_size": original_size[None],
-                }
 
-                # set annotation for eval
-                annotation, hard_prediction, _ = self.post_process(prediction, metadata)
-                annotations[label].extend(annotation)
-                predicted_masks[label].append(hard_prediction)
+                predicted_masks[label].append(prediction[self.model["decoder"].output_blob_name])
                 used_points[label].append(points_score)
-        self._inspect_overlapping_areas(predicted_masks, used_points, annotations)
+
+        self._inspect_overlapping_areas(predicted_masks, used_points)
+        
+        for label, predictions in predicted_masks.items():
+            if len(predictions) == 0:
+                continue
+            metadata = {
+                "label": [_label for _label in self.labels if int(_label.id_) == label][0],
+                "original_size": original_size,
+            }
+            for prediction, used_point in zip(predictions, used_points[label]):
+                annotation, _, _ = self.post_process(
+                {
+                    self.model["decoder"].output_blob_name: prediction,
+                    "scores": used_point[-1]
+                },
+                metadata)
+                annotations[label].extend(annotation)
+            
         return sum(annotations.values(), []), predicted_masks, used_points
 
-    def forward_prompt_getter(self, inputs: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
+    def forward_prompt_getter(
+        self,
+        image_embeddings: Dict[str, np.ndarray],
+        reference_feats: np.ndarray,
+        used_indices: np.ndarray,
+        original_size: np.ndarray,
+    ) -> Dict[str, np.ndarray]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
-        return self.model["prompt_getter"].infer_sync(inputs)
+        inputs = {
+            "original_size": original_size,
+            "threshold": np.array([[self.model["prompt_getter"].sim_threshold]], dtype=np.float32),
+            "num_bg_points": np.array([[self.model["prompt_getter"].num_bg_points]], dtype=np.int64),
+            **image_embeddings
+        }
+        total_points_scores: Dict[int, np.ndarray] = {}
+        total_bg_coords: Dict[int, np.ndarray] = {}
+        for label in used_indices[0]:
+            reference_feat = reference_feats[label]
+            inputs["reference_feat"] = reference_feat
+            outputs = self.model["prompt_getter"].infer_sync(inputs)
+            
+            total_points_scores[label] = outputs["points_scores"]
+            total_bg_coords[label] = outputs["bg_coords"]
+        
+        return total_points_scores, total_bg_coords
 
     def forward_decoder(  # type: ignore
         self, inputs: Dict[str, np.ndarray], original_size: np.ndarray
@@ -419,7 +431,7 @@ def forward_decoder(  # type: ignore
                 has_mask_input = self.has_mask_inputs[1]
                 y, x = np.nonzero(masks)
                 box_coords = self.model["decoder"]._apply_coords(
-                    np.array([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=np.float32), original_size
+                    np.array([[[x.min(), y.min()], [x.max(), y.max()]]], dtype=np.float32), original_size[0]
                 )
                 inputs.update(
                     {
@@ -467,7 +479,7 @@ def _inspect_overlapping_areas(
         self,
         predicted_masks: Dict[int, List[np.ndarray]],
         used_points: Dict[int, List[np.ndarray]],
-        annotations: Dict[int, List[np.ndarray]],
+        # annotations: Dict[int, List[np.ndarray]],
         threshold_iou: float = 0.8,
     ):
         def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
@@ -497,12 +509,12 @@ def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
             for im in sorted(list(set(overlapped_label)), reverse=True):
                 masks.pop(im)
                 used_points[label].pop(im)
-                annotations[label].pop(im)
+                # annotations[label].pop(im)
 
             for jm in sorted(list(set(overlapped_other_label)), reverse=True):
                 other_masks.pop(jm)
                 used_points[other_label].pop(jm)
-                annotations[other_label].pop(jm)
+                # annotations[other_label].pop(jm)
 
     def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
         """Perform a prediction for a given input image."""
@@ -620,7 +632,7 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
 
         items = self.dataset[index]
         images, meta = self.inferencer.pre_process_image_encoder(items.numpy, extra_processing=True)  # type: ignore
-        original_size = np.asarray(meta["original_shape"][:2])
+        original_size = np.asarray([meta["original_shape"][:2]])
         _, _, h, w = images["images"].shape
         pad_width = ((0, 0), (0, 0), (0, self.target_length - h), (0, self.target_length - w))
         images["images"] = np.pad(images["images"], pad_width, mode="constant", constant_values=0)
@@ -628,26 +640,31 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
             return images
         else:
             image_embeddings = self.image_encoder(images["images"])
-            inputs_prompt_getter = self.inferencer.pre_process_prompt_getter(
-                image_embeddings, self.reference_feats, self.used_indices, original_size
-            )
             if self.module_name == "prompt_getter":
-                return inputs_prompt_getter
+                return {
+                    "reference_feat": self.reference_feats[self.used_indices[0][0]], # only use the first feature
+                    "original_size": original_size,
+                    "threshold": np.array([[self.inferencer.model["prompt_getter"].sim_threshold]], dtype=np.float32),
+                    "num_bg_points": np.array([[self.inferencer.model["prompt_getter"].num_bg_points]], dtype=np.int64),
+                    **image_embeddings
+                }
 
-            total_prompts = self.prompt_getter(inputs_prompt_getter)
+            total_points_scores, total_bg_coords = self.inferencer.forward_prompt_getter(
+                image_embeddings, self.reference_feats, self.used_indices, original_size)
+            
             # only use the first prompt
-            point_score = total_prompts["total_points_scores"][0][0]
-            bg_coords = total_prompts["total_bg_coords"][0]
+            point_score = total_points_scores[0][0]
+            bg_coords = total_bg_coords[0]
 
             x, y = point_score[:2]
             point_coords = np.concatenate((np.array([[x, y]]), bg_coords), axis=0, dtype=np.float32)
-            point_coords = self.inferencer.model["decoder"]._apply_coords(point_coords, original_size)
+            point_coords = self.inferencer.model["decoder"]._apply_coords(point_coords, original_size[0])
             point_labels = np.array([1] + [0] * len(bg_coords), dtype=np.float32)
             inputs_decoder = {"point_coords": point_coords[None], "point_labels": point_labels[None]}
             inputs_decoder.update(image_embeddings)
             inputs_decoder.update(
                 {
-                    "orig_size": original_size[None],
+                    "orig_size": original_size,
                     "mask_input": np.zeros((1, 1, 256, 256), dtype=np.float32),
                     "has_mask_input": np.zeros((1, 1), dtype=np.float32),
                 }
diff --git a/tests/assets/car_tree_bug_zero_shot/annotations/instances_train.json b/tests/assets/car_tree_bug_zero_shot/annotations/instances_train.json
new file mode 100644
index 00000000000..39dadb88943
--- /dev/null
+++ b/tests/assets/car_tree_bug_zero_shot/annotations/instances_train.json
@@ -0,0 +1,66 @@
+{
+  "licenses": [{ "name": "", "id": 0, "url": "" }],
+  "info": {
+    "contributor": "",
+    "date_created": "",
+    "description": "",
+    "url": "",
+    "version": "",
+    "year": ""
+  },
+  "categories": [
+    { "id": 1, "name": "car", "supercategory": "" },
+    { "id": 2, "name": "tree", "supercategory": "" },
+    { "id": 3, "name": "bug", "supercategory": "" }
+  ],
+  "images": [
+    {
+      "id": 6,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide4.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    }
+  ],
+  "annotations": [
+    {
+      "id": 16,
+      "image_id": 6,
+      "category_id": 3,
+      "segmentation": [
+        [251.2, 150.5, 372.47, 47.31, 596.99, 231.4, 455.05, 376.77]
+      ],
+      "area": 53610.0,
+      "bbox": [251.2, 47.31, 345.79, 329.46],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 17,
+      "image_id": 6,
+      "category_id": 2,
+      "segmentation": [
+        [641.72, 255.48, 731.18, 87.74, 848.17, 144.52, 746.67, 311.4]
+      ],
+      "area": 23927.0,
+      "bbox": [641.72, 87.74, 206.45, 223.66],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 18,
+      "image_id": 6,
+      "category_id": 1,
+      "segmentation": [
+        [791.4, 443.9, 984.95, 183.23, 1112.26, 273.55, 910.97, 536.77]
+      ],
+      "area": 50412.0,
+      "bbox": [791.4, 183.23, 320.86, 353.54],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    }
+  ]
+}
diff --git a/tests/assets/car_tree_bug_zero_shot/annotations/instances_val.json b/tests/assets/car_tree_bug_zero_shot/annotations/instances_val.json
new file mode 100644
index 00000000000..bdbfe9331bf
--- /dev/null
+++ b/tests/assets/car_tree_bug_zero_shot/annotations/instances_val.json
@@ -0,0 +1,316 @@
+{
+  "licenses": [{ "name": "", "id": 0, "url": "" }],
+  "info": {
+    "contributor": "",
+    "date_created": "",
+    "description": "",
+    "url": "",
+    "version": "",
+    "year": ""
+  },
+  "categories": [
+    { "id": 1, "name": "car", "supercategory": "" },
+    { "id": 2, "name": "tree", "supercategory": "" },
+    { "id": 3, "name": "bug", "supercategory": "" }
+  ],
+  "images": [
+    {
+      "id": 7,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide3.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 8,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide4.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 1,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide9.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 2,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide8.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 3,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide7.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 4,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide6.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    },
+    {
+      "id": 5,
+      "width": 1280,
+      "height": 720,
+      "file_name": "Slide5.PNG",
+      "license": 0,
+      "flickr_url": "",
+      "coco_url": "",
+      "date_captured": 0
+    }
+  ],
+  "annotations": [
+    {
+      "id": 19,
+      "image_id": 7,
+      "category_id": 1,
+      "segmentation": [
+        [184.09, 131.61, 338.06, 129.89, 339.78, 457.63, 183.23, 461.08]
+      ],
+      "area": 51030.0,
+      "bbox": [183.23, 129.89, 156.55, 331.19],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 20,
+      "image_id": 7,
+      "category_id": 2,
+      "segmentation": [
+        [832.69, 104.09, 1018.49, 102.37, 1017.63, 226.24, 825.81, 233.98]
+      ],
+      "area": 23933.0,
+      "bbox": [825.81, 102.37, 192.68, 131.61],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 21,
+      "image_id": 7,
+      "category_id": 3,
+      "segmentation": [
+        [898.92, 490.32, 1195.7, 487.74, 1209.46, 673.55, 913.55, 670.11]
+      ],
+      "area": 54157.0,
+      "bbox": [898.92, 487.74, 310.54, 185.81],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 22,
+      "image_id": 8,
+      "category_id": 3,
+      "segmentation": [
+        [341.51, 373.33, 502.4, 456.8, 341.5, 709.7, 188.39, 612.47]
+      ],
+      "area": 52814.0,
+      "bbox": [188.39, 373.33, 314.01, 336.37],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 1,
+      "image_id": 1,
+      "category_id": 3,
+      "segmentation": [
+        [17.2, 166.88, 203.87, 7.74, 410.32, 43.87, 117.85, 331.18]
+      ],
+      "area": 58273.0,
+      "bbox": [17.2, 7.74, 393.12, 323.44],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 2,
+      "image_id": 1,
+      "category_id": 1,
+      "segmentation": [
+        [294.19, 281.29, 643.44, 300.22, 628.82, 469.68, 277.85, 449.03]
+      ],
+      "area": 59331.0,
+      "bbox": [277.85, 281.29, 365.59, 188.39],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 3,
+      "image_id": 1,
+      "category_id": 2,
+      "segmentation": [
+        [114.41, 499.79, 30.97, 670.11, 151.4, 705.38, 240.86, 536.77]
+      ],
+      "area": 24033.0,
+      "bbox": [30.97, 499.79, 209.89, 205.59],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 4,
+      "image_id": 2,
+      "category_id": 1,
+      "segmentation": [[165.16, 2.58, 344.95, 41.29, 27.5, 363.0, 9.46, 147.1]],
+      "area": 53173.0,
+      "bbox": [9.46, 2.58, 335.49, 360.42],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 5,
+      "image_id": 2,
+      "category_id": 2,
+      "segmentation": [
+        [524.73, 378.49, 648.6, 227.96, 762.15, 298.49, 627.96, 458.49]
+      ],
+      "area": 26526.0,
+      "bbox": [524.73, 227.96, 237.42, 230.53],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 6,
+      "image_id": 2,
+      "category_id": 3,
+      "segmentation": [
+        [946.24, 652.9, 1191.4, 356.13, 1274.8, 576.3, 1092.5, 715.7]
+      ],
+      "area": 55317.0,
+      "bbox": [946.24, 356.13, 328.56, 359.57],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 7,
+      "image_id": 3,
+      "category_id": 2,
+      "segmentation": [
+        [584.95, 221.94, 715.7, 223.66, 706.24, 411.18, 583.23, 413.76]
+      ],
+      "area": 24074.0,
+      "bbox": [583.23, 221.94, 132.47, 191.82],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 8,
+      "image_id": 3,
+      "category_id": 1,
+      "segmentation": [
+        [826.67, 222.8, 966.9, 176.3, 1081.29, 489.46, 931.61, 542.8]
+      ],
+      "area": 51362.0,
+      "bbox": [826.67, 176.3, 254.62, 366.5],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 9,
+      "image_id": 3,
+      "category_id": 3,
+      "segmentation": [
+        [698.49, 384.52, 864.52, 390.54, 872.26, 688.17, 683.01, 683.01]
+      ],
+      "area": 52982.0,
+      "bbox": [683.01, 384.52, 189.25, 303.65],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 10,
+      "image_id": 4,
+      "category_id": 1,
+      "segmentation": [
+        [69.68, 11.18, 67.1, 336.34, 213.33, 338.92, 222.8, 10.32]
+      ],
+      "area": 48945.0,
+      "bbox": [67.1, 10.32, 155.7, 328.6],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 11,
+      "image_id": 4,
+      "category_id": 2,
+      "segmentation": [
+        [569.46, 70.54, 688.17, 70.54, 683.01, 262.37, 559.14, 263.23]
+      ],
+      "area": 23273.0,
+      "bbox": [559.14, 70.54, 129.03, 192.69],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 12,
+      "image_id": 4,
+      "category_id": 3,
+      "segmentation": [
+        [972.04, 116.13, 1265.38, 95.48, 1274.84, 295.05, 974.62, 292.47]
+      ],
+      "area": 55841.0,
+      "bbox": [972.04, 95.48, 302.8, 199.57],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 13,
+      "image_id": 5,
+      "category_id": 3,
+      "segmentation": [
+        [200.43, 336.34, 385.38, 334.62, 382.8, 635.7, 206.45, 638.28]
+      ],
+      "area": 54478.0,
+      "bbox": [200.43, 334.62, 184.95, 303.66],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 14,
+      "image_id": 5,
+      "category_id": 2,
+      "segmentation": [
+        [594.41, 523.01, 779.35, 523.87, 778.49, 643.44, 590.97, 645.16]
+      ],
+      "area": 22525.0,
+      "bbox": [590.97, 523.01, 188.38, 122.15],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    },
+    {
+      "id": 15,
+      "image_id": 5,
+      "category_id": 1,
+      "segmentation": [
+        [1101.1, 304.6, 1230.1, 389.6, 1058.1, 665.8, 929.03, 581.51]
+      ],
+      "area": 50271.0,
+      "bbox": [929.03, 304.6, 301.07, 361.2],
+      "iscrowd": 0,
+      "attributes": { "occluded": false }
+    }
+  ]
+}
diff --git a/tests/assets/car_tree_bug_zero_shot/images/train/Slide4.PNG b/tests/assets/car_tree_bug_zero_shot/images/train/Slide4.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..236d848e36a377a1791144c9bc0325801af42b16
GIT binary patch
literal 26804
zcmeEu`9IWq8~2n`j%=M$La5V%3fXrmMG|985h~eZ?87i<K}A$TSwmSv_T7j|*~c1#
zA=_BTGIqvueRb|1p1<LFez;%vea>yZ-_LR_@9TZNukSM*ZFSat-1|@{6syM7%X%o(
z9{8WTTmRk-|8;@?+AjF7U9NiSmr&WwM<?JP{<6NPbrFTi4P)NCi~K2*)74w9C=^>Q
z@}FIGn6&#Sl#PbQ<%<TMW;0vNZw;65qI6CjwCKcyZCkrJ8mH}Q*Y?&#bvHvT3sy}l
zEbAHL-4fQ?E3;UOPS1*Ce?3Y2n~m+Mi|+Bzf732-Jl%D<V$Y%EzlinxlKy}F^X%$h
ziMHG3eP`SIStZ!cdt1#^n&pgdt$vpFULMUvD^NdCt+P~Cv9PYIs69teDF4ek{;-9s
zfBpY|_<#5eeyC>3%;Is2-u%99=j8W-jUPEMx->(rnrny?IK4ilDDKwVFz&ZK(LEr4
zgqH#3znf*vf&5Svjp-HJcsO34ga7=#AjzwHLTWVWF7tOSQ@(hvc!D>Tk=Qc+scoJ|
zbr*e!eu5~5LJ3yCj51ptZO+7aZ*E@1rq(-Tn&q3UROZaQ|H!NUNzqoG#qX2n;O`rj
z%*HG6aqF*=lq15&KJ`@D`_Z1ujSgE7m}g{8!V>+_@F<ngqe?{&W+MH#om=MHNE_L$
zwyOEUT*b<rd#B#MZb;Y{AzLuoRXP5od&1q)(6ec-p6EUD$egNjpz3PR2hL;*6l(Pe
z?Bx#ZrC1bKW<eN^DE#bmx9(MDo5GB(K!vtv;jq9HZ==BOmWV7O;n#A^+>kn%e$9;q
zg-X!zKVHl$hdXD`-m1%tGdRqCyU%jEjJDTAlqu(y;#}g5N-IvcoVUep7PRkTMY!RC
zxm*EJpA-LEe65++aGc*!XLLi`I|oLUShHLCZ*(4;xMfmct&P@?GkuFBe0pz%IfjMq
zLJc0=cYfr<<UvA!C3Bv0S8<9P_txdRWG^w5Ly8tZi{^BrvdO$~Gt~kdcYY{!3iy&(
zy(BB+q;J-29HzbHBI$n_9>=iqyUe6u|F>(ou?a1OePu?HbI-^J4-Qm*rdIq`j>y4;
znYJ92pBbHFGWVXk^S_NoU7jiSxr^g;`1IyB-9UHk(Gib9xY*jdjsGsz=hCjg*ncvY
zW3FW|aDGc?@bj6*rnekMmP8e*e#vw1m#;n(Z^`Ee7j>y6ZiC1jC>#Fo4wUvBak=n-
zggKKtS-o?}uhjlKWI`%RwQokjDb0c)m$fkvH6S+}Im}x^tw65)RdQWYOaIjym_8om
zmXQF=<vJ8<&U@eaul*72eD<C89yV_$0we+_#63*mjJsQI^WtaK^p^7DhH4veZr;e=
zkmrBO>YKwYX16PZ-h5{!Wf8Ni8Wn&pUmhnVp~*+A&h;Et{@J!vzH@A_(BJl8ELZSn
zg7+%FG!wUoPx1|xUyeDb>_y|{?vuhgF!2vhG*PeoZz<iaJp7KG_S%~=DygaW`_{oI
z*h_q!kw0#=6%7iT?=@lk-=pu@_4EqY$*6?gjxZB-{dY)0=C>L}GaIk$6S&2w5{j6?
z3dfzN>Dj}Fzb{v*Cq&m&t-LYax<J$7Zmp~Hjao+`F7`p!|9B{3TVtlFSnj3v$!qUS
zCeCl>&aLs#i>e8|8@iq=c^aZ>u)>#Wy8>m)TzW0CHu@sWgj_|27Q<r76xWqmM>$Ep
z2M5-l--<yNcQqNMIz0`y@ZS4M`SGxnm7lSitDAd19%A*8qjd7D1Qy3v%)OvMK4}xI
z-mkD2>G<H5yhO1G-REA-IHogsVQdYpM$Wj{S3Ws2*X7FZeHL*7Nw_KPRA1L#8bPJ<
z`k2jgMrDJ(RPps2toOdZb{N9{XVMqA;n0z2G)5cEX$vpo*HbmyHx*r8^kfAIS^wLz
z+2I`A$1BCpY25G1g-SQm=S(BB`VZiXr+E-#uF9!&KQF&`wMClhIsf(OYU{S$-j^J(
zfQF{UgmcppGL>p$Zu=T$qb04iX#IKA*Y*kQn6mA`I-=N1Ptr<7mEEu^HoL^OqcJ<G
zuQe&<4SGg#sOld>d#Vq;(A}H;NHXe&2|P|S=WJ##ZNnKC)HqQv;pbP}z^O(SZctWU
zF7Jp@tGn>1GIz1lgwx!#ALHGxXP}%vGg!*2ZWyQHIDu`G@c4-h?C-X$o?f<W5BAm)
z(pK_Y4WS11`q18POR*j!!{LfL;EalRH=W!cHZFZ|40Icz<Wi<A+jEwEIR<hcv<1U@
z6%+HadkPdgx^?b(Z>Lf6<ah%`r(Y92?a`E4RZcgN7+s%<7kesC7snJY+z@!OYvI{$
z7O^~hHytsUqorpmRma@qU-<s(lWa8LDX<>cLko5wFi}0feQJWIpJY__5q;AcA5IGR
z%{@7lzC3VHV*LyA-bP1r`E++mn@qpT?2EV@HR@j%MD+cSt98+ROwxi&eHjNsIB0Ks
z&Ub$?u5`wDmj7fXKMj5O^=7`USI*BF+;x$C5mFr8A!ob;xfF{nTGFp1?lsD)l)$N=
zTq0q8;*e*Im!jha_!f@^;H_Fr%EVx0Zq$t8T2d2!I*W4Ts=sOwd$Z^2daL#ZndmtE
zx^dPxjHz?IjfXiqR=>!e97?I&_2VL3(p9s=gGp73H7n1n?n*b^%oDEYZ(B-PIwmB&
z)gUpdudwp=w%Zs_{M74zg0JBQI#M>C)JVldswbfQZ-XO*JW0CQhw;UEH$HxXaocL_
zB+4|V*w6MJ7?_=CnjL3)cHVyNgQla6u5_|z<8Nk(XuD-`xY&B@A}y^w<!X`_MD*ZT
zI0hxe?o4V9%@obZbMP{(*?+I?Hpf|fdASxRvCppdWCg}^<+=KI^@P3teTm3Rci-;7
zn(e%Fx1w!xG=;~i{ZzLy|A6E6#qs;c5_pxCoin!BOj~7VWFqEf!AS279h688+jAil
z@|4=)tX~~v*s3N}(C$8zL^YD<E4r^3De=Io1q;aYE38!MT=(H3vccZ|elfSf(H6Sw
zx)W+-_nss7?n6`(*0fG3u2a0K+hfY=bckX;vn5zzv=hFMQ4Na8a8Tl9HWrBGCNf|=
z-3@8~?0O2p^4`fZ_PLU!i9O<}2dgw!)rP#Ukk8voJK|L;`VZwxtJhr^D0QFj%(Ww=
zziD|ua=0`#<82nEd<r$60((BH=ytN>8(XdK>w)90^1SNV`^Yvc1MEqdw^D#tSTR`5
z&KR&r^R5TRzCT%oe}7(F`$~&8_8E=MqdeZdAA`8}p2y>mm}8L#I4}wePwI2wiyxF>
zJ4r(`W6H}n+*n3lG4NM)ySkJwJTKyEk~qzkoAy=B#{V`Pdc|$IncSMZJrlg*sromo
zoT`_Clc|wf#RipJ1CHM$ypt@Bg>4enqLwPsn|_^o*Ce4V6r+1vxiDCPf#GYyZkG8x
zn{O#TWo$QMs<33m?Vpr#%XTC~U{+{jEAV6UbAHQp$W5)*h;YS8+~C?8+|z%$D86{1
zi=xV~-w0VhGoi1iVv44noNzVgm0UD{8aravz4;8IBv=VUN)^X|GbvjYPJz>7hD3w|
zq({t%@3Ur+m}_PmN)BW{ig!cZ)7Ta0Dd$4FDOf!dB%+CVEf*bn;zl49@p)uTJ9gx}
zeQ|^Rcut?&c8QsO@NmVNoRZRMPB|u){jRVMZHR08HU`26PP$57*_xfAZ(Si4*Qc<n
z$R8e<P0Eaa5&E#=Gq@Poz1V%<qdA*CZM2*RM#_b*w256$Pk|)}xXTP`C;LX5opu#5
z{hn>#qqC9VSGbZq7it-9781JnZSsY8!oa3k+6qzZWMf4{vB&n^sgT!P;~^E>o2|cO
zz%_}SsAIm60F#;bwm7tJ(`m}@T!r{nb(|CGPpI=GRXFsm`+C-$wQ>jzUA&S;8}bgo
z@$;&_KDs(rIeXBUy&_*#?PrJ=PD?o3^S%#_7xk1MGK&rIxt4I5p&M6ub{Tr?w((MI
z=lO0wHUv}*%tvO(BoF&I6AfDesiQ3Ce->=^@Pn5&+<r&TYWDA$DMt&sshsrx1s)v=
zSoB6mYM$qyYA0U2xb5yQH23s^(U;210XB^MN^Wt|aK&dxC_|d2E;Ea;M-9Q3qx$7w
z_A$Dm!wE!oGbgZN>58&5>!(p)Lm=e2FK!B$ETx@;WHx!@ovP3*w&n9Pc6IW}4aanx
zM8(8&>*dRJUU%X8hcsE=uIst6XdKvgKkZfGb*qks*sDEFGqbmy^=VqDr(AIUb^u!-
zV_#w5PjieYeeI#j=P=;Deil=45Fb8Wey?k6u8OC)il28pAXSFIS=9lC`YiOg##mNG
zVRy&xkbCdW3gy@+K+Yq|>^f!zw~+HK(~L<gZ}@3MXTLZd0a+Kr6dcWIZi3ZAi{>h>
zOO`T<Z?2K;`KFyM4S)DgaiXL1t(F;D0xF=YIBxBk^=X&B`94ggLw6!YR`ipn6?#=~
zP39MiyVzLhY@(VKN5k!(aKxhJOZ-Wg-4_}m`PtW=R_@E_o__5-mC0k(sDD$M;NGHe
z3Bztn%u0IS()Y3pPd&?5QX=oLz9p-8(?D58@%U2X%iuualJ%lG&a7V?>rS-MsdJ*D
zho#^~YNC)p09WyyXu*$-X~nhn^GZCo%)b9)vl-m7#{DzuH@MDC5_V9h&e!$%x^RZ$
zTW4)RK)#t%AzhX$NFwg1S|G3=qSS%s&NkmVBTB|q$bK71spe$5W5IqY=1Q50tCVV@
zbJbN=>70QQ8Tr7kzbb~m=^TT!pP+lr_0);O5l>nONPgNt?H2%N>+em#Fb90zYPIZq
zAr{Z1v~1FU=2BRb=cYlsE$v|9RelxCzGQKqdF`#S>(TiD<E7|l-sNxx*Ef8zx>}5H
znc^`sH7s0ieko)~CV0@_pPilLb?GVaqe$l-OEw(S`7W}iEm++`C=ywZM}39tFE_aU
zF4H8k`O}76XCQi9%-;A-^BIaecFU}&Z1{(v!#bH<wi4U=Bp7F8h9?aRNDt}!Q!Alr
z3T}G^l0fasdgnk`lEn`^U;|83%4j=ZnH3<V#D7!Ts9m8-7{`Nc+o2v?XTwrAw6bt$
z`*7kJ*Rz$5c05(Kw8C9a4+8wJ1~4Buu<8_C;xTBj;n3r%y=)bs@WM9<{3#1ZJ`_C%
zk-XWY%$fX<^4e$(mz!2`^ug(O1HaZLAy%*@I6G&|^HIwpX7;ws$llY{P{UmRNcwPN
zh>44R>;}i39T%1PTErl8VM7aTQv(K4VmZ;+`XncI`4>2WD*_~jy*zT$0BN(CPXhi6
z3q8)rEAQjf`8_DGWz|uz@?$vZjZp!wIAIZjz3KedC%YBzcx;3)x}F_6d$t-i{{f}i
zeARvM5uey(HEW-5roPLX{!E~v@h@w~v@hG|mscH<_`K|}1=qcfn={5}JcU$8faKSI
zx?mP^FQxJXRi&`uR>1Dq@+g1pQyhb=IoZ!68T@4d&qqCBQ+#mmw!D)?A(?hm<#-<<
zLu>6(D8pEjN!k2h;tArl(UKjv?f>Jp>Q5#WCh9Cgsjl(HlAexrWwBHxG1I!@ACGre
zc+~|HlXeVk<gEdtqw^-Nf&HV=m%YvN3ViOFHhs$`adTYH8ATj#$>*nGN2eL%8y<7V
z<~(j1$u-^ZFfbw?UsSN>sL{XX`6^bVI+?%X!mhnfh2aX@Dn9ZsUpw3TrfEiU{Tc1d
zFBB3KGL;v@%p<)98J4`8#wD(&eStLW(3vd$sS%L(r>bi=L=dxG=mZxO#oPH7Gi7-X
zbd0fNV$LKgv6bg&et#_YM%Qp04mkHCN;SI|m;h>K2KTCgliTm&x;Q`u^{bh0Pr5A(
zOO<p_Kbn=h{TKt`(I*P(61BShMu}sr3C{B$We!6=!s2-1(buy%2Vo0AIZl+hFE%ew
zJLX67i{00}VBlm^k$#YXM2{N+zV_EhU)VWmY1bW|E4xV6Ii7X{1Uz*10c06BmU&m&
zue!w>1KiPO3AW!-aqlvF+ubiUG>p0Z@H`*XM#$@bZ@VpEd>VBRYOL5O`9nTFXcw{<
zVh_}rf^}SjFs^YHs<IHm3)dPw4&V-wOs+sN6q(iZk@*@D&Q~{)TFs3`%yqDKg}&CT
z95wJuoTs+zqrUM9;Y1LX;kJi;IRX*S?m>d%O97<jo3hRp5zG>#IIG9_swgk^42oEE
z2Q2?88~>z(CoRQwX0-H$2#j2-;S@S&ex&N#$0$*sp(usqrRT-L**IyR1$LGC+=JaG
zzoz(krF4&YC}HgM2?Ais(v>AE1=8S4*C2Q_!)bo%UDYYLxs#~8M!^DujVfK{e|U!`
z>=z*@a1e?fM|vF1%?nKpP<3#W86mbN)V2f~35sM^;EC-#W_|zAkD8qv{#hCUkf68h
zf!XT65ZWsMkJwb+6QT{wJrpfnUZWiOeSa-qrD`Pbu%*QFDf{0mUlC9tUuyXMaVY}M
zYt|o-7jXq<JIh#`h>#gz>R-y^;j=T8mnf(j#+OWI|H%leAB4`B>$IN0S@m?X_W=}(
z;4wqLXcjTfO^1>(_U*=2huabQ%}g`b8Z1ljLDH8&hjO-nFnK`eM6s{mYXqzUJ!IyZ
zr`afM@ub|T)x*e1)cmGJ_bD5TqJFFI3fwFPy35|L$^kOmi`AhMmD#xWbrFVKLK|Nm
z8<Cei1<Mkp)m6f+8sAs~h0|t#KRdPmrMfC<JN2KHSlWhk*$WdgJO9E?!Tt1Ma-DMg
z{M=I1z|B?^mcm>0Su?HYR<C*7%o}nSy?5<#k|6_vga-9Fma-ox*x2G8Ha<Se&ob-n
zlf3#t&Yq1={s6wy^X7bG{2UYF4f28C5BE}6Hv+LSzi^aCxRNGns@3#ssv9r)9$<E^
zusW6WtrjFB7|SBt*=x7BzKm7XY+P1~<UU)6y}T0JnM;mMm6>^3<;1d6#`n~yBwN@#
zC>E)bFCe$him#h3>w%SwchrP14kysDZbu|%l59Fr2>45Dn%_0p_HN>%Sc{&xBR+<{
z67<?{w4iR8SC#`mqvj4ce&}%)AOJU(7j|6>8pyuSkNajG9GHF0V;quE|IxG+0?n~=
z9ShceM}?1~azCIL3g2y2yI#SMf0X%NaUB>t;0yIT%3VeSzY{#BptuMMD{!9LAc{%G
zXB1)s|CPV*`?!ewa=84xKtW@cY3oq?a&#Y^k{gy>$Z#x@<Ia0&Ac*=iS~h9fZ*vd=
zA?YEzBvpQ^mt-V0YcS`ZE*j91?_hLJTRr?f<kv&J7>3(RlOF)_$EwPx?t4*L_dyzo
z2U)x*=Z~;DF5GZ!z$kNj@IBwu+PS)YI0KJKnehPg*^tQhUrgOP^M}HHR&}pa+b&;1
z0O~?|oGNpZ4URR^7hAx*RDPT>tmPyFb-ERQY3j2LabXuy4(P=28TFY<&OM|(f4d^Y
zRa})u#=mgS8Eo~gwf>t|Rz6TooL)5qc`C&8bq|gESnceT9(YUlys${z%~o&Y7>G7^
z)ePhOp#Eo0c*_NA6-$A&3URq%=c|L-Htt?(bNrDE#h#hsLi|kmvbsI-7*e1P;Hx{L
zk}Pak?4^YX;Est;tt2ZSSVMW+V8szsJd3xFd9{*|P{whi7JV$iZoTie;=;1#_3hmw
z&{#;*60A<#mm*U;GO}9vXrob>VkU-}_--SYhZ*i#sWmYdvv+l4VRfWF?C%#SnDtXH
z4~dSNVL=@2No4=&!%tKURDQwY9m9>K>Yq{MQ~WH;Ki?*;NYa-UQEAM`3+}a`j0AkP
z4)>o~H!m)^n;la4`<>E|)*sW~yV8PPp77Ft?C8rzH8dkGgH1j03%%uYcisGj@7$y8
zVUw@RF0BB=ZsRO6Kcbt$y4d`G>H9zI^}#gazcFy8a@PEn0yroO*EHLKX+|(}mS)$E
z7laNafQ5Z(9Tlm}eSch3Nos+<sF_+YE1zhoWfHexRK(0HD>)QW$D@>3yb`VFGqZV-
z*t_c`Gh%P}9Iww>+Z*kqJL3A_6?c0cJ(@ifeB;6M$kPft2)};^;kWu1QN1cjPr4GF
z;LxWnDUX^H*8_JSwR8GzZYV%&gkxEXH9nI0qIIbDjg~#h-eseb(L=p~@q&mB>>(t&
z;+$^r#SA6bhz!_sy=|B!MuCPnh)i$E(iY_icRl1@_JsL<GZnj%OcD07^5Ii51%ffU
zo9Ru`9mB2R-sE0cqmRmBLl(5Mk6II7|4k;K<d7iY3hYN#{oxN$)lX8P4gkBMw?BPg
z%8LVQsVRRzau<{j49xpeGkr<R;BNWag;_V^BQ#<9Tcp{|e|>~ViszwLeeGnLkujLx
z_O@u4TkC#c%qPr`lDh;ikN~Em%W|=e=lx})Ad#$_rHOPj{Wjt}!pfUziBy&AH3;#s
z)*&(~L9%*NdG3_^&&-S1?Yd52Z#qn0anlz%CNv7k*$0h7ZyFrv^#*t$L-iChzKn8t
zxSOTJ3W~ohoIy&r3KQb;(pyvv6h#n0lOi(M6ITK`*%7Y_d@a94I?oxsPpW)swY{Po
zW7q9w;XWu9Ec(fjkgi_ucy?B*G*0Y}|Is9HCNErjsLtZG?^);^zt`6`5*sF0*fdBU
zy!%8IcG8LD(?G?onBFxX956Mz@<~BPe_tthn8U2(c5&SCAoHa*(N}n%Bc67oT<n`b
zsgN57{XG#|1o~BwV=(lD(ML{|bnR4K6#!9rRceo95He+KhfHZbcEiCZ?XLEAN;5fr
zvU&WVgwNe5oy+#6u6kQm|M@DKkN(a%=J!j0R7frW)=52+=|Y82b#<yudwJf>U{JeB
z!iUs`U+#dq4Led7b|Tia8xoy8znEkUgoGwu3@J_a_PWxA-VV&e`K-7o^;@Jb>lufo
zE6?-LU(OqKYOo3*Bv8S7)YrGWS$v*vD=aG)@6J{_6{TYMYb72rbEC=wS-rNU2ockX
zWRf3M0EgK}h4CgBX>s5aF9m&CG3|Oh?lBFnRIHID64RbIKz+$!=Rh?fb>vz^kq5GU
z?4p43Q)}3F;Wn~<q{d_JXG&)i;$FP~d_E?A`+k8_>MRuTwNpra>BiE79e*B!(`s6;
z47e12F0L<QK*8>6T36zG4_@-Wl0!P5TQ;S&6Q=^@d9S&rRYBpR-|&n~dlIHApZ@vx
zt9fGHZX^uP35^N&|4MSR!nKD-?*z;;LzbA;xe3$5Q;@)Gga1IpTOI%rA)@hlS_g&5
z8xbJL*6HX30T(;(sdSpw0(Xy%xw+rio@(n5*huxsuvAd=`-L~HKilgOr?X$#L9i1k
z2}G1UXXly5wYC?o7nGpBzD5p|)3-Dq+HG-aTauL%8d?*w_bj2q!5LOzjqqcKKIJ2&
zb<Ss`v_6ve04lP?d`tI04};FNM08g-g_?T0pfz5fxi)H_5FEk-oU`Ef#Gd#9_8@J(
zdOu3p(7Sa03F&l0Pr9w>)isBiI>Y6V!<_y%_52^Aoq=(F)b<<zUP`xO`l-0OPqzKu
z%>k-oRaIO4HvdeG0UN}sLLXAEgMu)Ri`|^O(KT&G&lqoVxRzSb{C#ZP_<->W(44AM
zY}9yp;oJI`<k(v#AnMO;N|2S6pmy_Y8aFrL%o0V-C!-h=vbWuXO}H48Bi@9VkTX_H
zzv0hJ4Lnq(vZ@@BcY-VJws$H26d%KTz8r7!Hnl9OWB}o#fkz+xdVl!wQVj8n#vOlI
za3Qoc%c|r7c_-*zL01uam;Yp0WceZ#ls-bF@v25zZmhQ?nguTN$UA)>UW@C{xD&op
zb}%H&LjQrALHWhyV)ZjwX_|}!5IN^g7&Z1Dg`XK_Rpz_t-Bt6;%n31Fh-(vq5~lC;
z@chXSn?=jOhl?0nB4xnd$q#Ahk79%Ens?V~fNK{uDL-Xg&;1EF0mU_hX@c_tB33je
zK8xoVE3j3Nx0`)3q&X$-`$A+e<_DNylekkZr-h5@N0#!;*$TJ^4+0PK?8f5#S<@vC
zPpegHH<nvnSi977{U}=m;l<q4QI;r=>VED}(@hmi*Xrz5NDkP0OdB$g_9ZWSAWonw
zMKcp{N6%M(OJr#A9W1rEh;~9CneEy5o`bxq(!B6ugxH&|`qURA<GzliJ?6Zx7#~X`
zB%2cW-Jf-FBSq>%c{xwY)r%`&8vNR#?k?AC0rwlr&moLwOV3WG$(vZ~?t%jMg#t=o
zlZH71=B;v$I>sxi*1&q<fnfCo(--gZKDqZ)(9{G{qv5v*>Js>qM;;ijcWm;72aDU8
zRDPWx8{h^E9&k!tU=e~}Aq_l146urt7&n)U^mGPupxA~mDw~v>U?2wbKUN2dtws+t
z0hpi>P6V$UZjN}<B*<sdbHdmWbx##;X7UHo;!5S&U`FLPX5%B0F83JODrFRXf%sFa
zdub`trywa=5Yp^1#s$^?DH+)<Z$$`AESk<p3Hfc;Q^q|<WcPFrRAqc_>7$7Vq(+6X
zZDv;T2QV^T#|ydt-6Q>A{oB^rscl6E@bydLeAlhMpQ}30F}P#rvZ(p@C<gaJ@K$d=
z?_|TCM+n<Xb#cLIP^%G6j;|Exxeh(mqNUi5<vh71ANCJG7w$hpMw$hD=m+b4eZC0U
ze~9rI9~1>Ak#3XugN-I&RH>#*%Wwf2rU^n3uZfw{kFs+Nu!EmFhUO51^V{`OOFayK
zG<vdi&r+WOv`w-}Umt}qhR0S(^lGKA`xa18Umx#g+1#K%<_{_>ciWnUel!7A1HIpN
zO=I`^)0Ic)9(@ii1(Uk(=8ar>HQWb@?eR>0RS2<0IPfC{ph(c}+f<odB5`f<Jg(nL
zV<7CAdrYZZ4raTc2xY4*Qst(VC))SxEm;W`aGEYX!0~4`Q%L4qDt`8!y}er0s^u=f
z+jW*eU<#KTxgA*!4;gQ;6>u9|R*ei$K>~Y0{Y$ML!Rlp|gtr~1zT9SHJ_eM}57)%<
zz`Z=U2R8Ytsg<z<<rxR%oos3(A|EiLnfqlWM<_w|=dJ4VYN{OnzMU{}{=t=t)KaXd
z`#<NnFG=s4dta^qN#P#S!Qn%@s15k}-W-?lsN&92K#9FasnV~q&St}_DY=QQky<#h
zWXq4Qdw(hds~8EL_;^2VnRq-@)-l+)0ox9=gxRg;lR$j1W|o0m!6xFkeV(p=Lkj0g
zMbA>gwmw40-5hUHFbnD8^*;jr8lx@T_A4E8kN~(sD>Sjy4D6XJ;7mkfyIJ<Y&!h4D
zTXS3+1PiJ?u;3vpuO+U7@9yb88$VER!r1ahD0DQ8sC!BPJ~LPqw|MFRIK2Z&fLO+`
zp8<3Pfl%BcTsb8-oU`hmNk=v3Yo`jskq|Uh625l*+@Go#NLC5u$FX<!HE<q&bE9J6
zy8`i&st5F;!WPStGeW6LP?pB$WJEAdxa*UUevvVB8ZJOXdEZju`r*q`th4cJ+G7$(
zzL2gwOJXyQo=+bE&bH`%8hUh!M1Ij#rW>%GUE^m8;OKeiPfmQQ2REe1kRp#>4bvG2
zUX67hslqdvT?1Ug@;irKF|D%dtj2roRL7uB;1ph9@m+}*Jv__1<>zXw%NT;-YN{4R
zjwoW0ELiBMIHXcE_bU5sp`83_bCx(ddh0C7vY@bC+cUlGXO=kMo0nDX@`4cIlc2r*
zCX{@$GTaL)lKe6_gc5$eWu#WU+he*d`8)n`Q0V1`|4V8x5}Y}{ESk_Gotzct0i343
zbU&{@`oNwe#~i^2X<F~fv)vQQ+hSrmjLJexyMeWKj1pCJDYz(8Y<;uL7rS~MEHN0i
zjU<SIE@an`ayzRP6z(eu9^ZHVLZ4uD^@QsI*~|$=E3Hhv2vv+v$v+BnNIxOr<fa5{
z>736!s~n*vk7lHzVTm`v9Q?M%vFB~z{MhiKhw>C|SO?f-C$-qWn(gFYtlI`V9s()N
zv&ix<h^wt0fgaimct2GA*;OJYQjj!jG&`wU5eujHooBs6wr^4D+rX%r#bf&kd=4XD
zxSpy&lslol+gjnS1K4#HG`DQN@pK{aJz;1^^Dq(&4B?Y5d4zwxtMK9pCzGgD#^h#5
z-Kh^PP1(D(Jb0DoW`}Ey9gi$T2oG0PeVPuo&!#tj=?}c|UK*B~0!#G(832d;Tzibi
z<+B6+hz?{GC005oyOg={q-)7V;6oyFA4B?gw}hJ|(32-VS3Ue-^=?=lV{GwCFJ^{7
zCn@C6$1-+}XB}B(B_$>v{_|Kze-8E%0??VurW?suFoMu?VxiKfww60a6Ya-X&&BAn
zTT~}r!N#?U4_{>BDj!*OqX7qY+SoY_lOy~t_tUhdLiYMbtFL|!w!b)@7Z*R%s)LlC
zjCJVUKs*4|I}xWfDb^{&H~`&(0t)MI1P4*+^SR_JbtA8Uf{4^U(@1%c`bx*YT+8L)
zVEBCmjiE3If%RJ_M=zO>6GFFh-^dF6-;)y~4!ya-!R{kinXj>x@<%(eQk9YMMP%O=
zz^j^#K(?d2;pU)d$Tj=!e`qjTe7Z{HZUDQF&6D_bBc=5iU3{j6f@H^K$ZvH(<8F8q
zKL<MSBUU<{jL3%oF&b%mpEIVNLqt47S`84)dBd?dEk_xI*Y5Sm(Dk>k2ZMZaD71@v
z(r}%E(aPoJ*1m>359oUNrI_{<a#%ZaOoulj+p;Wv?xiWCHWjb6qu~in#`v9*KTYi`
zTb<0ab&#+lrK|1A76K+OCr(As*gj1~Ae7#9fRuO~GE7H|rF}8bEtTzQ2`QY?=_>`$
zTDdu1sM4~YBaX)SidOWmu9+lTEND+UR$F^AChPdu1|yDwcXU+2jAjvyRh7~$X~ikO
z*jOh3+9;di)@qQ}t%FwHtD1!X5bR*7TgQx7rgum&Go**h8y;4l>VY87kTH*I8d_y#
zta@rSkAMXD5@=*}g_(6}NMGI0zJ9=-uteKWN*t1f`$mgyx>!ST(SJaQv{N!#x{6l%
z+`S?C)E#LdcnA2Y)r}+ri2zt@?)v1^(<`T}eYxef#eIh6w@s^0y-zoH=i16T6;tQi
z<Y5B^I!+pL(2dHz142`XqyHyV%JM$khzJviJIbR4;zM@wlyu?}lJM6}`oqh5&=ha3
zFabq~XiFQqQ8p2Y%CT*5KVBPWNQoNquuy9fc1Dz9M)YWHqZOU*xf*m=#TIJYyh4+t
zk2x45@DEt-sKFLGv{u5VqzYw|37tw7Q9c43gNe-AgiGU=^^T54fSqif4TZ#g-?{wg
zCxD~Zu{({)SVIVpIFs-PAArOF-oJ?2%P)+tEJIp}hY<^)yj3iOW>n&r_@5esP}%Ms
z=s7<8<s`xy)xOT|8DZgB8eXMPTMws-t($1-krERsmdgovtTj#gB@2*U+j_;9&rB8z
zR%U2K>ixnFK#jqtXt2F~))(nHO>Nsz)rn$jQ!L<sRfSq!-`)*q3njcR6JJiOWlUZQ
zkjTG#Z(^9C%BMG(AoUTaK%3>vlhtByhMMA5E~vYrAvHUHt*2$c;Y8(50f{07H(@g|
z)q?cH&HT2%gn_+vbH)E9lF(D+c5Mp5MGGX7sZ^?>@oia&CGm9={fuH&LfKqUk!tw<
zedk{agFk)xT%37V>au+3O{8ys<*f7-O>?xKC?D@j>KQ&R(8xF;kfQ9isY_2Rn`vG7
zA*VN=1A*}Tj6}2g8m_cDBH8$WLr~SMH$3$(<f$8MN6$>^-%@mPc1MR#pZeqqARCe9
z7;i}|Pi?8!G#~&({xli7RdcIK53P$@(p=p&CL>jV(gK{`cw0JEb%iNFJ*-Fm%e(sj
zZgK={@0@XwhbE+jI5m}5W}PQBg{Bp#32J@%=GeWyGg*&^q=K})A$RUVy3({??d6HO
z-AuxajmmKgv3?V-t{VzQeH>Xru@BF^pm%YA-Tl4i$dBv=d;W805)ua5Pip?ByxYk-
z^hTsA_H5c|&uauC8YubFCff&3W|gwET4G4|72j5e+%#~3COGh2O2#^}rnh*(QQ#0j
z^meQ~iL{&DR~E5_HeX5}lus&^w(14z)_0b^R}SsjVfs#UvH7eHSQgV&sV3;P*UgV-
zXMef-1m|js^P`Q|n<(eA^Q;S^oyxeab^?W#FBtQu7|>^e(?K1pkz@DzBgc|S2Qb}u
zMz$-Neq|Y;!K~=*s~dP8su2R{uNT<Zg?Fi4521JKce?vkbJz#%ZRPFWG8>_x0zi~V
zfQ(iIO7#{NU=Qf@4ll<=At!;5wT3x8*1kp`070|o9JX_EyZ-tU$h0_0FJiHcf}Vw`
zU}9Ag+S^t&Ui(x1(m7cb!7CmFofd@(NA;@16A@54;0(jheFP@5sa)F}uC8D`breur
zt!5U|S3Z3gn)OQ{cM24d4Ui~Hk2xMU@Z!zFG3{+VQEvM`Tc$&uwil753>!V+TWSa6
zzg0h9{%^#M>4A-l@VHroznI$LFRm{bD2mhH$j|*cMP}L~+c6kYckp$d2}y3Nfy=Is
zNRGazAn65FEw}de)ZyYusnn5$)V9p%`zP+cB0&ofM63n0R7Xld6g-dQo|6A{q(FbB
zdi0P3@T1M1=b5;mU|(U%GTnS9@lOz9^LC{(ts_uL*~tDe_@0#QtqDV|yf;t|*?{N;
zxH{rwLJ}f74c^2NwsTb`^mt1`#{qu&hkFQpn#-y^0HwHuvG{@E(J1qHW&HX#PV>zy
z^ybBA!g?mbt~<NG#CiC87(c<$G*0?^oi&`vbjc)@pSnGIh=DLaWv<ADw?QOgt?xlR
z*arR!I++wBsh@g!Jc(%P^VW2ur)*b?bY?DP3Fp|DM3o<1itjqDEAX&3nik!=oFO+#
zy!VRAc2xGf_wqEeXvJKtYctc8l~4K2*Q5rbKg2J;Yw@xA_4#h%dG#EZM<w7jZ(3Ut
zxHm_GytijW`Ag*1DKG6IdnKBJ@aPd=#HryfqZP@Yde=n!G7NL%phb367#}wSAb52i
z(S^#No{#r)Fh%rp+i5mPRW4Z_a&qRR)3!pddxB1hy}44KPt03VzI6+`@7!aRwQv>N
z-U8c<zWCP$2bEX)TvM;Vb6K4mqFhmV#l6)pK0C$bAW@V)z=mI~E+=9U<rxF(40QM6
z)|-STvF3lw?`Z^Lex*YlY6o`&Jo!89NiAX9o8nn(?Nhw*wFYWD>o>5J@so3N0j8Ho
z=7sZ#HZOkYD|k#($OON%oCI7|u}Dqu<~Q@rZzsYOU*qOGC{)*4LZq&^1s^H#rzzaQ
z--uAWwZ0coG9xMAX4PkXf*d1zT#Y1Ev}o(}=HL$NIrj{@WoD~i8+F!@QyAGAw=VcH
z`9XU(d%U+@@juO_2q(iQb4aTMbGbl%Bz+I_;aL1~PgAs1@Vfn$kdRPMnfp}WvyYj3
zeJ#oE^AK}{Am${)H8?}3Py>6#?kyq?G7PKNaP7LRY0ijf&s1c+hIoi#AG|EaGfy~f
z5B%vb8?~GIHCPwTfX;$IW$N>@+As?z?a(%+KNJ|*j#36Hy*qB?4}Kbsj_})X1*B@F
zR@|8Qi%qK;{8DZ4r|Byzp1qF#1w0EjfHBNhElsJWxlnb_>F*Q^Lw{)|M3aH<ay)2~
zP_Q=)Hm2jh(Ja{PsfZB=`WQ%>+@VfGqRUHcy!ULWYm1T$kAN>)^8Kj!{et?w`p`_e
zuNPS~o4Z}aSmCJ*xSAxm+GGbt`!3lV@m-;;hA~rU{hUw_yOL?1>1|ZW>x<MHf#Gs)
zt9O7p(w8dnu_A?N5!0W`0@|ElA5&L5cV<F4cZVLm^b*^ZWrA&%8}DfNiZPY6Z~Fjp
z=8a~lrX5*LLrIPbloVLPXL)1go!4JS(9{lP6p6T*nqKjqc<Aa8+qx%_xE6+RWAcH~
zAMNIlsGe4?l*$3bSZ!KH8dH>(DAHwFy6bz=Cyp<?_j(EDEA%ncH81#jt&K*ZH}qOk
zXFjE9$tP^2mi~H(HQQ>kv^4i|A7W1)eT;N>)QW7royP8Ue9OV=^FFkAPt{nq@X#6>
zBi;8y2odN&YaN@Qv-;C8uQvi@A3oT+S(!p$G0H?__BTHZfyNOS)r+j@#I3PRUIh;y
z`eFiRa5<)JI-D(5SJWaN-;KBL^k0kzP_vDrab)NJ{X!F>CHbvTgYi<@G)N-=JmFgA
zNxN6lDuB99cF*Yn)`O{K`&$wDa-EKKBA275dnc${=!<dqO9$E32*W0k{gLdo)fR`%
zf7#oS{wemOwhlce7M~&<m311-*<R1#v6@}^5G`dZzE(qM7JTR2oUhtqA~wfEz!zt8
zzV(-dr2ic|*ZPV%s8M)YK~%DUR&#nOzN7QA8JRBBJ$WvMcF%JmXJx~jCU+u{g|If#
zAA?zI{>)<xjrS;`fGhNZmB|K@Hoe<D-gmIHr5_64zI|hrUNH&JkCDEmG}pkTBoQyR
zt{Dnc@y%bL)wpPT1=_u-#c;XCE2WvFQD)fZ7esj)zqV5wIh85D?UdL|+MHDFzWus^
z*Yj)AYcTJytm&aX3}0qO#&da_Wcf_n@RsfF{9<#i6`m7ZTc@y)6mOq1DgOv8_>QL0
zUXMB<>tc9w96I}9U0Z)43Rgv8w?WDg?d~CK&H(c_?3<KLKZWJ_Xsrg<0tR-|*3yly
zmy9awF+uG7^uBRrPWRUD9v=Cp$s4_XnwS{pV4}H8^lp}~KR~DQfSC2~e`f`nd$&oI
z)D)9{5sqdqc1KF@WPE9YwNqv@dvo62UO((I?`r7n$?C-1pqAV7KRjn-2rsy$Ca3dT
zLigDyDynQ|PJr4IPDvZg(34Age&~kM8HKQ0;by-+__l+X%-HeSj^?_B3%_xeC>9$@
zR`#`Dy|VWBmQwOC`UzVf=sfdC&+C9gOt$+j46X6^D_IWZXq!p|M5YR$D>oM!UMR1}
z41og#9->>~#zW8{^w(1Wq|87Zuz)@>Yui6A!uhq{&aU)STYC(d8GQukHaC{kSuB6m
z3jImVT=_ySR$PWe0Kxh~x+zXuh}+B0hsv%9afUB^pL!p?a>m>4=;31+H|jv8?yR>#
zxofmk4{lASEynujJzA6^kea#qpo)Oniy5S&$mYV!PPM^th8$LEg5Ks6fj%H2kayN@
z5cqwK*tSDLxATF#k}1klFMgKA2vX!^FKL{bis!{1hi~Z)Eh1hKe?c65N*Wt~lcGgG
z=2(1LUU^+%?yQ>MPxpQ2_pARGPoQ`I0Sjer0dE2AgTO;zK|NMUb1h&D#9XWyJK-!r
zC2<Zm%Bke`2(kwos|2l4x&67i;kQv}^&`5@b4NkGfb?8w7k^?Qe})cCkl(ay&EJ~X
z$|o-N&E<d*vIAp%UD#~tdGIA`DXF->;&bnttdc%X^-sTPe^a(Z#PAcO<g<Ht6wG;@
zH&zf5ZZM7)bkYpgATX><DPj%q0ovA-dPetoaBy-NmnNExehs(B;p>sDQy6%*PpP0=
zk!xFcHBzn7?RKg)*+cw<BI(WS*s>*Enoiaw5=)$fWJ}vm2X0GsH}|DDy(RoXf=0eY
zU1*xkcfI!0eoj+Yw$D}uNZ(NA+S+<eG+%&f`)f3^VgV&4A4JiV<C{ApLlo<{wUiac
zrcy45OLJW2`xV#wPycXAQegTQh&D$C(pJJk$)i`7v_4I=r`FA^w@ftK^?oo~Y)RoH
zs%c2Bd-tz^f$@WZb%BAo481IcZn+4IVpW!E;77lPS0ju|*q@rM;})fIsCzWTgul}?
z?F)=mqyx3nFK{>#IC&5AsE{sWQ6^CnU+;2F%sGlwsi|i<&`Osv=!Fm-8Ik5ava6XY
zN(z<1(%0aP*6_wz4=C;328U(T+8*Y=Jc!=CL;ufjF}}FyVwZuz)4(Z{gR~o<h&G{5
zhXu8)&dwjK2@hT;6K5M^=F&12MWMSmUGy7t6b)u_sYt`j3O~_T>S`cqF+EB&N9QE7
z9!(~UfL{2=E~KuW)xTv>Ro(P3|5ZqMbvmQIUtrnB%)(FDRJ0@BjT=!V#+`LMlhAw0
zpzEcw3}=wIBB=0MgVe#2)xI_T__5K#FHUgXuJ^#F99k)xSbd85@iid%QlUqEl->5Q
z-)2~v<@a9;3j|Hf2Uae4Gl*#5%^tkCIct&v%dIxsk&MNhRrIJ<vtoL9b+M?t8P2bA
z`>N*$7y(I-2$IjJTGCxrrRi)jMYtpFKbBN&W}{T9FfU`Hpawcj!=WwVJjiVIGpfuB
zm$i$feqV+t7<$I4Lh8mY>2=L22;aZfAP+$TUKudNn7ejILQC=5PKCtolLk0UJ(SNI
zsS%w3;V1iX>BulrdANKnvKz9(n){#vhc^2>((SSGzIUmc)o<f_vr)3OD(gL8CW_q)
zL;n}_*Es;O#1Jr>R<7AijbEZHLv1L+OAxD!Ftn!4<b)&;qzrzSHkh3SDJrp=Q9S;b
z^;;mY3Mn^oU23P0&InsAA(%qm-*IuWOPz|IohxT22lGAp)ey1U2iv_JH=xOL8X|)Q
zfJrB$(Wiq|4XgY(*`$*zQa0H`V~f1zd497q<P@o<y(S!8ad;ls5ugRkV0bTt|Bb|T
zUBU*=^j0A<x;H*2DroGuD!J`3J<M*k#6l2Dn9K@$?>u^KSJiO2`PG~h*8Ol#fT?~&
z!jxo*2UN>-p$P<cIG*<M8}8TYFf8uaSqqG?DM6v59}&)PIc33E28X;82#T#LqGL7W
zLvJ6tQTRKtyEdVg@X97@sQm+2)16+Gz|%*S{wh6YWaZBp86PvCozDxeg7Jzp#>A4q
z^)k@9B+dq2iYa?F{ftR!6&Z~he{yf9OJIwLq!P2-LEQG;<j1=YKuND}&#hLbWcXT9
zg5}qWYZ;NeIH$SDHL(6S_Z0)N;hD7|b1l)do=}Jt@w-`K^N?A+Nqfpwq;ujf<4*c@
z@F3l8p+@FXe?N5f9LkeSraO`W8y8S+Br<ec0A4Q`ZXF$BNBHF%Ty{&OIt(j8%+JZl
z%Z)|uRL0x+WdMZj%yu^{Xi+Btkf8r5n?f>>l5{~&t8M#+p8F+6Hb1E;;!NV9AjgXB
zI}9^VDtyQ6cuV*Yo;6NAi_PyByxHbiq7^~m^Trlaz);i})U1DVS9ez}7lP|_a?(k1
z6MLjGA<15XNSe|y=qOx#Y+qPV8Bcuv&UyL5fFcnsy_dJdx80YrKn=MCbg&v>)s25(
zdkfAZ@texi(Bs_@a%PdlU1qf@>CiTkv`9Z!nQSGh=BL31SIv=u!JWvc+Czl(Bi=v^
z|9zP;yY7QXz*0RA^>Q$?l<ab*QLShb6$S;SPPiBu8X9&;Sxih@X?-u;+HSMATZT-D
zP%+ojSz8FQJ;G1#X~F{vW|gelVx@@nmss_RmDee$hlpY$?^nr86F<M0YviQgdQLPy
zaAnutg(qyc&4obm4&#yPp<NS^1}EEJ85sqbN5if96f6JRjR#m7R%X>GY2Q0?c6*Dx
z-QF$|UI{~j4>#eqG-1#S8Qev>;lgJ(trKY}U~-;-B2-`I-&5G@bEExqHx{%i`b}yO
zZ7%>FW#l%9(}>9{MO2LiDRT*UTF2&`_ab9H4qBxRMUMU4Qq8M7IH18!8ea7m#1fzI
zNo4e-LC{@q1sLr#!-?zV@8yV87$tL`N=`=7?EaCGV9_dBM}dbecD*C~Rk2G`DyeFI
zE0xQ)3-%*?;L9LnxB$u`WVAK8TZ0!tN}v53x*VS!=J2H!HwXlgV^k`(LGm|JH6HIR
zG8vca40x`mYO@n5d~WqbnVE>6Fy3>O1#S+V&+jQ>Q1bg>utY{&hgbl9`~^%t8f&e7
z-2n(P01(b^9>@K?P}5DIFQsV3sLkZ(#&s6_{9Hr0E}FR%U)k+DHNVSz;C!y+I{oEI
zSYI(Ra47EtaM#*EAzc$B)FQL`T*9Y(eh=bL)aEo3@S9zfBMaZ4EaMC}ZEcE{UhCse
zSVEj(!>-c0_GONua>2JD2C5pF|22^I3qAF$IWR#i2pni~|IgGEkXS+G?T0Tn4}0pf
z+1m@Jz7ejlIa|Au6m%>sgkPoVR=$-ktAha+yG5R`EF4&@r{r*No%#tmvPB8#G63VC
zGJ_eSdWxr>_dZ<X?Sgyy1)(QN1m2BR0GF`HpafG6rN6vVP0Ky3<J@l;sPfd_(u#3Y
zUO~4v*j3h1M8TWr-uR939aPtR7BYPDu!b<O136-M^e$0Z3G)1Jm%|d7)Y&f%^yFI9
zSt~ET`!3J!qn&@+e4|07esat@nXn(8^6wKipD%-@FnZbv7a0;QAoUT3z^{DwUt96p
z{$E@oE!O2^z2JMKSm-iZ@7gSvI;7x~vE@iuh?&3_4p`7QVL<Kb^L}rYRnXCb*OWv5
zz)Dg{&bH<f7sKh*nc?C<^Ej1~Ohx&?ib#;<due@Y^;Ow6Z}Ih_b^o`c53ra=Fl+V!
z5t_jG_4#Q5E=#mFw?*~lZ1wDl07oD>zJLAch$obl+cflc!QHWkEsigj?P;7DP=@3k
zmBT#%XN9Z5fD~7)C;Xs(A2t5-ZDQ}Mt2@Dj^j|RHy2MDQ`q5LHqAyEHdDDj){32)K
zTGafG$L(f0NlZEv6H5NKgd8TZ9)BO<)V<9O4`p%x442>H`Q4>{vT2s{MhkUa=9}N*
zyH?#8>GW+r#5=8J?Ee&dnLSU5=2ufu3`~W<ZtuHx_IW*LiKLsbt-|Wv?l0F$S|5f1
zjVteee|4r@!H8w^JCugZGtSObQF@q)Qn~Gx;h<GQMVI+@x9bh|O&3r}>4nfQ?S=Er
z?k*;f(#f$kdFAp@rRZ*+i~t>)6jdG%EqUVWT_fdwO1P1q_Jne~YOjs#jT;%NRjyi#
znev;RK(oZUTPdx^qCE_PCcW#hgqNKpUPG{>*sHtt9&mt0tq8PFpuL$mVd(}cJg_H@
zK}2WV#e9rh3DkN+eO1ZgWMgSxl)4{&A;1qmYrU;0n48-1q&8e1ElQ7RId@J;X4o>H
zGnwF|2X5%b$j-|pVR;L9ZINL2wt`u?hJg1EpZwaYnM7!|NasS+@HNPByxq&&8=HPx
zVH+p6h&@`F<4;;_yGB~*pKo2YBhjpzZ9RxyeutwM!ygk~&&wm}P*!Nqk--w^aY;j?
z3Z%Dck0kW^FjdSK88luhubb4J6WuNk_&9FXIh2Ew+I~P1^|89pK=|8Fd$cQ<Zj@Hf
z)q?qaWGf&VTF$Q!1s9+w=v`01utdm#w<8=&<ts3TlPtCShkiPdoWQznw?x6rey50b
zfE03-9bszlhK8^$B$84dKGOmjk+Xi3I`dyIQ_W0QYez@@iZ>WpogFehAwSZFAE99=
z%>%kM`dB6r=#Ga-zpZ}aPVwviZR3tw(6ytjyrBA>NWMOH4{mw}rewlzmJTg}ILe`Z
ztV~r8L<LIQYjhh^)PlAh4Lkf)Ev7rG4t9VBfpcshv{FU+ypYcqFp=3z0W-l^-iyO~
zjuGH5AiwOgLxz(97x=iiQ?5L#ly?K16UkDhVEm8&<~<tcYv7JoIUw;6!L)$vL?$@|
zgEX#^E^5X>@Pkt2>M=O)c3wrd`=FYyL+@etIc970U7Ues<Ww@@<#p(hn+gL3kcPFK
zQpTQ|RQm#YYxT{!S2j$jg-7u7bMSKx1F~QV%n)&15i{5Cpbo%*L}-=D1+yPHMDn8a
z=4W-<H-9v=gktmHUJpaCot7v@Vo1lgO-<*u3KFLhFW%L^FA9@Ps5R!^3MkX}-=$p#
zTL<rqc#=NYrDPpY!;g3HKWa_mJOF1diwA9^UPpNz(EYrDDv?B<#S_q9`R5uPMmf3g
zcl~c$(Kz9wE6qQu7=i+|d{)D~_mI1h2)O&*lr01hU1XXu><antu_&&sJ#Z;Fr%kgj
zTjK@i;(9=w&4mo3f`}3XU{}l!?EDSH!1ulo!0!KvLZYw*=n-&fiL2|kT8u$lNy7a7
zGCzdATL_K}j~jh~O$xSHD$iMbE5!w)iU;g`R{zbxU;}skZ~&(2M|36u0_)%K=JjI0
z*_>?eNJ%E&=It+sne5D~1>mWrw1oho(e59&MEkuy63veR6#4NSRt@7wxoK!iaA)L;
z7DVCU1Z2Jgc=SRQ@?ku`e;=)DJqQ-k2@gNl3ypHS^CbXDp;`7$SA6(rs^3FJC#RJn
z%TaUbb!G!N>O9Ad?5*U}qRdhyT0MQpD6p=S#S!H1=5o}-6rH_N6HJ#6B@-@4!5KWP
z012X)erWd~M5<nP(@3J+a5!YUeF>ktBMgb-dw*z&HkH*-L|<Oqwbw=m3OR!B*d%;&
zV=0=qC1x+@a%C#GWciBke*8WEqi@j31T?E0Rp2p9sc}vQ>}F~G`%)!~R0(_pVQL_s
z#O5ojZ1h|Nw5Atrzag4%ybk=xFPU(h;4?Tznc8IygHbyKYZ-`CAYF!$&!>}<)CkvZ
zr2R%1^oq#{%Hpn(_li7Q_<|O4cKI~bjL*GduZmayN-Sfb1k5-{=E0#NHojtLNFZ}o
ze!AZm&ebWo(|K;Ku4R0?<v$g?yHq$q&5!@=_HT4O;jKK2<*L9I-SJyBW^<hrt!92m
z7MPI>$!l}T7D!k2;}o_9=H;Q(s-l&R>$?}!eoo+s6t+=V9fI1*5Zf{Rv7I8Y&$~rv
zR(~R|<wABtCXxJQT*=sDcGjMr{4?m<=Ez_UwArm8D%l((4EUYw=m(-$2|h!jgB69k
zD&*hj`j;j$FW%8>O41uZDZ&TcTQhKLgyNbnz-jy}S^CDn4|B{(u&r#|o{=)-2u^bJ
zJDWFnDxNIl#npa=cJPoS1j<{;2Qqf%g3I9xCcMChaW&=3Co-YUR@FRvzh_Y~pfX<{
zLL=r6bLt2v&1`#@$W%P*Y9yGsBqeZ^+Z`E&QugjkU0A+?=7$fq;EeqHkc*mP+%o+>
z+{uLY{Ing#&);>AY(I8mumj#|Jqcf$+4xic%Fur4*dRK-WQP=&pM@<!ggD@NI$Jyf
z$G3q2q>QU**`%J8sqh6ceB!4RZRX+$Z){o)I6xnw{2KTeiL?)31Ox$uz`0iZzy>y&
z0`N-fA(Budc@<n+{g5paXj{K6M$PLz6;E`M63r2$y8AsF_=$c#ObsVi1s+g9Dx%iN
zQ1V-%guW0kwW}nZXGV(o0*WecMkwPgjkA^fPsDv-^px>7LDn!()flFBe5)TlQg(%p
zmF!=ic_uXNjrP=@{K?4@ui!YjF!wL<?Cy~&AMPzv@17_LH9z~G4arsV9||fSRI_e)
zP@i}0PK(&hq7wt-{x%?t`mcDlkuY|257hSd+T6?<vbC3~+(5oz!rxp4Fah~&ij~UT
z?W<{pQT=l|D^zmAnKGs9PiF|rU*@h}UCo?(0-3LU&Ft9rNIQk9dBXnamXd8`*oN7v
z$|`x*!+T>^zbl7sTk!9NCfNJ-9P;T2R%zO=Rx8!iCY#c9MEsw~6r_AK{%|}$w+pQn
zC0PO+==U8ZtMru@5y>GBO3U!ND<-{f$R>z>D`8EzY(U!K^-SxW5a8fMz2alWFc;M!
z{1n8CZ*|@Dg@AF`ocW|)EDDvm;!fK?f$`Yh*V5}IaZ25V(>>#5HGEj)vp195cME|~
zB$=kJOsEWa*rY*BIH@g|8+Dtf-&y7aw87k(Oie(Cy9~<9DQ726{|8s6DZuP}Q8m5|
z(iVz9+cx1Qj4A;*dk_?847l^a$Js9m>&r&4Dja;wgsb@2?qg#i^de{g;Wo6s!8;Ia
z_$sauCD%m3#4LDAK*b5ZvB6g1yj_I{@CBbC+J*Wd2I6DTs$8>Ea0{B1ak#m7zrYT@
z(uBI;sIf_*hJ--7R<Dv5xBliGvG?L{p=uOrge93EfHc^N9+U=H%`nFu6w2i>k=zT<
zc@7#RtzCOj$7D`=(Tdd*rc74SHaq_Gd;u=QuTjYa_oF-QdEgUK7y7uilsf<X|4aZu
zQTy%H{Lq*`J?GcR>%K-4$?bf9UUyNYfH(RwAOzG?(PV-(KqvjAnbwc`zE$*eo6_Ut
zoxguWB(HNrFAkjHY-&^JVcrEG<53rmf_R1u-Qw51Z1|qmXMfZIDAeyC6pS<&FmeD<
zZ9m~(lac-0tD|6+k$(*#$6>uOpEr6Fn)9grP`~FZ{b{cgcHZ`sNZ!lyr}J00^TU%6
zQjNa){@DT0VX#)_WCE)S{0IbU9eo})3&F^G&V5kxGv~IY$*3XihX%((r<5FV^Y6x!
z{m?HPkSB4Q19a*i^synx$-_orE(FQXSGq}7O(hEJ#e3wSDTI1IU-^I9yVAENuWkR@
z_UQGZ!CMDpwg<#&5h{Zak<e2Vr4%TosDOx~41&x8f*Bktpo|KLB118PD2NI|1T?YL
zA}A10Kp8@qiC_q1gg}zJ_G|Bb&hy-V;C|=_Kajld&K}nATfeo|)?(_?OL-CFcAIwM
z(Hr@H?bMMhC;+&PW@u3b2~z1dDiU31+)sG;)4cT1*$4gTKs#y;L&Kx*@JF@itcQP<
zPBh?CNbBHU%YbAvp!T&V8#u8+QR!&zBj3tztcKzk5VNSu=mP`zz@E;kxdqiS;k46V
z%lE6H{xGb)9mr#Tkl0Ecl|(`m$^DmIs<P{QXd8i<JxW_k*PBxL(^>JEe~?5r*)EZm
z#!;Xf*mvg>M6>Y@H1ptN>7@|dfhN`-Y0c_jd7;ZeurrC(s#fRMaLlbIuGC$F*-;v%
z#F|#@bQl|&mtbuh_Rto7zw5SpYH2}k$BJkWb07{=@I^I=-Y%U08#1DBIU;5}Vz0cH
z&|)T6LrFI5x35Nfc!H{`XxSpmeZ4|a+u@tM1pQ*gYgPxBHm|fY)4RWUe!JYp58z8w
zE}CoX7?&rgem*JpsXfSHI?KGF@HL>bNyix#2gRDpod|$Ou%=4zRX$y;N)6^K7;4{B
z0nnw8$ANEdB1?{{1J}>Qy?)ikIUV}_9$LyW6L_E!f8Z)IRM!9eKxYk20mY9W*0J5B
zG$nIZ5x!nwDR;sTwCNHWNO67~2C<Zfe9EL?KdDlcfIrtGQ7j8Y`|$lvwEdl%=qv~!
zYtf!eW_9>Na19p`VVI<mcOO=5j#e!{B7Vm+eUXorX}eblCMlVko78v!@tj82Q-CDX
zLlZ1pb`iv;E!2J?iE!|J;2}^N%^WiZa5$)>1&b)XvQvJeux9^a5uwD2@sU)GQ06_e
z{qO*;RuK;Hz-bk4+GnAEM(Ckit{*STyII9lneJiOYb!RwMPL|~&os<XCJuN<d<f+-
z#nSDZ@c5t>m9wh+U_T)gCj^&Z))YtPl)z59I5gX!O!wwOm7zW7DYboO=CEWnSU<Id
z(U#aCI?E7xxUgr(e-QEn9-r?i<w_WEXTsKo*2lMGX`8~5!XiJULZ8Y$S<5L1h?YY!
z0~WF4OOm{A@z(}rqFI0ay9zCaeQs+8A0I3U4YnET>kdF#%^?w4eCNL6lzQ%$Y<@cs
zerwQHx(A4@PI6L>JPYz_?Pa`Z$_jjGZfN^mK+h(@2!2_(>IdP)zG9k%xavS?5}ZF=
ziv=tHjLrfs-?9V>Wk9wPCjhh<Qf$PpUaPhl^%a9)1HFb}G5%uNEl@(kM`Pi>JLn+T
z-j!j#ik~Cjqy0ICpMUuiEd}45{a0a|Fp}#;Yi;EG@Aty~de<MXoB5OTpxEKpW$HTC
zt8DKOQkB$6s8G7thp!ecRl+jE3Hqcbz^v#m4ld&=r3}@ER(<7=K6Tu@XgO`pR5ixB
zrRTWW#H7&nv=si`EV6?I;C6p#Ex0r~cn0k_n@1hKSAGPoYto1Mj`iD5yzdDt(hcT&
z1{UN@GGaO&9e+E+zH;RXD!e%5^`0lV19hL-ow2-|TDA57mRxHY_Hh7_e(U4IE<@0U
zZP4Ed#VMSIIQ5Ov>;!?MqPb|XIeWAF&3B4l!Huj3LtPV|6vm-fp-^pcFJ<GCn5lud
z7~_Jd!Dq|ByzE_%Mq%)Ks&C)AWw_>DKx<x!*2n8=rjim<%5hG8%#b8aSt#X&Z=KKP
zx1ZN<J4JnpC>QUOPx*C$hxRIRKVJ^;8ykJl?q(;eV|N{u+`HPlM)6b?{ktuv%^6e(
z8vD$Gk5Nj_!sMYPVW!GF&^TH7_UtZnKv7+&-lpBHx`SG73CwFWjodAI9?isx2EdC0
zLD<<OcRnAQGz}G>DPiXFl?qpzzOT-L5S{6d9(;E8+_`g|j=d+;bmz<`EKWzw?7utZ
z**0e3Kr@{r-g)q@KmzvOsTU%U^1kAX45c*h6BCAQc&S4oKL+m3({>h5{whcFOH)qM
z9@ZC*{p4~~EGt@;9@Z1;09SmN<sFm00e~9PL*Ud4D1lYFy*yJCa$!*|P&P##&rKkC
zINz`!Hh5+o{P^s)mAntGw1FkpuHF*%b-iV15<xO2t_8ima1zA!JChH6jPU0l$s{-4
zvk8k3p^$ffe_Zw4rb6UTs7OgJ0Evrz*LKa!|A${3w3~q%A}@$&i!SN?tE@qsWhJRE
z!GDVxG53ipK?~i3dkMCYGnwwaAlDi$_;DoW`cmtL;z&hho>@lwW9G`me(TtyFF>mA
z9Je`H_9BM+Z_f@74<|mE@!((0eJ8o}DlO+AK!=T+I0Ri~elLN8LTR>Z#wS(>+pz2x
z50Bkl=1cRo9~b7Rlu2DV#)-{!f5E1)G1URZaoNFlW;MF?56DA~aUrlA#N%L*aZo?S
zrb?^b7E;rI{Mom^>cB=zc|$!d9hG-Y^4smS5<KNEl^OhYdeNLK+e)rxiFSE7tp9Oj
zU=by@(`IA{@&T7r<K`bI#zu~5N)yusKNHZAEAV#sA@^St&93$6bPx}OB{$fqM0Obo
zfh{|UbKi9N<8LTcl}$rvvbo4o7005DuDExw&g|}uZPCEIY81!<yPS;)FgVHXNoCE>
zT4Xv<Nx8pTa1YwEV1(LUc*&+YH@1gr<+NQ}u4_$_ebH78SnCbc5NCxYF&~S)=&YY?
zkZLi(aOf~Ta93=r)Bl-?^ci7`?19TnU3tl6U!R|qU9@;f7tc<2@EDTnIj;wY{{`99
z4^WU$DG<(2{?L;gAIpoMN0{gh!#<oP6zESY7~%J5U&{qc%aCDT<vY@tW;0b$eY0NE
z4kU3t$<#`@e&%YVFI(^l0V*;$_kC<eUh(X-ehWZrCg)^l@A5IPL}ITgsXALNnhYo9
z0s74{BBz`ytaeK17x!`XE-C3tf{{ym@lJy|a{)5t?_bkdg>Gir#z0?o-&Okl(8d#7
zbmU0lM5VcOmERdJ<hr)Sz*(31HF_JY8~h?&S-L6a9e2K3kfuyoLP|pucK2q4rJ*Eh
z_l4iGcFMKi#JA+F;{>eGClclCf0TWWcCHjEq%=HQPK9KDc0w$hi`ynwyA4z*mV5_f
zn+Vyl?wNnQN3Pr27Om(<{`-dp|Myc!J5r+}d~QcfZFHc41xT;oEWiJS^;&!rcCR%b
z7^o|)T*zeaL_pz>>j+B?PoX@PSHMBa5UU`ogNO)I2ig!c&|%nYSEWw_Y=g%{w90}W
zjq+RYnkF)3!y$`vvyr6bi%Ef&y8GsGHEYbBwgD>9XhU#fm2bvQ;oeTmkznAF11#7l
z3_<oQ1RAo=(1G_nQ$)zZspwiqY}k_bP~2ljDqm80CWb7IdlqlkAUs!;KK1y?qF|un
z`8pm=&$`}2*4L2gliLBejDZ+4k%B$!uJV6f7S`B8Y|XQ1QBDLV6*Bw=t*8srZFH96
z@X<fZ5UZPFz{;D=7Xuhaw|Kf4;P309r9IqEk&D6p?=CS9&>AE>18S9HT9JNrgsY+~
zu-9$<MkQpMEWC1L*h97n&{|vBWg-CRpIAt257?0BW+q6oy0XYdhw4)2HpNF;^Vda^
z!h-#@XwM#~VV-*)w3RBb3J;AR7P<iDa7k0Hlh%DhcNh;sRhwX|H`*G3v5-S+IQD{r
z+`{@pYVzqymCaLf_i1SRa6PoZ0sh?3E7<~!QvvW2TN?(Xz2)9F1uOEmA(5Gue(YIY
z{uyHX5&1vad%fhB45dSPgxQC_Vo{V>HZZRaJWrWA+HAsl%ivCMy=&KgDrf2`RN6+0
z8Af;b?V?v+H7yMtT(^<1bf)1?u4sii5g^m5c{^30&W;&R=W6QgL>>f8QDO#Q0^uz1
zRP2n;^0gGUtcf+SD+1g_&GbjJE#(e8NTuYComT1TD|F{_Hr-zsaD>=#n=HK!R?uB9
z^Qj|!enPBK1y==NO`7hdKVHlWIvUjdsF~QqU0j438VbTAhk71Ay!CSQTo_|w+VDY#
zZ_^gE=G>FK_F}g%HJx4oI11xw#h<B6FV>v}c!9LosvI&f8s{M#lNpK@Z9<F~{H`IE
zujOwl{^L!wxja}gp4GuM?V^Yo>*^jO<7t8JyT<+F+;EC%*r4cH99gEONvVMgNWmC3
zL-8?)+!7>m#5**V(#@*16bhQ=TZqnrX*mmqd^!P<R)}yYT&k`;`#c}PF%7Lm1f#MD
z`llGQp$Ypvq1}%h?^bj-KsDTaudlL0^qiIFwr|Q71-U78aV??$CE&M%2$okusN#ui
z?&z(RtDW@qXT~hgw=GB~RMuN6BoiM4UOGw6T%}#^ve7H~IRoZ5f6PYgJljtw&{8~o
z%r>O6+A;hnM7>uui!G+1irHOPm@c3Wu~oc^1maq9+ompA2yK2(4|6bjiv)f?OPy9t
z<2}$m`dGIfK^>oj*R}V;cj>QbGQxM}<7s{!Ql)R`s<&;U{QcDE14F)4X$Jd0C6+Nt
z=`SUB!)LkV(j#}nxAXu4+dcNE2V@FLlmpj51rP+e%^A77iqxCT@Z-}#onH3dC!o^{
zfo1TFu?l%n5~07sG0Sd81JodQkAg4w`F5HeIH(!-@?})wNLN)}1^ANDiK(J3n-!yz
z?FRe(J3E#_>LutDe$3#%rk_f^HT}&D`tC1|>v9!;c<3tc@-;}JG1lC3*WR#FwsO|9
zSUix>p$b0p`e4Der7+iKC1t@xd&g`Rzg^uw`^{1)?F)j-UjdBF9TUq^y6CLi*(kjf
zKhM43mp^-oe(QHQ>C3$YAGo`VegY7@S{m}bO)G&hY3;C>0V&i9LAW=_7}q&8OrPt@
zuy#$3F?jz4ypo*Z&^3c#{r3hqBz)$+y54p`^>z<!qzGd95T}4$zQ%f5C-7koqG9?A
znt9<BS0G44NoBuCp-AEy0}c9B-n;kV0ICu4F5sjK!HfS-rg&>W)l*!&q2~wOvM^k1
zC7$Os{ONHOhh|Xl2`2nCFP#4V!FV&F#C<Xn{a*5N^mJ*ea{Eoif>fd_elEvRmLHBV
z4Ct)LZ%4)xa0wrV_Gm1*fXi$U%@%X1;IkpIps{)f*(DWD3Xgs+96(&o@R)CgUc9hZ
zP^lF$z?_PrisN2-(29Uhuq5)4a(xAjk=U72@lb-Bn?s45=;F3~okh(3E0NxI4}5xl
zV=Ql2nF86U3Z#hJsunJEsmqjo&@5Q!W)!_mJCVF`57?L}vdu}^jY(XwUt}UbqyddX
z6PM0%zm;nTzGT?@#*W+$OHv*a$JN7Xwb5$Vr7FiWKez|f@)EqKNFtYdT{)PD9u3)i
z@xMQEb!AAQ@wAi<fX8H-RaOJvLGcwZG@6Rm-3v1Q6!EZMBeFcJ0@0Znn9FaxrCQwi
zr^P@g{wds+iG6p}Onze-^ae4eJt226jDjOX(}?VWo|<wB2YHG_@WTQeO?&`?7<Nn0
zl}4Ba0j^eAi9fEh!0i{$J-C{ozKH6!Np#ooeP7zh)kbyDY&wvBCF)BLbjzGj-=ITz
zOO7E~tkJX;pVaBl=2WE1iAQ|Z{LScx8toitXVL8>m<5w^-!?Y|$>nEtkhZ~+FXZ<j
zOV&h0Z>EmELaRV`)WV5w+=PVaA2>O*$_AJ?s|<yggDBM=QaD6t0|d|lVchK`isimf
z10wDv`96Hsl}nNE%bC~Rc4-4ZIh&7zEw4wq^~US)h=mXW`_klwev**<9REq2BZ;yf
z2}9~OirAVBffn7dJXZ%ATMd|h!kK<N2Dp}pgZw7k+f5&DLZWr{2@38<ZzbYbwh7u^
zVcge~__&2gIHSC!_)>&$u_l%&s(OfYCX<E)a;IAcKKXb@n0%AI1P_EVp*RgY^CW5<
z*1Tf=6);fnBa4tH#72t&l`mA<)bQJ3<0q?_z5}Z9Qr-njn2d%lB0)UmC-L{?#K-vH
zn1V|t9}<6w&7;k3@&?&-HC|Z|3$^fpCpoLLv=kI-?;tN}i4ak8nG>iDPFq>ii@yqy
z-Fe<i?9*2&QVdev4a?EYc9r?#e|&}h_=1Ql^#&sP<5)AUwtJyG3EL0}GH7Qzx)V+`
z3Mtg`99+#meOX5pbzkMVV#}dW7w!W857;yeyZHAXfA7KHa`3k<{A~~ZpW8%Cr4*=W
aT3LC0Chfn&e`SGTj=NoURepc?+<yZE=v_Vl

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide3.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide3.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..6cd6e10f7021d538cd38ccd3a5603db5ee0c0166
GIT binary patch
literal 21103
zcmeHvXHb+~)8-(e0t!kHktB}-Dgu%vBdDlEMM1Ji1_6<r21J601VM5V5s)ld;(!Q9
z&N&R3A*X?XVTRdr5Bh%ZSGBdZe|D>O>;3VR56m6T>Aw2v?yJubsHUQDhT<Xx3<f*%
z;QpP*Fc=y5-(y{;j)8w<*&dRBe@N^fE6Bm}I+z#1568`JE8m8}3PLGyhR~(tukLHv
z!(deP(0?RNw%?3lu*b(A+`0Y4MSpdi(&gz+A_{-`5<|XVhEsZmRO0r?BHCeMr2DoF
zc{J>tl#zr~NZRdiS-C4?be{@FW)yi+<%}NPk(K53IY#m2s(G&YG7DcC-NVc>D(WGM
zf1W-Jl03GB7F(AYU!J`=%Jjf#SKB?mb75af+W1^}_27hSl82rZe$Wex1ZxXBcI*NS
z=KbVsE%@Q959|aCc8UJ@P4GkJC2z2^`(z}i!7i_#dJcZ|_%@6b+(Pl6TmECQ|3iSG
zs88~&EL|J=fsJi&mUe2Tr@Ss>IxjQ!eL8UTC)c`%y8=3t1j1*a<K*Nag8}^YguHrF
z8*YKJX3%kn8XNp0t!j5by2Kv)K}&qY$n;lL&6K^R*=c5e&*ysqmot(@x8GQ?BGJ$9
zqfm;ZnO6YD??<gx8YB{4$-kR;zw90yBldZ%6h0FbZBT=)I^UVb-N9J?ePvqbq=pDO
zH&z2aQ?2Qj7gEC*#tFi}`_69%+f#jpGAK=n3{~z_Oe*KEjfU-m{xQktzd!Y;{fvj>
zb<wMD=pb|bS{XLTi@$01%0zyqygho%Q$?e&wEF|XOwEWK`Q&`~LEXZ-_^Vf5Xmjsr
z4zRFy3g#W|$=c7xyKIMoJ(n3GO~@=KOjioulnA!^D_yI0TrJ_H@t@`dm;Nrpgfdet
z>g?caN{Y1IJ#F*md*~PXCPwnN4m^SRCe3fsaf3*f)vw;$iV*H?u`-MDn{UhN8Z!Ht
zWxqU+&bLHWn&NxZ7P~K>c<%>j;F9WY^+87&!;?X<74S~2x!7;yCi<dQwcp}9ku?YX
zs1(w>5Fdn5>o~v6u|!F1l^$x9#13T|Jki+4OD~IzF-ehp-XI6WF)~(|+*V<3xfqic
zD7|q|S3bLDpOvw5P}jn(OMkindP<kSocN(v(6aydkTXu|ZQ)korTM~x4kkM*!Xv?&
zc?a{W5K(!TZjCS@a%`?<$1{!cguE8M%n<cd*I*ia>;Qc%Vz;hN$(^?@Y`N*36LX`B
z_#zwB_S;C!XDk;fqM?0ZqVUEV{lzupRo#!Qpm$`)3s1&L+_QLoRSkLub8Fwa@6!cS
zGGn{8hBB(YTD8v*-pZeh?Zh*lIZgHu+OwdHo88@=Vt9idB9{Ej*!R~j(8X_#u|u8+
z&Xo|+7b-!AR|YLrAVcF?*Cg`i(+h={qwlG*UOi5H#ZZy$VA!>^)M&cqxXvblnRP}X
z2g}8D0~v<ZW0y9e*X2U+#>Nve0T#6apP>bvhHJC@YuFD{SxBzLgMEB1G0l&a^e|1%
zxuw+H$9LhT-CCHr0<0NAgLx!a16Azf*TbH}CNW8w<*v2WsyrW@PKG^)ULWO7_&xQp
zm!Ocvi^osuR!k?nUsiOgslGgNP4+S5FYUEm1)H2ittfimYYC4iPl}(r7_JIAN$z06
zGh-SXzKJ>Al2zxR0?f!)gEKC+^1ed@0AAd+A{=35t32M)$tIk7p8A5sHq$zr81pL2
zC0A%uJO!SZY$UF1jf%i3^|Ti&y@%85raBg`06m)~B05B6D%eW~-pIqp<|OZR@(p_d
zbXR++88XjS56RtLCXM~ZXxCcXBr0`bmGhF}@td2Kh%NN2D*73Tw~Wpf%0ZI=M}5$i
zoNV!{|1y63$_E%HoShG&za9F36`)@M+BmQx2yU~IWODmq&xJFIBcoDp*ki$?Xlr3=
z&yw(S6;q;>m~{n}@v%Zb8SL%iBJV5bYJY+K%Np(*lnwUr(bTeuMd#hhFx7Vbw+}x|
z$h>x04hc5YkhIf@VOrZC<UapIq}|b3FMRlLx>eP`a~QKGF}2N|#~1s3X)I?+P@!|f
z0$;Wc7anPleR1mfG?zEYB?9on4B!E5$7+3zt2O3a*{8wwi)w&eVz75jjH0Ms#&s)x
zS!L*Cyda%Bjn!vRm0AC?9vz}5a+7<%ty^6*1Hp7m{uueu(`*5g-U5VYnI5Srg3hci
zF+h#xygv9vzNq7FmHghHl|kR#Y3D=FN6JJ>JUgw^3W#?F36}e(NtAHF^s~3`Bf%y#
z#=9y342gvGz2tc#Qt)cu`rTf)Gf6#qC(PPSFo~LWp3MgeR0{5F_fH(%TAZnbrNYD4
zMOym!6;HNIQg^(IZ`C_-`oU%Iz+~77ZxJYvU-qJ<+%5|4H8!3um31(^*D0|M7&1to
z{1W4lKf2Bj<QTYh_jwqkG70?nlA^G~lNLQ5jn}k@uNcLS?+@{G2a#NP3%m?){d&2!
z4o@h3x$>RqAOw)AM&z8K9G0pwP009_N}dq!7-O|@?QB|VYV`5HtDNoH9nT+7;?4(_
z(FmP$y~lfl$wRN~QBZt!i$ixC#Qgxm-Py<f^^UnFfN1d&A@*T08&}tczO{NYOvgx4
zI5~-~V6LQ_p||nbBdn-M?gX(>S)*7gxU+w~NFdMd$gRrMT+Q_?EX^<kCj3^BVt^ij
zHD5Zi$G~WX(M(rnT3{Bk&G$r}IqH$KPFsd?lK_K9z7=}wU)O1j0P1a7$1qmDe6Hcm
zVGKDVu}p!cHcaSP%pM3{^I)k1r+y`F)xFp1z+)Yg%N4=;;d#O!l&%CH8A^Kve>tG3
z1Di}{=d(FvLMl(1u{3f2*oE+WFj^T9COl4nMXhYE#k5+<f74W1k1Q#82BnDrrE9_Z
zT>tWoz#lR=&g-%B2#TerLg@R=?Qh0Ukkj1%^Q2zDjN?_ubhIqndfvRvaG>T$D5>7#
zJpO96*#1j2#a}-Z*AB2S@x^IBe#fCDRz3L6)w4+CJo1B{$i`B`<u96)|8?C^Eb?lX
zeP9z(5cSillZwzjqHk=!iGy#m*}T)|`PbTXp}f0O?4xzC^j6TrX?;ZJz}WgmA){EC
z$GE?HD#Pi&PZSyq#^^h4;*L>GHtPJNyqHu{@H-to;F@2(*@_|~4ZnS~k@l^_kM;C}
z(RA2rzu3LyPncQXNGH$le)IMqb@0EQ)e0C~6gPiN2e<!L9R%2j*j@MY=pS=)MrQa<
z@H86vKTi{|?k`i|y|D*l1#WEU`yxeIVWU5s7&=Z<{roQ#Ob%6h&Ayt$94I|=w9&Am
zpFW+`gAcFlY#w>WtnS?U8BqMae8#W~M{d96#GY+wM%SEHR8qEk4V6|UZONG<F1qX-
z_+!4B%O+?-v^M#O6efqH4q6=<klI&Bbe(k&U#b`%f5YY=Ii@d?oYp!DyH_h3GDSbk
z2SP5ExX1!}$O*oCQ~sixhdcZ!+K1%9KSz)?f=a3P^TksItta$_ay17U<&y1MTM@2$
zW$fw?kutAuxwU)$Itqu{g39?0Y`5YG<GY$~I-t0cOdC4ylNUQ)>7<ejNX{0!W8ZCc
z;)?$fL_@}bqc4<INpe?EIP>o559aWYNr5!!FjZm}Ko|h&jxgnq68@J(oAN4BkZIiS
z7Y1$vl<}o!s0&tNHioO~*+4SMN9-qm0dD%kYRr{T?&BYq!2E31<E!LzH+rMeJB4fu
zkVhchk(KLr_?W^@|FtS%c{h$<F=$rgKCMQg#9oDs789j&92LP70tqjv|JS3`0QR>z
zkqzj>Dl5k)30Q%zBE@W=@mS)9?g`b!35Pw=w%_G*P)$_*QdZ#9^DS%4+PJFpzOlS~
z{cLl@$uk!n2DHo)52INS*jV0bX|JBAcxl@Cx+$+RSDa=cjnKroTG*8)z0giZ{(M_?
z6MFOqeSy!f-T22&EJ7s=;0pvw1{WF_w16q1IuAv<s=n)7<Z?dZP|*PUxG981Rcfb?
zfbLdV>pS$F8TFi>omuXXUhS^nGbu5@aoL;Xeh5gbE}IGeh>^L?{wz%GWaJ&oOzh2Y
zyJs=b2)Vf0<qo=#1Q3>m&Pi(41R;wai;P64pX6K)s4jNPF{AS@ezb0`%8?@kHb3NC
zZC?N_W4IAWhzy<A!$&U(Pw<s(gqj9)wTthI3I0i+f1W3=7<EHz?vr|j!y?VxOGU?}
zSGo&St9hvHL0f%KQ;oQJGmwm$3)aV-ujj2lW#;BG5Xv1yagY{@oFDLAV<9rO%&gX_
z=l3s>D8;*MHFK%T;L)@yx7Z&J>7x2l#+wsIkPe3nMN2}|&`v*rq;}p-)W3O5=-X#S
zY014H8<ZA^*PPW@axaO!ku?a3Agf?%wiDNj55KJ|ERLn<4KH%pBDUBHFP12Sf+g%K
z@TSBICqmBH^&gC42!kD-;lcLv1E-Vw4`vuPD<<9b{!jxii=@j2Zo6}Rtla)dwsoq}
z!A#9Yvc#NynmNg7*)McWT3vqdL;F*E?=#1_K+IYlbYJ#nw>W-NtjHa+#^*6b$s*zO
zLA)f`QLIL9V&L6*fP{9CsLjS5GQOvGb$YRue}B30X${zv%gTKKSm7dxlFiCc!Ax<l
zO3;XqfgEXm7H%}>nRO5ZqV|e2n1S}|EA<(E?uXm_1GX1Ppsd{d=Iu#p_!OnVyDT9D
z0Esm9f*d9ApecxmC&Fl!vZ_n#0ewHj`#!B1MV+4={}5UY&pmdA{0g{pYip}2aUXXZ
zY-+DmpH>lAN7ib9kHt>$wutd+f;1b#>4^ZBgM0hR1igA_CcOU0Q&xKL{kcSO$M-LU
zs8~kXG_gpPmWKmlA+QsdVjJ5lCgs;<<7b<7@l1#DB~nwgmb+h=+%!6`;`g5Bha@Sk
zNt^vlsOwfF1D;BM%IZ4%u~dJ1|Cz&D*V>4^bo|p?F=k+10qgOC%`cf8ekGrt%(Yir
zE_mz(+tW%w#VLRdh?dro!%f=5ergT<bJb6v(A1to_o{|o6tYQqV-V|dPvghDB|u*w
zJx#Wae%+0eQl^Wwb?Oxi8U=$BBV3a61+8(OrTgFZ=dzQvn5$KYPgrYYvX$JQjS$2$
zR?>#pidHRW6#f*}grJw>xn3h^{Ltcdj=!*BfBD<|bq!~o?U2;QTM)Z-pONCuz=-em
zCWR(-+jqEc&T*v<l0&(Q5`gHY*iL+RCf-i9-%Bi8&tBlz1@9}P=;ehh8<g`0ZovlO
z=y7NLcg<PxiYE*Pa{;eLahOOiOz-1)t9Dn;ZDUQXWp)|)@d=5x>+kfhK)I3!y=v=q
z@QYUULysV%c(Ftt&Q`VAaxmcoaKRM3YV&Z34Xyn8sm8<@xBU$A>{7qrL!DO-tS3zS
zrb+14D?u1)T7_hbvPvcONVvH|WM8b-1IW<<M6k}-`DXwK9?nwI8T0cb%~e%SdpQ%4
zqWFLV)?iSs(wU;#uy@HTyj|OR8Wn_H9%2UabERGz4ugbDUo8TqLG5Wj2K`%$Un9Pt
zl)`BxN-Gy}on~`Z%OL#J&ZZ-s9anh*b<Spkoq_<~PcfNr$tg)uIMo5%8YDVif0>+G
zKr&oc?~A16nro=7+f7iOK}-DQ4C0!KMf^^8?9J*!P7LNa2q`*W&_UNY?Q_)I@+$r=
z?@F-<=v>m<#Iv#yTpDk%JOgEk-(V8UW3f(6fq;`DuMM?O6C0Q8-EUNYeh4^~w)*DL
zl5G#?ZNH2uC8@BCRJ-oPYLSh$fUu(rt7J{K@9yh4KUT_BM*H{=mZciyE2p;Qd8ljh
zbneQBR*a4~Dd#I;L^Lq7c~j^7k9A^wHB|O|qTBa1&~7l0c`Zr-8C<d6&rVGoDaRMS
z=QE~%$qm`#87TFEQjib7ZN+D7y(u)Yj;T05v3V6Ysze?DhY~IP(eR%82@_8hO3pj$
zfITmEHYJ^RwB2j5c9_`tJ-i?Zbz`8M4uqu-N8QlsNmGozDoVT96OL%PKBNJn^7G`1
zJ0`JpL8GtzV%S1&iu_~$->3ISu<~7?p$hP|(G_reh`AXlx<g?g_LAHW7uXaaFBqW8
zRQ~-2E}~_z`uasTU6<A2*!;7lW}dHYxIC6Yoj}bz##T)u8-D6}Ty9DVJ4O3f@pZ)v
z(?7d<2Q^iP^=Clo)z{wI^t6uMQ$EojeWorp!l+b1_3Fggys2d~swsU3`A(n!?HF;=
z+3poX)!utdc;&XUg&p39q<6hi-0h6F@3~qbou^Nq(#1X4>F;!C{<Eo9tHWj7=x6b!
z8{LbeaqbJ}>wNVd4R<+%%+)<{lI9+)_az>?6C@{?DeWj1kmF-L2ND~_csFB0{Z5+P
zBSmCt2E5KY`+WZ$&^WP@C>15bv5Sw)IiG3NmA(ed(ZC*2*MG*OK0UI(eNo2k0i|X$
zMJ^<kTuzTBw9YnCTBX^V@>Z@_r6AX<L4Vx6b^<iSy)QEoUO#*i`@qq1VK=_Xi-Y@x
zjvxs5e`x=)-VqGqOV)X0=lKAiyATpH=nQIWHpkqT^VrcgjvqVrC;S?jB$mB+4>nf@
zC-&yzb`%xbK3L~2epLp>802GPTYYiq<MEsAyThGu5n%A8uC*Uy6BHm?)pDqq#|SHY
zgSjr#rT&q$`Onl`hZAf~&VRgU*DJ<=#K{PFv5GGHY-l3&1Oz-rX5v}LKvUZ-bGR+8
zQ*VP}8@a_zrVde=(8@_X%b}kyEwFhgtJIe+r7U&vlFlQWU$Ta}+zUE3z4^&ZBV7lD
zrR6C$B<%o$V>^2Z<{7XD902HbZ%-<u#6B_D>r5=9D2)W{IccPXSYN6T!x#D7e5#K4
zwh}%}fC}?&R+8|Cz+@4e%I|@nT2y4jvY+dSfu2yh5tQiJOf8MPii?3=DscZ}PzZ4-
zMma3?U9)VzlO?{a{wG2=HU+rk0_fV_*MCVq(rg5{WaLxM)(VYM@>PY+`h8Yi8Cpx<
zmPvm4Jr|@FSTd@k$%YZGjwseU?Xj8gW{?+gWrR$2njG-fY=mmaYe83)TQzz%>u>1p
z!ANTh%%6!B%tt|L&M$Su^I8<6fea+};okhBtu4M2K3LbXPw@<B8ZhP{h>ARH8h(#H
zF1R!RIyAnx9{0`~z3zyU6mMYZ9IP&rZ;X@3#xv{fG<Ob^G-^uW&~2$uZkcYpoodV1
z+0yV&JSaoC5lAx#$h{{he8G)Wtdc|J_6t|VBW6FTtg_`EJpOxcCFmOIok`7>oPke*
zl5VQq0@=Y~t$e`=6n9G<zxJjx#LY^qRKQMTi|Fh~vkk7AAB#CZQsrvjW8?4eFk`p%
zdQDh=o){yrOgsHWm%THj_r^vLF#~?AC9fud*tNs}^tbf6rR8N>$Fo0zsR;gn7Q2p}
zq*_D3_xF}8@Vw;*2|3|6en9HT$D&R~!I@-dAWFHp$&%327LTosS4P-#yo#O~m3m5;
z%X!t_58}|g;AAo)8)%$MKpo{UA!!H33Q*>*1kxxZJV|kXz6so6()0DfgInh+JS%5f
z2)mzhEZe7`dxN3q0BR;*S8vKP!|87}SKW7`m%nh4(g~0#Wix}S?So*)JNzA#x<naL
z@{n7rKdv5H^Fy#^0?;lgmK{)_uV>-6uSIk)Etw+JgpS|LEdfQFsu#BRQDEcLuaMP|
z=>Y|WP@s4uJS&@;0G)egWv;;%HQ6BFbtdr=G=PxZh#(y;i4xVzi7K(Ypf{DwNFvJm
zv&GneEz1kHS6K$f__GSx%QeYDwZy+WVZ9a241PE^HKJ~lHpsyun5ZQ;YRS10&OWuM
zNf^uY93RdwX|U7*N!)vE1k?whmii17NT~bTG5a8B<GTMYmehZ|i?FFoeUF;%D(|tF
zgjiAAi9oKT3ndn#5}8JoSw+&3PqVCrKnRV<q1w)gQe_?9Mt-DhotC=h29ke?;0X=U
zPk<#v79{1*(bmrPfEanTl$kMasGLJ9F+bC`kPey3Nd%8AU*}~MbFf^ma5ag}`@Kw~
z73_@Bd!_*>7X2YV;9_)*8XDBPrNC^6gGz)|rJ~PhnZz3CQ9SyybD)O|?5>PDE{|A&
zIJUH5Rs&6|2+x!fW&r6nF=^$Fdp<{>x8sLCTRS@eC<AmlqnYX%32YQV4hI#bWv>LQ
zC(0Kl9VNE?laY+Sd-*slknXjR`cDR=r1|XF7Le5dUwwM|ohRmOb!znK3$CNJ6*?zO
zDlkBcvj5G{12TYy)5?$?2!{D3wGGfqcx|dDvN0bzYgxo4liS6Y%*k^V&xsK#q^AH$
zU8Ng!MqnqmorQA}o(bkwE&0PQo1L+28yv+3gOLPnD0P4-RWejSj8`>Yt}YRcq>g=z
z9p|{LDQOo!Qype{ET(h^@CSS++dvA`KH`o`bR{qCi*21jJ2pz_f6&BNwLxgkYWDys
z*&F23GHzch)RfwH%^TkmZklb7A(yfIf==_IKQkbKEOvT)e>Vh0NwUc4-5k&!FDfb|
z<gpoKaGH_%#jUK-(1E!<CoadAgyD!#w%+ahY4Ig@GzbnD(A#6GMylO0;T($hK3uQ^
z?WrGY$LyLx;!uiw<;+jE!{tn3v${aQsAnV{lruCPdTQu5-U5(;z4r!f{9q&@S~Oo;
zkeicM_+P(liize2>xOy%LlnL`pe_;iSP{}i5;}piVPB&l1>jaPuq5n04<shOk_Pn+
zEKD0}$tR&r!!DH&b&7q6g=H;?vc@*V_TGIokodM>57-cv$pOg}|9QcGO!R*pMfB=0
zhF^CMSOBttq>A`LKIs#)q5B0-1%q89>fiqxsC>`qT{IB)aOJ-q^MYbB$DNAf)qlYK
zpZNa|MgGr89^rFp(kQko3gmhCNtpL1pw9ArHPU#+Y5nu^2UczvOzr!T#0`^WJ({?|
zyd^>9Fbb;6WRI<8e}Dhx2rlZ2La3ag4skr61bjzO&&~7jYc|%Cv+Zp?!Df3e-<ZR^
z3QA9(PXQs<7!(|Wrfs)oJ_w3Ct+;>^=rAMH>2<~~7pUU|x!An!CL`ckhpk#X7hC`4
zRS|{?Oi?oa<w8fI1jz2fYXug_wI31Ur3ZcFEOO*TmGO%ZF$F^J80jJf6<&F#=bA6A
z%DeZaY2wc2&fpE|op@c1=y_|15a>vN@;OxHx(!rh$DRjVcgf=^+25*46x2f+LB9nn
zDBNMP=b$1fSaf2jz^Lh}jy;d@2Et>G4_Sq6%eiOW3#DJR<DeE3@z~n{OH(Kr5fOJZ
z#}E#mMtkVf-}tlT*QUrC&$P5OYv&b=`RXW(^GG0zKJx)aD{bW7U$eZv+_%0UL)hP4
z9aZp)V*mw4bn?L!XtoU16=PR<cvibo<mu8~9iW>bnJs_b<L^GG`UVHfb{r9GX`o=Q
zfbaSqX^O<<7Z3Q5Q9>orjyZD$Gcs>&rMEC#xa{wHwN19?+cyVn8v~olDtDL&JlWdz
ze=CyorzcKdI4f*Erm{L#v#R0?L?7aIc`&u4)V~}T*H0$)l;cYlJVdy4?f+IR>2oST
zO8zDtF*U2QT4u0Mt+g5#@z+p1XmF>)Vh_i|PGb~Mh&?f?e|n^&wRT0<pwqrdL|hSi
zrK3y!*M8bB0AwCW;NtB-YY%lzk@>k$r4PF+3+tr;>t6=$p-Ute-1~rNYZ)8>%>%Jq
zvIu0Aa=TL=E1BV7nf&kfyaaN*Qpb3xG^(%TsHr<r9*CeQlG>;x<w|mS@hQH!w`SaK
z@O0benKxvAAXC&H!=EZ2K=sImvj^e9(G*PrG+&~ROe=9)t*M}GXHAyj|I<xFbzS;!
zhYSICRVZS~K7;vhD@<JaOMpBLAg=U-it<lSO|);j(}+dJUC_xViWiT<=VXlUzO?#_
zMxoJb83{CgrjHYc3dXr<JAc0)8UTIE(SA$$y!KmxlJYdbsmbI>5F*-DXw;j86jBs=
zRaV0OG~-7r*xA_u*Z0fdqG|s~2E-u!t1Mv~k%YT{WSJx5<UT8yhU8Ce!H>M__`(0S
z>{6qq5MI5CJGrrvUc+hltQJK3k$myT8@;MBQZ4PXU^FluubCYC?+c*;O2iTX_>%;3
z!AVN4KbqwcD1Jd|LR+i1>f+cTq$d2kC*>f~hc59!59sWPqK|=&%%9RZ-LD3}YU{i<
zCbraJi~@?y7D9^SNChU=+v$Fwxras{3X8F)A@0e<^8Gh3j)Ay$0n?g0JB(#&2*Oq*
zqK#ruH~6oy%YpoV*kFLWOZVAs8Egr1#=n0ad6L!Tvtl&wqqmJ2V9c{r=_UT}u~p)K
znvrJR@)^VjjRwxH+e?B)S;7JTip%B6y0>NtcJ~6PSiZ2fSAve(Hb{q!Me1*3aMEdg
z0>7y0CBD>+6E85TJQ+aH`*%*TsoMBSxi53B!al%>$a^`Z4S(r0f&5HDCs~1biOi8a
zbJVvIn~-SGxNVAhZHq!<hJQ-hL3>PeWJ_Nu)1Q!wFgn;>T^=p79ZjTy6csPWo=v_G
z{K3C7qV)@)Af&vpPpX3~izC37cDV`c+-jq4e>qEM7_>@Z?DDT2#T|e|;dn3>2+8!^
z%G*G-J93b<^+Z*<X1+cbPOGo5y&JrW%jqSiLDjKifgZsI^)e8MBd8^uR$};#?v>e$
z1D<n$xZLre8*m+Oo0n9{3^W%Ue1uGedLkm9_k+|GeM3v&#jkfDWi1C8;9hV#6PA8a
zv5c@V9G-oT$xdE;eds0QxKy4z3aJ?M7ia!jZck71&|m_h|9PdpfBn+0s*R%K>tAF*
z$2DNEE!m^dHEj0H(S#QEoAl_U2Q2L5(XbSDiTZz*TcMLPFqrGnIfwu6AA|^tZcauO
zfTO`=U<~&E;$hh^u2Cn}cZbLApd*X2(m8w0ObfVWY`^OIdujhFa7YR!8?S@dNmN4^
zCJmu0&1q*7A%Khw*kY=j>%KC`<9XfocFbvi5f2MnqsB2IPgI95EIDsYYkpBaOAFz=
zII>>h(4k>0HwCVsiex@yw)TsG;P*Lq>@~7f4Z)z~wwYTU04^sc#^ONmc8VSM!>65u
z7o%+8*U(u;SUGxY`G9AK?xo8dCIwzEi5^=6uT_Quqob+>qt-BcCI$N#UN)L;lDv>N
z%xdj(9NJSDQ$Q%r0*!7@!REzO3b#{A3-cU!?D#}~Q3i8=5Uj6_z*6OQ{B|kHE?7TI
zYb)%$Y_sRuf*}RNP7>IqlWV$|cr{OGjb-i=K_Y2AW5T0?2X%#=SHd7R9nglX<7p^T
zwRK>weq9PI>pSg@1Em*G`C}ewEkW%&OF0$2<(DsqHr;(|lKkKXv>To>XTnsQk80Ui
z8PCY4hqkLypR{(x<__Kf4IKuK-a(7Tweh52ZyZQY`zMc5Ux5PC`+Qu*N{O0hlxK}w
zo#HMj8Dw=?MaXL@*bc;!YpZks{`tNY1$4CA+tf;XR}9Av9v^FxT>sT$fBE#u+6X8>
zWw&uoo$j%7A)2_^V6f_H92VYcz_z+X#;3>coPAnF09s@F*)7hZmF+&0{JZgw7lDVa
zbK0JQcFq{9URN(o?#?-KXA5%VEe){T#rNep@(Ul5%Q@OxSppWM-LUiD*=`&pd=#U{
z;j)vw_IL2zoHP%>+iMFd=4z}{z@gBWW0izg#n&P^`QRKyYzb>jIN)%B3?G?-Mdxl#
zpz?!@`L*!qytYR*z>;_3A6s<S=Z{h=PmH*ud>-)mv(Pvx76%@OY_`!1x!|mr;p3W>
zZQ6LanT(Jtrus7E%ps--1_X`uKf@4zd-x#NT>0!jbryEU+Iy<ku*K<CPo{krl_Qm)
zuTq1yvv*6mRXPKA)u6Yq^P0h7#RJpf1xn6p^Wg<BPauI!UIqw~bi<lL#tz$>%a50=
zFjO64S}@qb@s;A)NA9pgz}xsZHYRMeR5<t({vGFduQzj&TB=N9<jx!7hf~mp4XP8r
zS5CG{pH$B9>C`Kd8o4t?{E!{`5I^AjQ{|M=@WSWCCmY!eRdm*0K0~YfJwKnv2!i%E
z&cIt@5YIXC$p+fpuD?k#C3NJ~n^4=XC%C&N^KySVl?MdpchMUMx?RxLaVKNrE46oB
z0kC%EBe+EHyeq*sqda-dp*T3?Yn3)Lj`A#J6NroQ)CWd4r}cx}CHvK4c;OS<9E7I1
zRs%C35I?P+qhFNih)50wi%o^D;PXoCcDTMrZ(k-(Uj-T&80?k(;BUYaW?&&EDrEUd
zh24dn#Ph#lP5}`P_Jf*bgf%mSTWTb2&3*J-+-0F8ljP7SfEVD}LjcfjPKpGuqf4t>
z_V$BHae3NhCKD|TKXnBUx`PnTwmXChv)_^2Etd1iu2UaG;=C>vUGWGyH<&|#<lWA#
z&+(j2YnvJNx}-g`wDC~}0NZMaI?A)dT%6vco*NM=@@Nk-N@mn(e5DPt+!>$F68oFk
z8BdyG+p56ZcD*Ouu*^AIn=XA}!z9fyTcFn&JyK{FYMhSS$`D-aAG?;kr_{^PthA0j
zh_AUjA0KtvKv(?b?E!fT>MbTs>5DZ#W%Nqm*hD_b&0g!2VCU6+BXmxW*=PqTF!@7Z
za&L6Y*jPzJtF*1_YJ1oF6h}V9R9zkIy5&-H($j4@OQ(G8glnB}BlrdFVnSM|iGZ+}
z;Pam|Q*ZnN1~U%dH1g0U^JMCrWA*5zCEr#xiZZ5R$=5F9c(R#-2_ifmsy?7LQz0u`
zvEL*rtxqVUJgFKgaVg4!ZE!S>UonoRz-$huv}_tOZ8}oel>u$bnmI``)KF2;pU=jv
zhU2Ndzn*93r~S}?>r@YLyEUZU6W1s}Yy=R1hxZ^4_n<CBdrp)rgoa8UEVa!o%Ws)%
z<u6Za1#{61=AO2TseM~vEl$ekUbtDj^f9}>*H+T6Qb1foS3}V2mvFWmwG1$gAAH&Y
zM4CH>Ekxwe9imHmi&wS-CdPJFJx|J^^?|$7OETS_@SvXBYXWQcYE&iPip>sDTwYt{
ze`yigV(uLIq&*YAmKEDN7%=|Bb78`~?7R<PawsHA6Yn}cizgo6@7wQ+ugpUhUltnm
zi{@*6qKj$qOIem1_iMCTlG@(v?W>mx;9J&W;VF5lAxMxAC7AbD@bFZMsZT&jC)TV0
zzAVwr`JO0B>xI;mFCUTy3`lSG#4N9^Yw&0Dy9IPCgC!!nt#qoWikjnOFdJEGPlGC{
zY{u$=F*${Dsy%o7dGf0nt67`RPMn_NImi0S#r;nF)q4;e=qxJAmwyWL>51W|Y>-%c
zj0Vq|)N4n=rc4Je(jxU*rLmsHc=dR8bb=ng#LBJ>ebGI_D87|#gzmtiXT8Sx%3rd+
z`WSj#@lfe@t&HBjE-HZ}b=_)ZqJ=EQ{iOb`v3g2ul`=t*ZCTID>)n(WZu{fZ?VX_X
zHZNRYa^tJflt0;eiXCL6Gmt+T*1HZApp6dlmRn8wa$!TXbRL)WY);H7AquE5S=ACn
z{u*WPwP-I6fEeYl&3=FciGAY&V1F8DB`<aXcD6NYx@`BX^-gv!iFs%Zb1a9b@I`!E
zlsRmwR>)h^rysZ9Tb@}vm9G~cwXu}p?6+<;ya2{LK<vm^mB6UPmFJC?nL08FWE9&(
zswYyKw{S`K+O|=<QyD~+YwyL=M)P=?_l_RC9{UCmDpPQ+yH!0xp7{3drCp?+xTE!{
zj5YZM3}Z>7=LZVspH_n)o>}XoOwWJJ1JHT8T)HCj>Z)6sP6{BiyVM9)>wn0Io1XzF
z<ZRa+jH!vF#0Gorp*{(aNzFyejGOhvt@SJZ3}P~ss~=HF@)UOKkIx4;HwPI6ihQIZ
zzWY0&HU*&4`CuTYh<o1mv3NzI62J1*sfis7hS|ye_vB&8wyi;R3R%WI1CWL`((b7u
zUk)!6Lm^97HAiiIJu|Mi<b@M(yo!rT^XZrkmj~l5ZFhfas(9@I?!4?K>JeUURT|)>
z&+l5V<8J@zy0G6}z}jlH6BEv#cPq2yj_qVsgZ4fnl<G=^KSWx|kF4A@cqrE2<cC(y
z(B_|oqsqsp3=+NixA_n?G9!25pvb1^$0|!a%$j##6GyfOD*+I$0ce-87r8&}h%MXl
zqi&xxczGayAGy&S<*C90>IK;;G4$a_i=bq2FnR-lWIs75q&!3a^RlKhn*f*bE0+Lf
zPuC3{cWSL&XlyL|W*G2>0h4pnFntk^lZmtg@MuKyB!IO5BW}%I;ojKB_J$^SxEORb
zJNb^pH~1q=3mGY=B-~;9LWJ<66V!==t2#xNdF~=F$<Hu9yyeFAkw~mQo51D5`e>c<
zT8qy4p^VON!ctibyKI=$MQ8Y4N!`}~kXAn5#I0!BetOces8xLjA2+YE|I6&*@|slT
z3n#619KgdJ+iN{g<D~|Z!#ici^glz`Y+Gp)=}dEVJeG+w7f^|t?MEWFd>!?gR3nV+
zOJ&(=BGPwyy>OxK8hqr8x(DM=N1fM#Cyc+og>u;<V1ZQ;<Br<aq_*pfM%t42M(MFs
z%iCqP?G=C=P@eZFAoJ5$@5{$`RK2=4{6nGeSYyn9DnI~;or`XB6ahZ!a$s3ux;pyq
zMIb052Xb4M&KtcJ4mP`^$@B1M3OGy0wm+Od;W9g{xsdBI6*nA8Rf$b1jN3QUzcmvd
zjTD}3p_k9scHP9Ro!L%|_bLWiA>W3ehwe0|f{gt-IKD7_v}<3h;e}EMSEGDg<t~n&
zZhf$E{G7I^pcMB4$cs-NmlYP<JRNYz*QYsLT6sS_)%)=Jf|B_&rV7tiE8BYpL8?%-
zbpIR3^1QexB#6}_?S`1r7?j$ph8$wJP?n~ZVOXA(*tFh7+UQ1BST(HZ3p?*H9c*}$
zhnh)cg~lC5OAoluE_Kz%URW~_l=<|%lG*DYXoYgZd7g6{`OWLTXU|B9+0IdRSAIWJ
zRe^V&NzWli&anZO0jXh`z@YkWO25Q8fQC0SPSAcI8#d7GJWUG4^@G`s{6H(WK_t>~
zBZoUf#shpSvwM_ZVVCVz-*;5=M&p`(`F8-E^6Fg8{W(?9q{~vqDuCqF=VfZFI@wa!
zTOJV~L${Tx(C?FDuo>Lk`6<|a7FYtA@Z&}RW4YycRCtTnxPKe@V*JkP>s#zuOsc`G
z;(W%h_CmTb^8|?8t~>%*dmJPazh_$O+vE;%oPi)>XPE~o2~srkZh=s$wp`~$;1+Xi
z`|hbfE)(!*B0poLcfmslGQ)IL^}6?VvhoV;ixT6#@SZHOT#HFpZiC+m+hc^DrotMy
zjw(P<#hF;aZ}5*MU*d__S_n+Gp==K;GcsxP==-GV!P#5OQz_NsxDBV7Xs%vf@5C0m
z#&5{?6}dfdIbDF9yTuOeG*A$jc|4$_75j7|<{78Zt4_l*>MVo4oaHyv;P84+rd1_X
zQlu{}qg$vh3v2V&*yS%LqhCNQ_^aG>Hx=Y9V?2$mlCC;@zjvIQf!UpOUFMZJDw4ZG
zqm38jwt?vPR)Ak`32hWNsNSXK&n3Q;_Sg?le`z7=m}qB=FG49|@0a^!zxy>CTP9#J
z%OHPf8C;oIlyJ+nPl1Haz+RW*KP>GA^4A-GNkFB?!CiZq!#mSqO@M0@?$43XvUXw@
z9>f3!4mhEI@Jq3~UK#2;Wv7*52`-^aqeXCMQO?~((fI&2wlMb4N{u!ZQ@m+9z&eOg
zZG?pa`kgwYXolSaDNRj#=G6BG(`Z1k?hPx{dUWExcE4w4ku}+1)#KMqOAZek{q7U5
zF3X?jO6|Gc#TlqBy&0?DA9T1Mz7q6Co+cpcv}?l7escorQ%Xxj<#jid1u=3GlJQHy
zf_w=dXSeSw$8l8&0{fPLNbNq@>5M*vdycV4WuNzx?7+?%tAKp^Gj!;3XU#>aH+@84
zsr(KER0@2NxT90aPaX@R0Fxh!bU5qM)8=T#b<l&PR(a6Am6|N~5CTOw3OLZ6s`B!(
zrjrqk0A#quOB>mvUe5D=g4huK_#H9w<Tmw~au3gWE?$l!==V$mv}+*>nIaM_#7~WZ
zu(4riZ^w@M#7qFz&yFd90wuYE_<hJL-MfkIlrL`UMT|5v-%dm<HP|4p_x0|yf;^sT
zS?B9t(ygj@#C_Ls`mAt?PHg0O%&by=TxE*_FG|{2B_+pmrw%xJHI|{n#U1$um610D
zb=fr6?Jd&Y)^Zl~f{a|mTb1@%(~5F&0)#b0;xdj6Dnz!VIu?ILwKJ~r&~78w&wWo+
zkV9Cp*IxBGk?`nI*a}V#kZEl!#<OyQ$wXs|KC2hfMM0F^ofnHYuNyb)|7!TtUBJNz
zqU2JBG2;%O8i=tK^1cyQu?mVL;8F><r+nAZw93jKm1oFmkqKJNwyoiDX_1<_@8GlV
z<A^z+bbV|z_k~9ONM#Y+N|dv45mZEFS$Dy=4IFxk&KE<;2?n(<)3?U^=-TO;$ZBJo
z6FRErN<>qW%ZT3=y?Q#rG;bN+t;;|fe?_TV7d9p|wJ5o@lHE<Mvk84WkIo-|lIl#-
zj~s<L$ulosB)_PD-GZ1q&5#9Z+hpsyYB|}}wab=P*vm|;Z=kL**0{9bno^UCkDh-e
z<qcnzhZZK~;?OTK4oB`FQ?i^kbXK`ln<GX+#<T&_yVRZweGl|>qcRw%1_9UTlhjk;
zju2T3L7C;aKqjG3?W>iaM<dV$Imnsd=6WbqT)bsvbXygaAi${~;I;3emi6-`zZ0x@
zTZ}V{WcE6$L<`Zv)H~3dYx(HocuYm{+61Uhza-}!t?PO3c^O9l$%jiCrg#xsrTC6v
zhIYQVc$gWdWFjS>!FjwcJs-3|L!A;EMG&z|Vk;ppD!76l>^bDd=opiXW$&(ytmnOl
ze>-~jq&1*gK%7!p=h1A0nH+=VV2dCNNwSLhJb(i!P{=TWp0Hc>$&p##>bW02bM1eh
z_tq?OU_5>&zM5Yw8I*zVU(%z1n?Y|Sgtfb_57FMV-NDiszyUj1`m<`^Eh<6n)iz9D
zmLz4YKd=jo71sc+f>E@AxL~0Wu@>hLZgtlclBj9|=t6@U-yM_yUy{EZ?OQHZz78xA
zWSA?<44y{Xf}Zzsp)*dDYX=s4>rgd;9X$rVQX+jmsRKkU10F|5G=tqpjs5yND-Az`
zWl)dU-79Wx_7#0RSXY$J1GgMZlJPIo0Nhb1og$lo`Q4xP)c#9y!WZakrB_W=aSo$0
z2Co&&BLR0Xv~P;jw)$dMABp8}Wzh#r$k0q^gQn*FyvltJVP}1K{o7c!*lFUOaL|+V
zAxsZJVF!9N$I9ouK8s$9Mj#VX?Now-xGjDOlq61U;xfy<6m~yV@3>f6{D9qpXzQyS
zDhkxgqIu_}UfLn-WLhj?M$)D&yjl!2b^v*1A$ZdqWbsTJr5Rn_@Q%q0>w{*p+BZ;8
z-97YNsIkvCmz<>+_yO{^7RU3plNrdN0Kl5UDTb@vJ50XRt(Nl)by{5@rLGxB$hc`7
zQ3J32;7e)*ny~Ghw5Y*fsW4X+R;Th_giRRWzO(98qr&(lXAm2d_#3E=d1}|>@a<$$
z2x1k?(pwqby`ljIc-tS#{h;(+0&U<ym>B42qUz>Rj5nt?0nb-xw&?>bq6hZ9Ax4*z
zp%ky$uDADqDC`=&1oWgONRI>Es>j}{7bN`TA3i!^AXN6%#^iTW>UQl{@CA46UhCg0
zwz_Uj04}Mb?Fv`3W03&LIbQl#p?s1gzu=GYLK{Dpg~;w+sLc@Vu}Pq#)RdIjJwk@n
zf$Veh8@wA(=he*eB4C!#wc`;<j4s0wfClSFpM(7s@nZ36km~?*9>Wtjp}tImpIa)x
zi@4i|<3Fa6@}xlFn*(yn-S|Tf<Rt&#I2$(OpvyJQ5mVv#$_>&+6qhta241e>cH*7x
z*uGPx`Dfp1W(@_RFeg6>2lXEqcKgSUd!j^lKS50_U1O3bzX7Spy@UdgFwA%j3_$zc
zdD1s}mub5`wY828iDoUwf>Z<WW@Au2^^6L~%409{)yF;J6NtgmdIS*x)%tKxulX0m
zzN8o&z6RB=0zqenWJ*l|L@_rL=65%_;rr)uKLGOhzD|NFV^yZhMcRWB6s@D4DBLk^
zh1`{JM>8+0fvQNpJF1t2p)8kKie1!8ISI7SAxNxTBn5rda}b#|Vs=_HncaZjvdX#9
z_1$8I6gF}>yNo623MMnR57)<VsemdjR1fjc9J<~|0ctF&Q51kw7IzKsP51UJRCh)o
z1l(0f{d`gfbhCSGUd5o-?a-|`9D5XOW#Y&to)0<=GEd3`2dH7Vk*m``^QD|!D?b*G
z;>z3h_%NUAk)ozie?7C@UUn5%oQO++zLkEz5@<h~N)u;)PIfpV0!2K-f#x7)ayWyx
zVhUy3{gyD8J)Z}3C6vlf^5nxAuQB_*!mv!d)qH?{yEO@`8{zk4U~2_+V)TYzuK>3U
z0klbXLM2QfYv9(QuFw4KcD~!;+#p6FJ_Izg0QlGi#kEr$%;^i?B|u5NZu&DYZsRrN
z9I?@T(3|IiWQJ;JEde<xQNr?3Xcif0pc8j4&<MmTdPU0=UvvAlNSq#aCoaQjDN@{8
z8l90=6aUmQg3iMZs4&sUjpv~F4(8AmLjaS2z_)vLE03Ihr1J1Ed@88E_0eNtGl1c;
zs$EmGWl8LY$b~0}Z=fVTUb6;uSRi&l*__%G2la=re>B_1FNzyNGm20hOCbEQmkC5u
zhz<;XABE7HbUCq)^lgKXAs2?Fx)k${J~oj6x+L`PqWpV&O8QWn!4YZy&@#$Y>bncq
zN<`jUrhHeMHRYuD#A)_#(BtaeTx?LIBr3jr;!yY0r}=IwKQyT0<C^eGLr~Mm1z*8E
z?DmN-R2+!x%e}$q<UXU4S8mmXkal1F8O&3Y`-mG4?Ct>iPT-xKlZ=DtJlm?sL8vwk
z<!d1N!eF_qhSp~asOyB)B+M6Ou`Qsp0M^!d7uL4=>F1{>d?@|uub^<9iS41H67@Qv
z<RbZ1uhY=kfk{AOr4vXG5RgF8h>Zxg^+#9-F>DTjo;J1MP4@W)_-z-Lw|<HCAcu8d
zRuly+1tZjf*@wKyhaXc5K=8jd^2rKH1IH|YDBGw<kkwM2nN~7%(`au+zXH0DgfuP%
zVkV%O`szWwL}<{w>8{RD0H7E&=+Ey;U^;EeZFTl=n}TJ<<3W3Ype#CQ4{Ytoix~&n
zOltDo2_m0ZM)~XYIW5EKiBnU;?fdBv=fZ#3H6Rq?M}|fZsEpnAwwPF%pRnnPjk%<b
zWZh-o1t?6Om;YrFons~0CgGvgyx|us)r&BN4?O}@xWSLvBIlca#{_3R+%G_mWbrTq
z)I&;V6Emw<Qgj1_1W&OuXy>PL>7_10#`frfCSG7B74Y80p+3o6oplw+MPKz$Q&6N&
zMXWVHF^C}+1vjX!MT$Ly+OH6*r0JM2EXdT&%riVJrI)sURcf}eg4hQs4w62b;M)_?
z2FPUBbMfw0*$I@?m0vAYLR<EyhFS2&qvjy(4W?$<4Jzy4p6F`c=7OqGZ|ZYBqQ$?+
zBJ@41Y;ol0viPoSZUu<=ovPY%$0P{~h@Zch-gMXgAmmxNn8w(JMxY*F&nE1#vFY6*
zPdIC#oC}n@!iDdn?i%e@Zja^jb<a(E0sqN~n<kEq-B2zoeZP4J^#Lk!4#pd;?0^zX
zmEYZ6`A>l0Mj@A?WmLerz;ql@(WntsfK;lDDI49*b@uDi)gfr3m!Otv3+#=x&yc@B
z=#Ak}J`)>zczGSAhp154VTqNbqKy;$WM&#-lfkUgJGd^oJC(jZYBbc-1^2Z4LR%F-
zZ)Ftis{n6bQK&CE^)Y$-C8+3J?z>9F<7fQ1ZlxGny9`KOUkG5BMkNlkkXKefO_o3S
z;(-%9$@bvMGoqt@bXE7Oi}+W>RGPXMXPyO)y?-jV?c5B1ZxUrF!6zHe{Ez+KSNE1s
zaR*%8qFwjc*)|1L(8~f%!vXEfqm>*H3;bx}S(47^*59%*Uf-r$BmHD;giD_yl&8-f
z{c3aw1h|QcuRfUknX8ZKj~`Z^#Tj-wfXUV7iEbk{o6Wg8JfCNdX1m@vp^9}#l#7=#
z|A&2&OcLb3&^`wii`QG3T5)SHalg`6XanVaOvU#<ni;e+-?N#6h6S;F7dY>@psi^x
z*^9V?xh7E>=cvt-@=7S(%d}Ux8}Life;j^;e+;NhtrrLDYD7I=V^V^-f&=o;I@Qo7
zg?=QysR(2)cFFK4)uz0~7xTwFabo<go7{$6y49Bgok=M;aDdQMVhMYLP2(Q}r`ph&
zT76!#Ecm61d>b7V_bR?xc0h%rS9lre;$3-AR@8F+p6w;(DTcVrR()<%6Vi?SkQ+<|
zC8kmqoD|||t`&_8)5?rHp;LC!wLq)9H;PiXI0tb@^5<*G)76<TgQ^cEKDNymK=UPa
zETpR&QYQfE5i!24UmS=gBB~i)uu9FI{L@$FPCUiK`OOST0z?!;<biqv@!$X-m@Fv?
zh(_kdR85UzXo`RK7Q_QzyWpT%pXD}~w``#6Rjzhj-Y+Iv+{h8kp4pS3G>iT5S!m9{
zaJ>IoSzOc(mDf8FqIw4l>7T#{2s+bW?+UUgTa^rJMjVs+D8VJ5u$7(h(%~Nm3gR@^
ze@6#NzE3AmC`-+r2WGBQn+$7#V`BX9`Ka%-P!x1Gef5VdKYtA!WFg4)T2*gX8*PV$
z#?$ln_prI^cjE7#gsjm{TT%9zn%y7Qn(!Ns@|WsPN0WC-t-N{P%mgNP%zvb3p;5p&
z(PF;OKA(wgMugR|`==@|3jT57WE#qdSnSCv8ebGJe)VQT%Nd$B&LCc_4WMSw!=PFL
z*fHcfnv@eOU0F(E4JRF(R8xMI?*tsoM`Lk1n-QX4loyMpB9RPeLc`~BD2j^JQjl60
zhS5V{3gYCohilu5zp-;{I`6&(^tZx3&q3AJ?R+j>6t{MnzRNTebBV7ODv)5D0kuCc
zZIhkeN;zqIl+R;7_*n@q==vd2Lkjdn2QyUk!g3?ER9(O<vb%}Q1Wd(Z-wq*$b^`z^
z#>&@34XMF70@NK8v7`805_kU7^{!rXXkNWBS+9Od^~%U6ffRR2EO^*u=wV`ET58vo
z-mofEgY9ojX7IZ6U~hTGt*z>Qj<&bI0qtq_-n`JrUkhJn1~YdDaDt(J#*4`3SjbL4
zJ;}t*LyLdSY0oT-69stjMW3Ork=5L8iG+r*PI7y;X#NIBVVCV#8TYzn&+G`c`9+z`
z1Q?$}bkhExDxcWaayB;`=i<PBARXsY^nk4=*KF^;rj)@uOrM2zyA>Gxz7t7P?d9GZ
zm0G_WfA#2<FW2H$Q$B;K_>Et8BL%Zv{cxkD+p@*iJRxWqa)H9TLGR;HiNS1Dk)iU2
zHwzlX{(Q@3SKp}~?==;u365Sk$-S8vd>=}DGbU~=Q~h1g^#B~@4ihU5WN!eVZE%jF
zp7oUayNtqZqJ&u-*!HD`%<Y=Rw;H8d29$NM{Xg~w*`O>>Jf#P|;{+YxXiNJ_=EiH+
zYaL9v`EWkrRKE87lidyRd+}FaLDm|1c=a2|$qXo-k3K#%;bL5IDkcUHJJR|o0ZYNy
z*1+ij3e3ZKT~{z`h`l^@H0%SXbD=qTT%Iu%KeP42<o`18(Qz6!Gw5i>VK6vT0DK7o
zLzmxGeu(^Nj&2C49JNgrEeIhboik{fn92Ewf8K8IIFtUqCnurDLw&3*Zo)#EMjsq=
zg$@WnB=Q?^tU66y#g|K8%EYK8)R#dVLXSsbm{$s1{?LymL?o@VHGkdtO4Vm5<|ku(
zms-c@Er<i-v`$c+Mr>W4B7=^dz?!Z{cv;k3vq4RO0_H33zRMJ}Bw0b@En8w`*}DKv
zjM!hPl7$}hnGu{>xEkRFHD+L|0ZF5Y(1V~O4uB1ML}v|tPlAnLz!8iz@F;fTql)}C
zDEFi1<EPigy?(sd^8ds}&PP1pYF?U-Q`}az8?^OY0OcRVrYIvC;@K;n0uab%D!wY<
zV?9z`VovOmgU*=<eUG1hd=)Z9mt2awJl_9=le#YPs0jEQP@tHvyvFMX4HjUzuR$35
z!oa7*KLiEK6$T8~RcrHJ@h5RBC15Mx+*<CL-Hsr-|380F5cvIn@%I!daO(s-fs}P!
Wq@YJ{QVCoOdvI6fPM+LzpZ@{lL3vsL

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide4.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide4.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..22960a010347823b0149f7cefe9c434e187a8167
GIT binary patch
literal 6123
zcmeHL_gB+Rv;QWc_g9c6?XiFY3et;^pdvPqB9Ks$AT`t=y##`Yh>EBPksj&Ni}V&n
z1PLX8f^-mpAk`3h$(!eYxIf){_nbX5pL1r*%<jzj?8cdz81Zt6a{&Ord+X+Pa{ypv
zs$dESgei7&E;=y<$j{v9Dp1xZvBU%{t_H>i08p90edxe61&5p7y>G~1Fs6R5R<`~i
z6n4JO{o)u=JTSVDQQGYtTM0#_D48edpfc5L)79;=bRHC`+h%InrQdeT!al9?dtRez
zm8kQe80ubk{oxmwSChWmM@^fQyY9rI`T-h!_e<Y2GVFu8O{%VQ8O*a$3zes4ov3E{
z2A9}C{ytaN{?i<rqhXhE&G+*amkMoU79ya)z`JF4|3KX)P0zJb9i9wBr!242UGZht
zu%EP1xnZyCwQZBstdgJ}_38-1{L)_yo0J<a={mMaKd1g^KPc3+&rq{^qlwJXw9C@A
zNiuX!*R)N)>eCjR-lk%eXyKdVfGf(a?o+w{ruEyz-ocTUO_HWPv7mX*G_32((25#7
zx%AU~Na_!5)O)+2!p-f2+ks!HtGoJ`n&$=M=`}OXk)_(GYz^Dg;Tf7$WVdU4zq)nO
z9ZZ%sGIN!-7n3vkzJ8W6vElHv_gN-M8<`fFIf4!^(nr42bt>&1S+jfGbImDD-6|38
zpC9zHR?{w3!-i1ZvDn)8XW}<q1D^air6svy>Q%`E43+Y0X*(vR#V>g%Aif6XL{v6^
zZRV4$VwHT^D#`6p*>$J1`@T79)=8@H6k`vf^8GhnuLl~sm$cCNmo4AkaVKinXRBJj
z(?n*gS-<suLoq<7+z#x3x>c(qQuOWLz0CfsZJ(jzkgsK*_K47|gUryhC1^Vo7(M)~
zg-rL1sZ1#yfVx&)_icxH)L-}gqH6g@&$+@Tut4?x8(s8!-3JAiEnZtb%+*7_({nAq
z<==k!{+p{FRUaCLb?lOiT+;hS7Gtvd&BMQFKlotcNxb4xX%*TKk?}*%sn{x@)FYu!
z(}uvjkV7Y{j+x@SzqyG8^B(?hYd|#qksi}OPV7yD9{}*R|Jy;GK84K86cW94-Qa#O
za&@xBTf(RoK!Y9`5ZePNlBSJfOIX2t9y$vf1R($tekQ2sEraIKa=?NvY+SvsmMzHU
z0(l|2ttEgW#ZkE1!*j=tlRS?$TwSz=hT4zxA<K_$1$y~oi%foeU}rA=|Ly-^1QaK>
zhiYnkY&`B|<uX{X3_V1ctqVe1ez{RG%`A|<Iu^JuMmkbItc}CXX9WaE(XdkDmo<Eq
zb#3mo0a&PBYge~Ut%DMjAyezHapg;*+4idPkua&HdNWJ*P(2PMmDWYj9Ag{_u=n?0
zvv*<q2OJfK34WWje#xRuWBceLE6GLsu&hTaAH`O0G+b4MILIdk#gGNyuJb1QIDFWH
z<+KvM_r`Sd=BBHg)dJnLak?C^^LehybLYsL4<9dU6mBi0EeM&A?xodnRT={p#HOi`
zljNLIdJs_Tt*6WRVJhwOgmJ1fb|H`h3%%0!C+<xff9muLxp|j;TfC))B6)NBH;M}@
z=;lZ2Tq*}yjU<-vtb&8N7DLV_3}!(}%5thwWe=K!t|e*nLP{FjOSmp~)L_1>G|^e{
zKj>$REv~2CemZ;BPm~RgF`DPCn=botDdk$o>sC<3vtcqKxvg|l-ldEcQtY8n?0O4;
zp3lX@q_31B^9Jtokxg^48Z9!cEc-K<E!p_|hXj_2f0#(-i<o<rF5>Bs6zK4J;Up(p
zC>qK$=*A_YnuLd%wHl+q@Jt7(LAQstqb28N`1`KC1%vfA=IG9^Vn-)WkTykC5}d*0
zorn7>U1oY=oV(FPs6GV}_WtLcc22$SUy^meX7jKR8#U@k%UR4xHLZwqCMlcy6O5l{
zQ+m+t>T=Ko*i)OEm;U+FTjy8pDu$K19P<wQ;-nKTih)CJD$_mRIhMI=$TBegl^Eq8
zPwYY#Ph&H!i)3I*D<%P_oTLnAngxI;q^K}EMlG&s3eq6N0wYTE(Hz8$(U(BV<z%`x
zJ}>mxJ9S*fj|0E>zHe{@Nd9BUp)`mESL?;nQ#P+~;bG7fD~q=V3j)A+B4x<~AgjM%
z!tSP;=wqT(UO7l{)F8c?=b;)Qno>l9-Q)Ru598nTJW=S&TK2?E6a-$?5>8X;l2`+0
z#Xh`<<3*1K_Gje}lm=GmqGEIt&P<~5t-)5tc5Eq~x7?0^C@D^dUn*U@Ym)+YETB-Q
zXL5s@YptAAV*#QO@=M%wXLHycs#a)vD=P)k)RF7_s+3kH2P(x!Aw!ZG?{3tEhK;CW
zLy}urDBRHdiOB}F>u~m$qlxr$c%<m>1B`!hU5-$<i^XwfD}=(Y^vr#5T&7(&4}aPA
z-ZY3BH+kqRKKaKj{n4o6ZKcm3+@?EiTt+wOCdfk<4_jY35~v;8usWq`M4}q|6JyzY
znLCZnK_q<sKG#C$#2*J|#Ew52T9~Veb>;6Lo|5kSDF$`fCjmHryDRr{6g*4YdTpGS
z3wv%<D?W1E*1+)*t&uUy-q9-(QtI!^UtWIPd{j{M%k2q=sL?NJiUstywWV8~)XVW1
z>&LSD^U}5sF=D6DX2;$T0qTdNjQ}yH0>s|%dQp-4iF%=~H<&YMi7+lsYU#)<IUnxf
zMH{)^rl5;976Cn{-IR7Cwngxi4;F5@mEJlo3Y<L{>E7`hx=zy4N!^`2hhM{mp87oy
zn)2g4hh|6pKw+jLi*Uep=@FFJzkM~yJNo13biMm8)HC!*q1+v?BjOJppijt5x841&
zkucWb!-dj}lM{SQYCIfbMHAO+9h=yu_KmEpemqWS>#!aER4L`W#sDkK$h3xT*dvA=
zp4BO(1#El^_(yrzIizr(yW|mJ@2_Fg9^u!><J7b3MQRms5jlTt_ElihG4QV!+EJOY
zjp3-5H|D>g%#>>traceS#@H+5Yn8_Q`bQ)Vnuhm66i1{z%X|*+3DWg-bi%!txJzDk
z&Z$^kC^9}5-n>w|wR<!++P;{UW^x27cJeuMDi%{Do%M9oWuJ@mC*<M%zTx?gCqFi_
zTaD}>w7vfBSZtLq=>n4DQ<KIyqiW>o4}HhD3*S@TuvbWwU#q`Uom7gd67924PBY$C
z!*LTGn>r-H6c@zTQ{$}@W%3Lq91qd4pB6W*&gc%+Xl-9ZnJ@3~D3T2>-_5~Dy2u43
zL^9KD5op}@)8A`a$Qo<x`h$(A+wxK1!jr0+U#cb<)JCMtuC1SZltUrWBw|lZskik7
zB5L_tdktQ%=K6sBS1sSQ^3cvI$a#f_rkO@$>2pyU1oAINuQ<&qXC0Q~wrRM1dDvn^
zS-xWQmELZlmIav?k}NFAE(eF?pkAMC<>E!N<6XDr*R9-?mZ1;c(Q1EvOvWQi&qG*}
zmz<9s#f@h?TlY|z9#2^nEzS=%E!2f&^-dkG-~Mz$6rO*ij1hC{JBTbb2UwD+0UU`f
z9lj06EeBh5{arv*o(@O(AoYgVVVd8rhYD;Fh;j}^+%;#z47Xiy>%C_dFP+~HQ8euG
zQHoYXGs=^2gjW6vp@WW6=T}>-za!(XYKo#8j+3M(%-Kc}cMr*q+h==$3JJz3pG!>t
z<ouc$f8(eKE%Z0!LMO&QuQ%ih!}pxz`zYt}rb?bF^-M<wq5pB?UB9~?K3WnyF1UV{
z8A0fLwA#MSIAiZpe->)}3N$25Hicq!YG*3_*tA3bN$?TF6(7f2-o{I#jyy!J-fKG{
zyAo5OHzCGja_?G?<9~{8bOj}_N5}Q~Kqh_>Lz8cyg8g4xZo1<cH3Olk3%Bn*>1T&#
zWm>HsB^Gg5ayabl?#=wz2-s)EYX>7T@&`otsvkJaDz)vK)z5A$p&4{~^S7NbC#5Hc
z=%65FQ9ik*Djwwwk@xj9Tu)62A9oR;seciSN>-oBI{vw~rtVwr2A=-=_jqQ>^aBK5
zQB0Q6U1hEBaK;C0SC;si3RQ+$$DID{utRqx+Yzs||9Vzukm)r4;7{AF$#A`e^-E-N
z=tt-H_x0?GpA3#I#hljN%-DfM6@K=nrh8uNAhJ3~6NYYP3m~122_iX|DMv8;ZHsBR
zuTsz%Ten?5K~ZY_<Y9<7jv|b=$hCQu!Yq(IBh?!SF(<#-ze>{gLjJ>tHoLm}yK8dT
ztbw{qAZSiOK-9jWU&i9<Q_0*@fM*>LaE&$eff9e@BwNV4ju<(7jSC*;vmz|*6YY7a
zkoCCW+5YKXwhI6L&!1vH1vV54Q?j~;^?TST5cZB(l`cbs;szV`0k;4(F>(=F!_XlY
z#caWjZ1q~x1&Jca-ww-FtUcf`K-REJcA1Tr60g}oX1ARWj=Bgi9tNo1$2*6d<ZILd
z{VY<IMB|a14yDn8_)=P(+Bq1MhgQGK1g8(0x4@(-udKRz6#>LPsk{rIeTUMVU6I=v
zXncfNAeME_{5y|0HN_ALrV(@w7NJAw0IuS7fH~m(_7vte9{ICJXAIQixUJR8NnsC@
zREsj1fUxb)m#oNx6uV<FCU||(OUEa36Z}7e_vZ<H!laNRk5oY-f1mIX&;E4w;tXDY
z^pdU)coNIYM@`}rA?^MA2Szj+*3||7Oor7Cl>p?rNqbhUexuBGmeAasOB8}Xm}URB
zuaP}bseu(t|5omt_%3hvqnl+&vI{G`wWfr63G`yrDpeT0RQol+bq|GfW96lHIamb{
zKspczy(_V7Yz8H+%Tzm+oRc0|us^6x%x8r+W+(>MLEa^7MS()+%rw@@!0DZ@Sp=x@
zTYY1&x<5H;7fFM9e`qJLiGVu`2mR{~VqlUDjU-@f@OjH0xnD3%6Wdif@=d<aBU44t
zqqS9yWL5kL7!k^+>j@E!mSk!=Qy{vq1Vx%beaNn?1>vA~x(Hv3y@Rg9M;!-`V>o|a
zL=&O4RX;JDa4?7e18gdrCbf!15yaA1Wo_B6);w%gcW%a0wU@gB9{tBy+`f3|jtmkE
z(%H=BsGHF75G1W7leW6<$iS=vpBG^K<F_sa0nb;IjkznZ<)012?Muh?@p5x&7h&uX
z^$k3O7_g`+VFu5)YR%*dc_!ne7gu69{g({Q^lrB+toV;|H1=iYVn8~+65i*d?fiFz
z>@JVptOonfAp&q>P72E#7wQ|sU+1$@zak53%XPWy&6p7stzKuIn!T9NNV37#fFNDG
zqZbg>=e0O8$%g1J>Z=8oJ(G215}vewx+vEG^2Yv47L+7@bnJz2nQ{wPSF0_Ztxc`h
z4E75(6*4S&u%7@7g{7b-g*;TlMiAg&nrkXucDy}9W6jdi^B2<cq+wS=&<^_)g~dgL
zrocbr^)_V63=E(C&boX{a3?}vkN2_?EnB`wp1-Lx`cv`jGMyVx9q(sk+{hYRUK7VE
zmb*Ptme7}MhU&!|fA^R}zFwRS4pI)5T?6@jlq{2vE+J${f=fChdl}Sg6S_}Vw~=S~
z_Y6+$Jh`S|k3Qa(l6Y*5IgIn^?mI#9Ecmv=EIh^&9#t)-k7d_ZkJW<Z*FfCcRlAKm
zRY^ZC=wGz6pC?3>$A$N*>;y+{JPh_Ih`{gt<leI`*ch%VJ3(rk^zbA!ls{qIPK5Zo
zFqX$<MC6#mY<e=~nPgOz=bw1}(7_*ETJB(o;z+`t$olK7cyqkX$Nifr5O@(`&sY51
z+}~JcI%rBou~A6`=U2TibwI+S2Wn?XRfli+2dRVYFWpIbt}uMwsfdVEzjsvDvro==
z8dm@d%m0@7_*L&UQ0IV6ssK6;X_Ym>hFEuEwSxnYmW|Uwk>ycJIt}c01Y3+gK5y*e
z@ug@d`zvhbt_ef_Osd?&Xym2Y4}bE&Uakgl&^rvaDlZJ*XxL?;u*p!)L~qE(f0g`>
z@0>HjXyEh4i;MUT<4Q|#5ssC1QIJq)&G7Tl;e-xDS0+DgOXng=XP#syF!b$$<=(+t
zThcj+N_@}AR&=B1Q(#3p3A`_xWFS1nUDH<<{!|nlg8Wi@=Ceud4A;#J+yL$*dM|LZ
zfM2R4>4NFSr-Briv>usX>Po$6tYJRhm+qCyOElUZQ|TJGS_e!5piuW$>rl9x<{@0V
z0R*NGE@z8hQUm^gc<IBw!PQ+Zh2A94(U$z7*#h?+(DKTG8b6HjxS~Ex#@GLy3yAvu
z(=S@vJFnbZz&>@$%?zac(jNwdu6ATUBN>Lghkz*_&vjS@$*3V7K(8hx@2R@$+!=Nt
zD$^Eq6L=AD;TVLgt9=znag+q2j&&6lBY8}Z*-IYBvpV5vQ1m&^^xQ&qO{Qkp0|mMn
z_>-5{&s7}q!^JlHfcZ#L_d6fpEd!+B)C&4{*ejn!QNS^Tei|oCWb`dqv&7{*FQr@J
zPb#++V}0pX3<Y%3N>8!a)l!CP#RHAS@B9_p$LtmEwF`p4o%J?Qv0SkQPgAs(IRDV1
zteOPhAO{-ra!m?w)b>4CES>F}>BXFXCYte%E!ivQu7jt`*4h3x^(Y?(op_!mT2#cr
z-pL4b5`foc=(avbq&s;mSy+!s>HC<s7P-wKEn~2ULo@SCn(z<FeDbe-Pq6t~t`crx
zz`1+JsK@ct?z6+)U}aI01~ZNdbmY{$irvwV?dil>G&XdsB;dK}?>(lWJN2;|<&LYf
zV4;iP@`@sEzN)#BF()0;GyX#9<Av$X*|tG`P$*_Qhf*n_NuHi{?j9)#*pyu&Z$=N{
zE=X6%N2FBp*^NZ5tUWpX<xnN&%9Q04VAbmxFauu<abswkGLg0R5rPU&h=6)o=ojqa
zlWO$?xJ1bh^^Pu+ZtJFcizp}O_*iP=4;;eOI!@OYl5-UL^&utMTi=+kqZ}p^Rlpyh
z9Y1@KWDh0vZ%0C~P`kTtAj3HWPkNy<Q+unmI1)UeinqTNv@>u}()thQ3%MNHU4Dz1
z%&`kbfBy*ll#LYOI|l9ioU%bgB(Z+gk~hlX4VuJ<m2pPp3>3a3q@Z5};LKe{I}Up3
zj`%RI`fb~k_^PkxPaW-Tdm4?3?E280{OVOFCWQ9?hd!8gm1r}c7?A$zj~oW5w(Wms
Oy=7=}z3eLb(SHGVSUFPw

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide5.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide5.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..fb979bc12208e4c76a92e302d179c4d1d104c08d
GIT binary patch
literal 31177
zcmeFZbySt>);_#YF_73I2nfhFL0Tzk6cGUd=?*~}>5jFq=#W;VL68QKPNf8-yOnMv
z7rlP-(H-X<?>XoF{`&sfdyGBS-ea(yxbHcyIp=j<_wsotBThtco&beH5lP&?E003q
z!+#!WKY0W`MCc^(-~-QAUi=O!yOm}R{^yv{ZRy)6R9?{O{YS{3p0c{HW{W}**C7Am
zHCVoXj6ykUOWeJ!=%l?ga!OIrF&2wU)!&f7Cda)oDs@f439d&|jc4B_+)MNz)pvC9
zdG0O0_|_v{@4>@MCTAm!Kaw9Kx0W71PoZZb((yZGz_nUW8A8@i&j%&FJGp(N?mmb+
z#638;qhcX)`}+LoZiUW_t<6{R@1q*A>AOxC`zqYR3Op7H#o%#-5czy~1fLT5j6X)e
zj(m!oBGW}aJxCd<kx$gshaSjVq9o%{N0HAk5xnEbC&SZ!F7VG9{1XTN<ibCA@DC>b
zQx^~%_$LnjS%ZJ#;Qxm{vAxlhP5TDjd-RQ*D~H&j&o>MG-O=l_8v(4J1|p7K&Ge00
zosNFixz8d_q7r-XBlPrI5Yoiv-XLIaKXrF*q!gQE;^%iX@ienopP*ZT<*@R?HhmQ>
z`=g-kmyTbrt(Ix0>_;f??<Xb-ddYBcjYW(R!`D5C9(p`IamjaK_WDAWRE}|ptdMJQ
z%&7-MGm)yZRC)LItu4DgP^?Q5tHg$norf=oq$Z+{Vw2K#wt6e*0`6<F_-2UL^5McR
zN#`#ad&;t|OZ9($q<+2fk%--=C|dG0dgLKzuaN1+>z_ZLGh@li$d_(*KihGm{!yqT
zm(mvA*<03Cvcu<ATQD5oxw4NV4S9_B5kk9a@l@e51$WVPw~0dHIL(x~>=??9jEYOW
zCC7|pqm0glYaBYk-`+}OYgLNlW{bof)qp2kDlim(X`HMboP+<4u6sj?x=#4r7x!Ts
zZ2K^*T#48@c;zI<xYe>dp`qI=J*6BrXU(4<Asim+S&)`sBQN!w+`YQkZl|b0Z`@Q?
zX+2aM#BE@HuI!@YNw`;FVv19MaaC%ZJ*(^@J$kh<0$+z`J-bbrp}Ja7HA2w!#ng%9
z<(`jStWw`<5A;r;IldQ^O6fF>z-Z;MI$~aK9yo8yT!0Iwi2E%~hOU??9b0LK^E@j)
zr4p|o=kc^`z;;%Vg3Z-}sfIq%*)xaGpY@iKtJ*RR+?#N@Vn?nn&;9&|;~E6HU*1tY
z54y6+!pVDJz0*BSt+_w@7?xO?4sG+v{XsUYr!+-@%H2~~AN3i1_4<(n*ae#EmzT@#
zc4WPoAMQkJT;e|0u*D+dGCut2$y!DS6+^Yif=rM0Ii)U+H6rQL1(wfJ$`{U7-0rqn
ze`;Ei1W&*oKTWN3j=t=IG*x_BC5y?I^KQ%qRmNXdTKtL!zMQXay^VKVUCwc?`HWA7
z;N@QL&uMA#F5VD&u=YVRm}`y{aFKiK4diVUbsK4Ib{!kq(yfL9hcb6Cwu*GqSL_K0
z*yHUtT@<W#T}Zkl%P`AW)>U84_3rJSVG0PnYVY$hH1NhssU+RqBZczey<;>T`EK<q
zmA4vhR-d?ocl=|BHsi+0W;*UP#V%g^@b96`0`)9$oNdgGNaP-0wYTj$y#92`=YW$1
zyPWoRSw5E2gI9tCqRq`~N2EZ?s*slcLV+;e*)4Cak=!!*Wj^?tzX+$Q?Oj<pV=nR{
z!Mv{$gThz8tqX@YTY3e;_=s!ocH{rOw3^>AG{tnDyV6=?haHAakEd?d*@HAasf^OZ
zhZRBW^mDf%&oHXFg~R%bAK<|_1r^%z+$dEQ9i5SxwLNc*s%nU0EU&ID%ad0E>ve?i
zegQ>JaX{myRDXiWYrdfO3)vS-m|wu0+Flhk&h_$I>N?eP&L>-kdV=(G^;4(UjJ{<v
zRC{nfYL#jG5gz4Y;_cL6PT<e_?Go}hvA=@vGGj(cgODz)dL>&Woo&dV4}`w;x91#=
z%kEr{f>JcjU(!7G^A^}?aq@;O*yBu-+?bFEUI`UbCs`*RGTm>1$&m!Gy~e)po{m!#
zIjb^LJcZ!0oBtfF1rHUK)*3ZF=u^V}<Tu-pWw`wQ!Tjbm?2Qt%&&w!flX${7&1ur;
zikbH7&i9e`4)~Xt$}IFYkm(%dgjgu>Jwo_ppsK!fXB|61zB35uG<6YfBf@i7Jie3o
zT4?hPmSY<$bO{Z&8HBATtU4SB@#wBdtdC2tU=Bjq4Ze1(@%OWdJ#$C7xO#fi-pn(r
zW0!9kO<z$<LW$ZF=N^_^wMcLtCO8e36|S_ffQ`)+$UA!+H!A)*bi<uY_kHV)USW-K
zxP7v-p|dVrGPg*!keOY4BlGR3m|vU*k!?t(oJT?_;4ZE1LH<?rDSo%lk&Q<I!#uRy
zt+Um;Thl+vT-bNTK-@6KaLcdqO2B0;7Q3ZKZ@lG{?u-YZZD>Hq2)jfG<gL0PqgC6C
zM=~KIc%Jk7Ny^)RMVt~7o<VqU$=Tb|tLNpo+KdQ0KRzsLxJjROYe2dMxgtYW+D;)v
zFJ^5h>b%MD_T8#aKQFqtU^ro5s~D%BR;s$cf6#4IN+VNgYnv&|8pr6MwEXcrL-j-T
zr8=fW)QOa(uI~F2^*j~}#WUQEHOLzXmh%Xw-zt<9?bVfD?^f!L4!w)#*$IV!9RC)8
z<Ig+QKcNmTR=;oo)~0kA;Dtb1CQ;jzAeGqGq+WRXidGi?Ws>RySfh_wuEJ}SMVeg~
ziDE8JI3B-C@ddewNS|rts)6-~g^n=1sPwd%=oTk8OLHBu=_|;00I>x>uW`|Kwo^qt
zb~x1<;dKI@D|_=#2>CekZEVl)l5GgQ#~_p==QlLzs1pzRFBAOeS@0e0gU6~4{2Mgv
zq>?5#@>}(AvgU2U4HMq*ttd!)4dd7p(er!}a@F>cm0{ul^d77~p~A>3oPOT;ho!he
zM1Pgt;QfXRXC5Fc;AVBNTTI%E)6P?x@g33`*j4-)78d1oZ4oqa0`cvWll=>E_yp`>
zTi@v^?rPa3mx=79lI0x(B#QTg*!Tqwsn0PhZ+0I<YROe(^geE!u-1g#TuSPS^C`V$
zE}xci?>rAAj?nk!$?p&(@ls(%g86Kt)qZg#^jJ0-2iwKEwsXk-FLy@5Cdv%wqy^Up
z0OTsv8#*J^F}LDxtjOwQ$Q#~EU#=l_79b8dLii4v00)v5+MDQ9<UT?jIkFm>%p@f6
zHExIe)epSsZF!1GtZ0ib=QYW?QZMmHUkFM<Qs;)6Mnz(;nQ5~11g-X~s;Tz^-0+?L
ze%JLRU;RDb^J&*3?ltZ;9ym1c`kiMBYUqHy<ddTiq%|hXh0+g8{_6{!zXFC!A%+W{
zjrZ+paX)U*k^9Sp915Q@J%9M*;Em<y<gZFqXqwa1l@9UX$*KdREZx+i7hO0FJW`a-
zUnNp|j;+J;b22Gj2tuMc=CV6sKfOh>!h}lm&opFF%4%fvozsXj>HbuOd9Y*&6}Vv8
zU*G~4aDWXvl(~`DF^!%zYwzI=bUBFV1J6Z8r+1w&RevXE^JwsQ+q*euhioTT8IT?U
z{<@)f>bAmY&xVA4-G{>+*r&!Mq_#hC8yWtH8+bJ3bNs@-#;<50KaMj6%&+3%)OI9k
zL0~*Bte#*Y4)(FM-e-w-+B??1E1-0HXL}M4fRl+Ue9cJ_kZk!^k+><UbIkC71Jd3_
z<^1_0qBE6z^eTr{M=EcrJPDjv2secH*?hxVmgu6A)fm!{L`lGI)ez%Xt#9LX?G3V?
zr(WlJZMH$sEWDF_vvNLU{jSxPWsMOXBTJ+Pt-5nB$KE(m`Sy;L4Q~7~4?cyk1}un4
zr$D~Zwxo0(r(|s`eH&w#IJ{l^)w>h+r&_mU(fn59=(gM=s|yrH%%#>7rTb6B757>_
z?}T>k_}>O}4~s_ksx4whysoY}b&KL1PgW?1WNG^rWs!#4zIv|IYx?FP^Z<5O<?No<
zSB{fc7ETE^zFOti61CYEomm>KltKsVmkc-!(Dclxrpr+X^E@VE^_^1)etRuQ9ANGn
z9n_kDO^I*u8_#1WBG^<jYJD%m3)1BqO?#K}yV<<i<UhMwl}K6!9pe4BuJ00VqtlX{
zFmig<<~nkhx#~I@oP|rqpa#fIOsz5?#YuGJ9h>{-NaqIg59kYh&b6b*?@@4a@~x>O
z*X3zo%FujHv3K>H>F68P@i_U(3m<B2H$E5a<TMNGx={o!GIH|vJh1QmD48^AUO%CS
zGaDQDJxE*$9+e+ZqF&^5!qKj+l7PmMT#Y)7en1NT-6z{LIl1Ifj@NomNea)>=1Y`k
zou}Ym_}bXD7DT1I9>C@osuh$%*bdiBzG(KNavSYs__Aq2juM};T^-s77Ea6lWcil$
z;!#r0DJ>}VNvWBUdql`KMa*QnT)Ks9Qp?-W75d-pa~k>$EI#=K2+^mRN-N7f7~6f2
zM0|LlZwd`pT!y_7#I^qsLSw)IVpnv&W36cnhO(bYFUE5F=CV6b{iU6C>obD)*fs)&
z(w&Vum20UP$BLhfeU*sPZ|9ZxkIC{Lz~$o|9NqD)z|q+YQF{P@lNT1gexJYuS3x`l
zUcLFK5T1;-LPmF(S3-w}X9?q=>YVuW1TNz078{hwflpQEp_{djAUV2_e{`?@3He?j
zbPOdqe4+UV{{4+p{AN=)SWWRf_X&s9Yggb+H~B-i8oX^oWX#)cAobazyRNN)o`ivu
zS1av%#0^q>O5;!f9w%R{zniFvSwbC?SAPZ<KjWnKd}G!S=55JVa(`ngM<C7n$4>yw
zh^Hh)pId&(8110W8hc~MSsxX;QCDIL%a+awOPJMWH$dJ~W-1*aZ~=0>j9=5aSCF%h
zLr?nQ?(LSYBCYYb{X(<*%hAu|OE4}rOBq8Fe(AL`fl;f6KDH`T$|%$~hR)Qlj;9-!
z!g$0EuZxXq%wc5Ye<tNp{<gFGtUlGDFv*2&j231!2{XUij@!+~xPC0_k6qlf9B(#_
zqf6e&XW@+=uDUf4P}kOr3?R|z6^qoBUoU4?FJoVB`3#MdMrXBc?afH(h<)X~<{-bX
zvafCdyBawD(4@qPuG1&!Qr65B_ubWowT4-VC4Oos-ZLD7nURIw3?Cn{;&zjk9nV_^
zSqSGsw=J;$YX4aS=CH=D3X>DH9W!JX^LMWQAOi5QgzCB9@}h6JfwgQ!wp>6L83qGY
zm(D~P2|Z-NO32DYm+_g_@97DLo>S7=n=c%OCXJ>pfV%G&81(KQNTl8^FR`I3)AX_p
zvEbI`7T@HJ7=%X4ERHS=-aBG+@40LDXs<mkg~e~p&Bix>+Y<_apWd$GWJ=@2ry`hP
z$}}rRA2;<cuC+YLD|O=Ch~c`E;9Mo`p0zbjUc%#I6Ti0BMI~N~Zq;;M>s_j9%z~Qo
z^lR*!iSr>G2!LpWwMsTeCy%{0oHLT?-18&#Ln@kymco<ZwwqeoG>50<+{@fud9k+z
z_S;GBo;{~k%8=go2c&${fT%rMxP0)1-t7&P3eozNZUM)kONk~L5rS2(RXV#f0aEyh
zd(|AcjQDMP@K9C{{=%5Ta<M!yJ1+XUKO1RB)^~%P3!-#;J3(cR_tp=H<(`Wf$_9ov
z9A!eL>1=JwD20cNu`v~&Rfa6YHhwf3B<_|5wX75Aq_i2deO`+xh^QdvQni4TBc3Vx
zZd?9Y`D3|Mp2=i%@{OE+YZ>{!!6vLnMz;H`aPZ3>!;sKW%hB@dB!puy-C!pamkgSH
zjoZHPcwT$MZ!OcAEAdXKgIZJ6#;cBO5{CKR!u?>E&&2?jWO)KcJ!d}+JT=+N9Lw(K
z?6nO!KhN59FndSI?1Qh~Zm-dx)ep<1_2);T(Zh2T=&F_PUUUwQjyX9wkqyTQ*nj-K
zMZ-rPTA(CwyA0mKR~`2|R*bB|@ym7$v?hKl+&263&_s65L>YY_R4Jlo-MPQWF%gJD
zS(#I`#bGHc+IcK8Hs3~M`&3L{;flVcZfuh-gKJ$1<L9AIy@fp(xJ)uKmX*zy?TTF%
z-f5IgPfwpb!x(ggP~g31{Q|zm7~O>`dAmOPI}tA8mK#$2XV@RL2{v5Aml<9`sE{4c
z$m=@$YQg;!MZWSZRXMEg=yyKW;vg1X`UPh(#hqEACT~JHxg8w(zDn$tmYCr3YnskJ
zxwTznI$n*zd|VgaaCIaR+-Oe<5^~vERCA6tO+p<dX>yAnZ20)~GV>U}VUkmn1Bs5H
z(*~(j0=YGH*&XwaNuL{18kM?MUr762UCm#^Vw^=+;@WkG&!ABA{gu{bPD!M|@z`Hq
z+O?aCPKIvf>lwj?lp)jAb)z?|?wGCnsl!VpNu`b}qxnW{_4W6h-xDOEYAhIK4hxdU
zL-y(PO^wTAH{=tXp8W6`O5ACXTdf+$uMysPGFF_WEw>;xw;4FffPNDGB*N9$-F5+&
z1#{?SQWd)F?v<*qv_gjz^|W}XcyA^mqRNXozK(1xD^C_*TaGdhxXE-<%%r!-xu<uU
z-`aLujK>s6?G?;+B#A$K_#3u15mh6U?QYL+XZX_TUbCi}`_w#3NzX%O3bbI2ibGRH
z{L1jk#Wk*ucX2w3$u+XLo6FFdze|Q*=3EGM7&>>V!goY}fmdQmt!>e-===CgiR^)w
zdE(Z3yC}(3div(0%H;Tz4*5GVrz4nrlMKr(i={rD?v3!9(SaJ)+ty4H%a3(E*}`F9
z-rLg>N9A}7<snZTKI$CD!Pz*xc1}z<uOm7c+EKEo=vHLp>TPYxem_bwv@)3EInIMT
ze(a&!Tt-@RkVb4Pq1|<-`j84(8g3V*lhMUvmzgnl$LAFaw&>Vuk?c^kov7aoo-lNM
znNxtqBxZd4Dj^xmUZya8<xg6u$I~y)^&ok{V$aQSq|6cii4@PiLcqS51KTv!E4f=m
zHVj<DlWet*IzYEA(2utw)h6dh4~z$kd%?O3*5Vkx^_M?kb#U<wuk5iI?+}xN>APNY
zwY(B5Y?91@OGB@?RJ+vqnE+NomlSwLT^3lJf6OPKL@Ba0E5b`Y46fOw3`u=TJh)b&
zID0;-4?q9g^OSPenO-%0o28zoKUe8lpo((Iakecf=8tHN#G!qC8TpN7EU*xN0#|EP
z@1W1zW~lh%m2NMKky0X>TMsW(xiD0>pNY{SsI*-wk{Yy>a!ceX_L_D$P&bALKVBh5
zg+djRBwaFcewmev=1If}e2Gr}S8^bRl2-e$;NjLKc6`d+M0iLkZW;%o#-(vNC&%og
z;+cSZ%8qkVC3J;-H<#~6V}6EZ;1q90?8(;UCd|T0rW-%~^_jj$Y>vh<@;y<RiLWa+
z-xQwTXG$z;t~!V0rGl*B)wXAP@Sv9$q&}66WkhN$%~YQ-(tMV`^nvESK!WRQWg#xK
zecv#Jdi0a#HUlIHB}<$R%Nu520`;RY`ht@6=~--9pxQj|Zcmf+=_t$Lz46;MK1tBw
z&u=`9Y8v!`#zBKy=3s8L!c|{CNq~Y8RzEe94PRDsEtpo=(<OcD!`^GnmDkcNvN>AO
z(CS`zv~y9RcgRJ^s<}9+WkvJVTbMY`kB1sM`)i%!oEnL6{M}PZWL`V++v0vL?o2`c
zt%0R=;gKczAAK#-dvUbNvBou@sxHl;aoKTm;^#+7?KO%lVgzWB8~3TR;tR)E*Rp11
zCams;p2`ro+^Z^F;ouIV?IvZ>eqA|fiAKy4RLT}DjQalXKNmHAiMq&dzQ2xcbAEVO
zjg!a0`SGNk9RC7`R$vMtbX-c4DI@0Q2LUaAPB~ACx7CtY+BS+02ks0gwUg=2j)Z3v
zowLZ!6DY7hc=v6Fg53E9`LTC&=FTn!D*K^h=W-sL5SrZ&Z}jIsf)X+4t1KR;UQ2h}
zxVOM_!P5_@ZES~M#?2$@O-3z>Gr$IwO+>fR4qqeL)X<5}fVLq`GXMKGo#4Tl6OGwv
zDyG@bQcyLNfOuR8*f=aSNqur9jwxNwNl0_eiql@ATy2^>TL~*rI$YKIV<$YNe+NT0
z7LB0$`|IvaZbh#&1@~=n?lY~a81b3OtKz1N#-hF7^wSPbOUpoeuAQ~lp!8xb(+CJ0
zht>Qq>ECb>VhRm1u>#j+eQz>Uv-_Lrj4Z=l6ODlQ=wD5$8k21ga}<=uhf#WpRVbH(
zB8LNkw{TiU7ROVq7or{)qMIKTJ9(1(->guKn@UNGj{s^!vs()~(5~sJ@mI*C!MKWy
zU2F>*wWSA&IdQlHuMZz?Hjf>%f9UauUCZJ6_S?F)9hWll+8E3?!bYw9q&v$~M}HO$
z-!kio-7nVpf;CHSIiV-~b@UQue|W`GC(Ok~DZbm}5#L~bvslIP0h^~d#ELGw@^wnb
zpt+lgYvs8Zp69CFr+2y~+iH<T&wc>eFLVv7xO8+R-4|Id-L0?D^v`Ue^)qvR`>5DE
z23p)V;kUFEgfv=SnmoUIg4K6^05E7>qtJT~vmeSF<+52rcW5>J_05{)b7ZL)hqMSH
z{Nn80m`(z;J!-RzC>ydy!T|MWr?hD7{QOa^lUOFvN%w-Ry0$2{TM0S|jcJDI`G#cy
z<GX%zjfrV0K>};{ZEQa5pH3hoU?+TFdwsjMBx8-|!q;a7Q+#pHDP9-0UA^VnmY`r0
zG}PiUEGJQSwS*k@+emSe83wdoOLjtvW`BP{1$354Y*xx4eq_v<AX<qJI~6(VYnDL{
zkzxutz_(shirzP?jn<Et`P3sebK%ECCQ(4&h^Xx*E++x1=h39oQBju72Noxg%t876
zf+OMwLq-S7pH)vqt!3}M9&XH=iy9nImeg}u30zc*Ga33+)w?qwEXeGaq78%*JENS7
z*k*wMb$u&t;<+Y4FK$Xl3Rc*bs8Q>~_uSzX*SrgkcGG%B#PYC8^){>f&Osfg-Ny@6
zs|wgPm;JqH@^i0si7W7Hw5&EBg(e?iM2DV4;jnFjk2|_E#4k8s;pf)?Y(0)o87R7;
zE!A+gq)O`3y^?)1AIo9xyrLzl(_fJy)_zIIB;w*>36AyD{gDNraZSc4rDx2!A`OdQ
zUv`P0-{>$eviCg1CVeoTCwF{$oNxt?4Zlr8Yx<n&>%QmwOqhDsx4_0@D~Wn10<+zF
zo?h*O_IuTQtXQikiFJ@__1qT@a|BGJzne3i7<XMEbWDFcYVQ}va>{n6&kG%oy<P3P
zwHS=y2)m8vd9)NB5?Sr~Nn)z1{HZx59PnG;SuY2x*VD2LYhLDo6^BV<+!qEp*b9wO
zDaspw(j($**-a<g$y$|dcL$zYzKKrnEY<swa}6f0l@93@{NnU<$-}9ii!mp}ir?0R
zY2a`(XtvFeyR=o3<yXyO<bi~`$o;sRV=PyA7!GussZZ5hg|WM@gV(t+II3_!*+G1Z
z*=H_mpr1=^w+!cUf9wEDQMLVY|7`*(0XvP_9vlAH&PGlLD^9(0UQE|I`DcLl5{F0l
zVF^V;zxG;QA*9HG+5Q1rf$fhKR|)r@v)OTzLPzOQ**P}96xPkzj?34s-Gc``g-=PG
z9TWMCXEB>|dE|Fail;g5_OA156MFned+dRXSmJ?im#8Z1cUfM1ACFXJ;M-4_xo4bi
zm~o~8%M#j^r^wHYwrIp0_@OP2jbv#QwY=qNuA)N<#3>_I?IQnba<$@;={@W-0&iCK
zWp>npU-i!?9^+Y=TKtvo`RLP+GXKu^n?sV;b~%afV_`t!8Y;VqEN;q#L3=<zz}kwJ
zBv5r<rX5*Tm%V8C@)O4=b*6f&a0f`K6}2Hc2>C`^G7qJv!>KO82lune8e5ltXyS6p
zlRGY0NWSu=?^%6iXKP&>oy@EX^!ab*GY3ZNu7fC7meOjs(EH@etBY_OsN@fu<JjCc
z52QZHag1*Gx@Aqx^?0w!qKEkB3v>kAkf4s|Vzh~XVXb$C1~T5I=x4^i{yXF6QwJ;}
z_xBq{4CcDu6pgw_9UeR+#HVEZP)Vk#>YywTcRYEy03E`m=UHp%V#<gtd9`DvcImFq
zx5es@?op1!$0Y&mM;az&0;T)xnhWJgyIdU~x9xt-zOX-JIxVzG=c4g*)t??c(Nsl;
zbYpTXk<Y|THUe*ML8|YQo<2zI<L0tm0`lzsAG1y|4_{s_ze^Hk_F2>eI_V;Fu3u~i
zB956CsR%XO?ex(JGH0Y(Fu3lh#G^z_7NlDDg2q_Wlct?}y%FU3n~IGu-%eHoE#!EN
z?IVd_orLt+24i1?j1tb8IuO|e5`qlN-19daaaUbai5bwpt_U@hR5s1C@*vzEnz<K(
zIpa%<c&KXZo3UZ*8C&hMTgId8lnjZbuAVGG{?XZKVNa$++*V?~%h-3>=WP%09iX?q
zhzhqiOU??qZhfh`B$)`0>BHd&(B37`bDk#xb1AFBNeodJ9`+QlWp8_y&yyEuuUmXl
zT@4tT7ILxP{Fo3WG+-(t%>RQWR6!|9@C2?fd@~`aamz#n`Wz`)*+`6-BGjx>(4PHD
zZ*eweI;V5lm0wn2QPp4#`VH`YekU!bn98k59~9f{B-^!+%UlK@vK9e|NxL-i<%A`$
zdyadfL>b~i7Y+{(A*en>Q1y9^gv-8|Jin`OV=PGLbD+oU5sR1jD37Lj*$b|GR}?;q
z!|+t^o-sjQX%ja0&Vl&Qoh?l^p>fWw{`VFKt*J=&!;-f#l2+K|5!|{kF2Ax;`;Z)=
z_~r>@bSyXQZ;#1BYv1>3_$Xa;@QG6}T(2YAB_c>{X+-k%x_+e-sAuIFFKz%WrSzB)
z=7;916w$)-9j;?Hd1<A3CdQ2x+#=>WSB!$}R(gAAi=pEablJG&vhg0;U2Xz)GJ9w<
zos>6auXhRz@rl6pU8zuH2s#Vn>~!rGjqP2wX0qX^2XV#kO~wU-K&*#8Gzv>t+5#+t
z#YNtH(guy9hO*>ar7CiYZJwE*qL%m`^wd5T##v*%%6{-6FDAC9v|q0R%+$l~5F)RU
z2G%aP(><UO>J~IbC+wYpaTlSAL+`H^-w54Gb~Ads9DBWLi$v?2(6v68K3j&4yVvut
z6~yP|e6(F%_d2^v9u;kx!e>^vzzy@f@Two3z4VE5J=j#?0}~~s@clhJ5L>t|Sr>r(
z(UPB~UwiX6r-jj=$pk2h1~2Zl*qv&1psP@|U3tIG=2Y?e)g^Q@lUqRRivLuc{FN<*
zFo%T$OQub>;zfU_#ryKYEe<g=`41vhi{hF>OgUg6!|bm%w8aWC;cUK9gNXMY2&L6p
z)7d;H#~|4aYNV0tfID$P=Bt<2AO;HRq859*R)#JqXsjNMl-1%6O^0cUUeQ#&ppD`o
z;gk3a6pBWd%95|7V3)tdMr@9HHbeV}-cF$FhYFs!>x(<wDyn*7ZXSl={lP-OuK4kh
z<2#V7o{!)L_%fdzM~QGDkZ?-v+u5SJHA7t>hgW{^yKna-J1rE{j`!*L4b9fYKEj8w
zq^R7vU7}o{d&i6QedB&B>2jfs+nUFnPzC#4krs!f$<H5e^Hv>VvJaO1PZlcnw%)Y*
zCL**kbPjGx7Fy7ASubY#R9~>38h!<MECh6c#GDNY(Xk=vX?3Il5lQ2|jL2+86nUKz
z6<)&kDo5Iw%LMso)>Frlm&pU%yJ~OrmJKSiwpal_%Pc#aG++@Q_LJ^iTbNFP##Kj0
zM^3ZMe!<Wr7khhO?K|x2T&exS<k+c?o<|6$I6<tU3uV$K;cP1*J{LF3^MLB@)QzJk
zl$Rme*6V#>R$$RdAk+aL_Lps~&vL0KtH`hehZi#YJ@!%a<;$rWE8n$jey_x~TBVlG
zd1~%6%f|@iwolr<OOY3Pfx)>-5us-jyavZ!uvDn*e;ykX78b@$x4c~&N+1G-4H*M#
z&8mU^ib~Zw;b%Z@2g<?JBe?$w`z1dn(T1<<fi7C>hShwD)UN(5>s0v<ZQYd=$6vfB
zT#daQ_IR&VV(l-?>MiS2{tL545zP8q0`zKc8pt;i3!>$*7zUomG_81?0=LG7Tl-N%
zqH%~77v7a?&cT;j$pYF8a{+q?&(mg_EX>6-r(0l#Pg@nA2<34}`-@YZ;NwQYl_lm;
z>h0PmXS8@&O`}qbmL)Oum9`;A=$B!0l1gkhwLb8FFp_v+*>dmb)K4L(DCh0lNzs1K
zB=MlukC)U}*)Tg&tA5Vw)8XOaH-={qA9^H2gS|szndE!_BXUk)7FV7>U-6Y0I&5>m
zPhQ?6(CC<tPWCVE=8t@~bc4Q;$!!V}V2NgbPg_3j51nY~o<xhHQ6A2oF5-*VZ(4pw
z2n%lQqpmEdD&N+rKnGzokAYXi{uf;WCYX-$6U9c%wjp<FFk`*ZS61<(g!Rh2w^&E>
z+yM19V?UDrfqGo?<1+w7VmI6wZMr5!y;;TaJk_}uv+amFMt|>+I~#_6SRTE|YbSX@
z07ygTb?sL4rCT$H%jSupyb@rz@u_4%>opS(3)Xd1_$Io5jyYfvrubmCiobOSbo8Cy
z5}QHDjG&U*a=B->ts@aYo#P$OHl0@uhKUm!a}9K5$Jx-;QUzd^xUH{aVlNZ&zM(KS
zbdq)_YI|-;bASK)3-lijiKTmu1O5$lhkcd#w}oEho*-cFPHa8l!2)G%7wj%;-|$c#
zM|a}BL<-ik^p$UniT~u%U9{{*X2!3@Vm>vO=BiC!aSqfnObQCxNhKRm%i0=Q9#Njx
zu^OOS?GH{yN5Lw?i0Zz2K-@wQ9z?hJ;t1hO;Kr}SK3=kS6)n44uIVFcNJ$DiXt>$Q
zpAiW{m5KK53V)(sX;9-*6R(8haa>>vdDhTw!^qXEH!2*(w=&x}=f91me>|LiFQl^2
z3Y+Hu;dh7mtj7}nM7EzkmFLGtSYg9VNQ$H~AVp-qJd@Gsc4*_3xS$_FD1FoHu-m->
zj56D?iMkrN_`)@b-T#R`)85+jg)?y7Em62`g!!#S6UKId9X>a=#9rwrqli=xX(@Hw
zI+?K#HAcEQK=v`}x6x>mAb2oDUl~4aYOC?D{mN7MpF->xSK>o43h00?{%ir&bDR(D
zx4~HP<CM-dS(p&mZ>_)OziYn%rLcV5(PA(v)J}IkC-wr5_<ry6VYh|W@tNGIQ*u<j
zV>_(8zm2&X6)TL`fh?XaL#JwNQz~EME1-p3eL3N$$0b$QD`-kbRd|2$nE*ITs&lj5
ztqo>hJCDlcm}xIR39k0pTe(RuVP3^SKUXRAg1ISwZ?-N<srSiLmr)E<Z>4us_CGA$
zQ=~wyOXpHsyZE94qRtnT)H)7(`cbCOBG#u{G+>sG1-?L6wN2I?N%iMBA{H1(5;T`Z
zD|Y5e01-y6ykvehIvGrMTH-Uty((i@%PzY{)Nb^eb3P>gX&%sD%|}B%`x2EgiIH>*
z;4_iteCgY!1iR2Z?59tTV&BOE6xZ)qE({w0h~|yzAAaL<FxiWI3=}4fflK4=+u^G#
zS2wS7^2W4LCPS)pKd?Wyr-Oau<90IPC{##2ncHb9ljmC$axbj{Ptc55QrJIG^JE7R
z!Z2b{#KQrZa+q-X!ef~b<CSqHVJFrvyZpb>S<r#G=e=JX^@rmW4O>}Xq(ezAZouIj
zCI6VxMu?zkHPv8tA~@ue^cHKgb@mf)&9!h-?WwYxAUbeP_!Ua;(&298xcIBtM4s+O
zQk%_*>-y*}$>`Ppl-qacpiy}l<yidzGNZ(EGKxg8*%qRMaH=av*nvoU|92kk<=T%%
zGV&Q>Bfr`b)iK!xRuHIi+9wq`derhq7<l7u++W0ik~S%Nas*=H6j->tQ}tKfvYam|
z(+dTg0a7=0Ay>V6%M_V8q9`2qQeb%v1=pXZRqbvT4~)fyPC`F2>Tj}~-1A@Zc?0tK
zD$Vvcf2Y=hj&di_T7E;sH|GqFH4p1dZ5N|+!Nl}x!?rRZD*Xh_Q(pYfF%8o~6g+uT
zX3Mc5uSX#mdX3IQq$#QYDF%z46hsET9gRSjHw2k+AC&ySPc&{nJ8y6Ci3H^ltSNvQ
zrGNYsyoD`HW$4aWUI|b~4@4=9g}8^6g(T=6B7zIudn6L19ihnc<q3)ly=bR`wv62*
zb^vJ~cGS}d-zqX{KSk`MMu78_x82rgu<e>LIeKPqLc`8P4$cF_Yecz<7lli4+;v@4
zi)|Or9}&4OwW!1Q<jQl5J(P>90x{(>^w@x0eF5nmc-$rqXuRy&_NR&EXorOL`YKC6
z!K;(F(W|i>a&)!A(UZSNng@yl5U_t3EiS%*O7?vlsF^Y6)%$3Tb<Qkxx5TmgaaB^W
zWH1t1@>pon@fpTCb}vLx*a^<Cg2^mYHqIX1JgNAn^8)G&V%hQftrTog6MM58M&myH
zmA2~}zk$9`75gJf;5u};W3h3&zOG+F5X&Fi?5?Z^xZ>W71Pe;(xtMXEXYdN$qo6O#
z+g6)lsJ3#0fy#!cu}6T5iL%)}7a@+;6RCmIuUl|!YsNv`U~qd0oi?lKf{*uQS~5zW
zj#C#&)7-hInVxdBFdQ4Q`-fIx&Lat(vfn|@UhZAk0Vbdv#=wi~J9cC+1n{SBp*?M|
zd98f5=4n&2tP#zbGRN(0oylt`RJQ|Cb{_#R*r!?*NX-h~qw#<Sge&mX`^gn_={LxO
zXQkw#nU%HqnC~#>&Tn@bkVj{m4tGj(KRp{!>^W&MvKK7C#+QpoFSHt!`N|+cia7tO
zw+-!aqvwT67JWL9A;Qd8Pvd&vvJzR3Lic@WL6i0F;6|)ld8jw--{56|#Ih=8SUE>q
zA-7(hY52#~*5+3+WiG5gkU@K50eR*y$6w4qN2<zCPB2CX_|ogXS<|YfRnXf)tAX^k
zFnHzR3RcWt0g!+_p@AYS*O%&kgwwssV1bN>u@rkH_Is8beX(SF>vr{Ll!cC>bXk7o
zXfNBCXLI|Nd?~9f7*Kl2>Xa35&zf7k8v?IvcEL)6A$P@Z#nD*#z1x1gE1B*sp;a>G
zSZ^-b7SnpRN8-A6PX*W<{<HwJ`u}YK_<^m;L52lO7#55U)}CYn0d7YZ%FoNKr9d_$
zDe+(h;$(p%fN#5qF{=vPSm0KMa<DAt@%`I8tGmetMTm?)Ml$#owMx#=((*p0azhE2
zNIltFjBwTGS5PYtzmM;%$0kwi);`HYFL7&@C-Ln1tM#{Dtd2`YtPwxS0v^fvlro_2
zV)|n*)7_!O12$>dYD!IR=UFF-X>3f-R~Y%m&E4#FIh-9zd%?aS-r}UQc&{@t-ArfB
z;yjFuTjX&94T79JU<&Nb8L5}7ejzv>Q=j6a&ncGIqZxH^P`_>Mg;3PeaY;2jT&Kxp
zs(mwNGTr_r&;UR|W7-kk%C}z;2mT?-$S8b5h`WuI=a~hSBd6_Bqec<#TG0`dhvn<B
z;fv`>@V1*7g?X3K%Vzowf`hxbSn{%HyjLnXhSMxc!7!wfO2UjdKGMQSE)qUFLa5QG
zsm4Ok229fzTv6y%YGPNUPrjW&6Zc+eJo-z1my!cbX8!(0^pe{h=;E@VIPC$+K)dPE
zyijmvg1V^3ePcY`A0;hk&J`!LupiNw#@O=yIR>N-P_&i7ctvv2AFN?-iy?Y4{94sw
zi9nbB5*YbR=PT2y?#Wyww?v(|!eZ6f7}h*DqpT@$gy#2978Kqj$G*A>m8qayo0=6Q
zbIo760_4L0?LhvOlp@2Vr#81rFxDTa{o(?^p-jzZy==O;FYG&WxA~oJ8MB+g$=qw;
zB1OiafxH`_p`Bo)A|gSFFae7Q1}GA$s;=G5L90~)jT=oH^86=om66@4Fz*QIzcMV7
zlZT4+%c*toMFQvr8oj_cQ7!N0@w7c~so}9OYtHAG`yg?@ud-Cox>0<b9+6)?3WtlW
z-5>Xx12f;t*6Zo_mu>+0c5h>}=s^&z)E{j)Sk;-3HvD8#0xF5=qb9+*Gw#JqZ1^Xx
zW|GP_vB<X9!jQj8+S6@5^p+iiewSu29DHH<<#CjUP>U<?p3k>vXkD^&G6N7BcPwQL
z2a|Pq3PWM|_^h^=oS;}hH>4ipwT&^_8Mpl4cTet#;oYM+^A5XNF~fqz`>C}o5=r0G
z5SC_&1r}~Rx?8*alI$xY?-NA^!yOQLlVHO6q4kHRQ2!(!{SNZuQt$B>t(G(6Ga&5%
z<BB*FhXO*tXkGh9tIqyA0efjTP>Kd>=a@S2BcJ7HTap^E>3)ulm@y`W2_rhRCfsO>
zci!A92v@h(M)QU#qijuH7=`jYOCva>KWwuQX;Ncg9HLg9)6X&07nRs=`k$M2f3T?W
z_wTp9%q}6*1Z{IQnFbC^Y;thrjWB#0S?SS0A4BNI;<6qol`-eYkiwiP{6Q<kS0;4L
z(Z;+KS+9Rr>&qK&00*>~WX6h**>1#N?bMr1I8x{xa1XD@pZ~+&4<jh2%YQ+KiL*Vo
z-Y@7-4V_Cupfz*@u`D8jCX5*)b`*cuBt=iwAhI}Uie9am?{Dp>loYHZ0~eX@f|2Q^
z&bcHH%aOyO;my=KM?*w9T{`nOaIm%i0uHU-{{|cjD2D1>4iINg^Ym%%rE{#vlmm3!
z!x;y#jML^{zh;&pAWrhtWw;}*)$Jkz9NdVwpeT+A<|;<|Vv3G!dCiwDr<$qjY5cop
zUg0`GAQe7kscw?gC!`aKf=(#M$J=taH?=+|VrKS}8~R=2&2Nl)t$85YHG_R9o(P%+
zERk~$|GfP|q+yL%d88FxMKR-L5d|`hmee-BDg&EzNe-BtV8{o}8Qd_C70<y)%VutU
z-3gjxquQtldNV!53JG4sM%xY4-3g)?wTqP(k=->;-MGF!A{s%td;IGE1r)TP;Ousm
z{DPm8I%Wr9(sw~lMXb{^vuGdjhPC?!*_q43FBjn8OuFd1sqw2$;IHY;Y4-F}>GzdO
zi5%qJ##ST?!R{i*3n2Kx(65JE5MmAnqyw|eNQup6#H?b3QU8f@GG>BYmG0e$88bbf
zscdqh0Bn_B38b-;2U3@<RFEc$ihGF3&AyrG!<h-t9Y)1yTr6HdSpH-x|4oxm14sAy
zl;o{Vy8W?r>*ljx30BQY3Z9c1EytrU!qDaWMrIHCsZ5&8$AjF)Jm!TcF(1c%x*d%C
zRK7>3Jgza=p8P<6Z>Ie9tHkTRPu~6HC}qL^2`}S&>cFvqcg+0p=a-_;>sEp6?x!={
zkow><$Jj>TX#ZNKG&7_R1HlikIJ@FX^&tMz6-+p^aku{kD8&B10EHD9sN7nB)T?8G
z5v+0^?h}RT!q#V{?&6Kw&V~s{FWG`AP$80z0;4SOTOs8jAq@P~w1UmENmSt9t=sd`
znrlOvt~AUi1EX4f@sM4ULh669I}4$$ekaqVtajHSAeK}W!?WPh{=TDuMPHFs;D->s
zZXy@H{T$I~;*&scTO(vbOewol{UZN2Iv7EC!f6`MN5c*8bmcq_NF7v#-OqwwLwO1e
zU$R{lszLmu(yEA+aLJ+k!#r5az+%BY%V@kZ8z7{AuJ`dS^5i>L+(kf@<#Ufm$_>v5
z=m-lyE$@8$ln3?Z3T*}1B3;4f+0f(WWMi`hwgi$>CZ!p^p_qdGvgfzHBai;=b~zt?
zpZ>-jY+mzGd8~RB>w-YU&Y1=I22;3x38(1ChFAk1cdj9#un)C|@4D{Y-0qW;^LQix
zwak$wGgL$4a%|@H>a2Pu(}Rz7b!(gj($S{PeQd*;whLQhNCB{_VvQ9RW_5^g@<ZSQ
zk@BSSEzTmw+W!kC*l*E*t~}S&qY*K5#_xgVS_f2xtC)$h+6+I8q+Bz{EfXJx>eZ^T
zYqAQ|AV^kJV#|kN;^C{QFL>)Fdan58<UB6G_b)F4(kr>`X0B^oOa0SgdL4Z01=pm6
zWUTJOR;$)ER;dfU-L<O^UTI5yVp_>UU<I94uk3g&X<RW2kQ-q}^R5;Eu%BDFB61dI
zV~X%lBCeVy#;;#-tCXv(66k<MTCLgaV>T&>-#|@|CsRUW!}M=>0ex_Y^@BO9>|MGE
zD}JBWZpe{gWjKcy=xPxvpfGk-CT;3@ZxSWqnF^>-P$cEtBJX=e<Bm3NW-_n)kAz<w
zE=^O2+v*|o$j_A7@D)5(NaZ}@&Q}fEBxV-)^OMemBH3*B6uPjgTl&FHU5Q1!5-@`3
z<q!+<+yOKVn60pUbAdm?6wJMd@l$_K8H}Ie_<=AKo>&EukL()|TgifNawTKSEPd*!
zZ_K>j)<Tr`_PWfOEBMmReY;9StW>||54LN`5U2ls{$OivIEq}=OV9m!vF^;9Pf3vv
zJv3VWK<NCffCy`G$X+9)ia8|2@k(f)F*|x^rU;aZyzYNt`M{u|hijuPZ#cO9of76G
zP6*AT^vfg<hWgeNWz;Vb<ifEX0fRAreiyz8?|vc20~6$Nb;Vs?hjMZ@GCSz1>~j3K
zCV>P+Lkp^X8c3+3+WR8w<}&J}KU(g(HkY`#uQ)f?BPgi>oPhg)HLxrr%lbC-4Jkjz
zE}yvyAb;Ic<WAs}5zEJBiMr#xSeILTFfA&#>{JocT5{KV?_>wPL+MnEcH0!VVO=-Z
zYGDecHGUI1NmZG`G;_rntddl}>;+Kl=;!{j7X*Zq&ancYm)g=iCE;}hUR`zg*aGXB
zR0FH~%lT5mq*Z9XiIYfuYfl9l$k9S~r-EWbCTJ?|OW(V|9+<LhO+1Z*1e;0sObe+Q
z415lxQUf0U$dN~`1l#-=bdKiKq@RYgR{MeI<oN3=E`eI-sCc!yNQau9ZjbWU#LTBq
z+ppT%?^Xlz93Ca<oR|PUQ<`)5+pK5s`}tk;Cx=OXUf`l?i^QO560gKKe_4>WX84(S
zq{;c;d|XgpTf_`SQ#`8Xk=)PVt!w1PswD}i0nc~x!_3qG7|V1Ag}?^j7peyd{Yyq1
z(#Y=+`OB(J03FgJt%}jRcos|%N06@0`2*@!>4Z@b?XJ_x^iig1i&z2scWGH)-7ZVG
zNt;Zuu5R{OYKLPeJz{@a?doO92U7oAu@=TDlnA6Ewtzr?NLw(e)SM+!Rvef!OB6NQ
z^G6R{=B8(0byWTO;o_-`BOzmyqspWzC%kifT27Pc&~f)UcCTo~v7FRU`7nEM)?nk2
zh)Q8ao?5GfB=97R7`jy{W-?2{|C*5sr|N)S`BJN}0|~<=yLuPMs$0o^L`FOjtZ$1s
zRXImrFMdyKx|T)F?Y#G8gBBYXwppXU?o~hIE3_DyYy9|jNBQ8c!rBa8ZboZV=#|68
zvq`qv5AJu6)fLLqjASsRr@jwHTF<%{hhY7(<kpUuIpdpvstMkC9U<qTL4;3fbtUrr
z!rc{V)eWhx@`M{w2Vw?2Vg0Qz>(Yy*1-kVJyWIXwnzG&J`PDt}LxUn0x#v*Nr1wN*
zb|2(H6R0o$o7uyaJRh2zdk#iMtH-xbWt18k$9{(cJ2awGsVfHA0kj-L8XKdhsGo7V
z_c-KnKI$sJc)g@8OnKR~(p_r#H<WAcR|y*;Z8tb~iKw6&&;DJBXV@Fxqm3eGl%)v+
zTIupx{C;pXPy57r!|5E+voCEe>k4P_Z+~AuAXUZ}_;v{lZrcSliTOt_P484yvZbyL
z>P#Nu<$6I%30Z3l^3p?1iwBEQ*lb!<ME~CwBRD*yWO4xB-0a?%;8a3T-fMybVoH6E
z44Xo{ScfYwlfL9IY>aVLGns#1VKej0DjT!SBntaCVal$nLmLIMMn#9w*UyOg02z+|
z4W9sKvW*DRkt-u4&e<Zdbxqm3aP>2^!mi9$v2Whj;c1l&mK?m<p7!o9ul9Dv^AZ|m
z!cuw&f;o20R~-|{D}fxv%~6ZwQ$~8}cT|ajTg+068HjX9Ps9+z9rCFPk}=rl2y^T%
zmn;^EGs@=8w<<Fhu{0&@znhG(OD}b6+~p0@7GQRFW?)Lh+biAs6Dzn`1*HBnRyf47
z16B;I`}s0dXMq4odH=~l+6%ENf35Sz4Q;y)oIXrmFWabh4}EUJ3Esu;KoLTLEd4*>
z10Ou!HNh3HRmQ<mzS$+1hdgs#YmMa#I3uP>8*lP@SGH;Qdy2Ctf$#Kcnr?Na`t+3@
z7*NZJPJ*b+$7f_M?~e?2KS@MaghUiqI8KI;h*WSC@YYQ^kC7UcqK`tIq^nKhc*pIQ
zcUv;pvSfXBpJN86>A6Na8Qy%{OnV*e(Or%!>9<n8K#p>4hXp6lSY5ErU&tRuI+61K
z#TDRYFpQBiN1BkQXBn{Sr#H<?**kNl*_{ox)FP(xhjneaN3p>@^{rv$Wo9q5fiXar
zvpH(97SxV%yx(|lZAu`Kl8CJ|a9RB8cvs+(1|08_#e)N{`k-ojzRaxCb>Qv&4%OUs
zOvK3H0$Ts8`O2!<Cb}wP%jmObytj&%;fmAKsdkSMgSO_k#=h5mR&;>NP@t%a`&Vk<
zuKiDDVD*E56Z|SzgQ3vFrE^iKbLhJW?EvS@Wf#U8iFDYzi<@JInv%0*>!yfZR|{6{
zbgP%|UUaiWD}q_$zWkUJ#~^c2gZ!RPj)|;nuMrr*N%t_z-?*Lot9cL@@M3E7{dw8P
zFh-H-ig+<Eso050yV||3WuuijaolbkOwo0X*$7z-rwY@P?jM|H<-dG^Kcrh0Xzc-t
zv6c4Bht40X;p~dlW4f5b2i@h!?;>mlc^{z(gEr<GYWLfZr9M(V_ISD$p340R-x=_n
zB1hOc(7+-8G7SEYZ1y5>gT_T1@j}aaAgbOx9mtB@V2luuM%n4keCK(ven3-spiUae
zzc@vdw)O?DUjrP<;1o6toE^A1hlS}rk-W&op#Qfs2BXlme{?PpxKm94^&}I{88cKv
zQX2of=<eH7IgMS%sz(`x-7UW|vo-7uTWY*1cYoYXw+~Pe+CWl03J0rqA2%P7a~YGE
z#Y~qph*7aZ2AhN*mvrv*_(=x(DI<*La0BAvQ&7OcvE1M{mQA+Ds<T<Sdf#YMh3wZ`
zNlV!+K5(;rVKVv(D;w6Bj}kcn$aTHV<zDwUFcd|RcN~`l7lZXl6;sA3IlM8j{5(G`
zsX7&zD+*Tx{|X%2KrV^S9xIP3-d3?@+UGiW_%b<@?gHOXYT;p+00qiU#O_<nnF{sZ
zJJhUjvJ$pTSX)@*0?7|SNUP#J1s=%4o^IiB@&Y0ELssWNRDav&Csa3_MY&oY#$3rJ
zri?o<v+%~BZAef{fU}iuGc!rokNSp=sP6lh0ZM(k+B$r(DK_Qs%-C$7_E)ChnUfwd
zis)VK?fr|PMP_CR@;*9zZd#3(obRFfmHs@Ab=(9_qW&c7ekNeUQN#>!g5ES1esn|n
z;Buv$*;)}zI9y7QQ+rQpyFB6gLeK8(3YNoULE(|Ddr+L<wYJkJemGU>>=?cN*Lhx=
zp(5lwFG)Mn5JkUt);_}oG3WGh#<IVq<s!;8s?6DDcnu{@t;iv<;918?^XZOpb>&@q
zGJN^&`^fRoyY>|L+TBjh0wUB9xmMFhd*JLt)UJ3f?-`?ev6oR=*@cJQ#U(<%Z0JH6
zH;eY{%hXc?c<od3Ld>8T{&{HfCo4ZgY!M3Y1+4rDEF2XCv%QxrYZs?TI5LM>%hbMs
zl@f_`MJpvoIif9?&T=0E+-(xmt!9PTVw7><kGd(+#YhzPht}P2bQlhP5N9Jtq8GuF
z&=V#S&H-nsZ_%fks$D+T->^sD^qGQiYAxql2k#-Z6c8@Q;!pqUG2y=}<F9K4%7#@W
zNEk|=Qsl);bFJe|E8BP{a2r*s+!Qhvpu_(_0r%wv(d~<Du9{WON|3Qy<C4+_zeZo=
zjIw=~*4Hx?GH)ZBC-9bE0vKWK*7!gpiXwe)v#;QL7cau-gh(GBdP{Zk&a!)t%eal#
zs)TMlll#c9BHxq8&LK}Jg#BVDysE$He58XUWV4<ohDjUwx~Kox+}Tz>(Q2XP;r%zV
zxWeE`9?#TCq04P#s3HD5!3LU+jQ>H5Kgz}DL-3F%HNKleFCgVuIt@70WCL_q+K4?^
zyPcNU67t=zJ#m***uO7zNoP~o{ky(%C91y@v<FcAl}zr@?rX%&$&+<qCpPn^R#8i!
z)K4J5-4&K|{V5~J`w9<wm7N_>b_p?FW8N*NyA_4>Vrs){&OK=hn>p>l0?kb+TV)(9
z%-?m37SeTM2zU_~hLA^!o9#<wVTEO>K5AUP6$!nu5_)$OdgHqvw548`c0O+0;OdA7
zSczjet7Ds%C*yXIkN>FAlCTfCmycy2C4WPojSF7C%zw@zYFNk><edExgi=j8!AK2~
z0Gud6_M!o**=lKsSfdo0n!AtZuWtDgq=FB6h_tzW`&~_4-|K0jfV~=pi}?bf#klnf
z(h4iejCj1arT;0_&H#*5rc4hdqW0*kvgV#P)w`R=B8TDmxgCHET2t}(#5Go_q+VjS
zJl&{2vwe@1&|8h~@n_H9rESYx<#`0Diauk8@2q@#lbs?qx{813Mp@i*C6lsMM7KKS
z-=?NG$ot2xT`YLKJ$VyOhmZ)D{4N{70teTw1AW=DVprF8#9UHw3_7o1zx}hy_z1$L
z<l1v4Vy|Trdw*ReLJzQTV`e9p2H748!1~6Oc3zPc^k`MaK9cxlxJTl9hU2);fv}_1
zTY(Gl&p}1MGpLzwG)W!B_}=h{Jl6NujALp({k3i(frF`5_JO+MRUW+K8;3q0UgOL9
z&NHdMNl6uS>f{zSQJ<u}*(_-BQkbis4+0V7*^nSSk5({0Dp-FQ>L5Exayarw$rKQ2
z0AIulMAH$^{;>SC$q-0&GF@kF^J!V5N=mK?_tlA;D#~M}2RtR@>|sr^V^V5sWk2vm
zLwg6d6v_ZaM$$gn3P2lejFuXxWhG9}623}z3Kf%z9F;b&`gv43OLF0_9}I|}fOq>G
zE1C)xWjGUMriI_*Z#i6e#;1_>h^F;g4kDxqn|+@?KgdRy9&e1m`b60ePbFA4I-W5p
z&PPI&%r}KFwW9>RJV#;{|1(6K#;25opZ(cs4T$gBQd7CjH9=c(!&x~&fQ|d1ymjN+
zbDP5ToP74=1)dzCqMMia8iX9*`!=ngj#=pF7&7D49zVT*+IPCq3;L(+(vmd0!7po9
zj<TW8@v{*>{Z!>70&ccXHkW4;B&H)Syf=4C(s>KVNZX{0hW1<Wn`NDWHYhp89@a%6
z6sY|hS)^s#yDvT&NADPx_~A?onZjXG<LIkz62})Yy7ec((4G^Hq~B`(fpg9rk`=_Q
z#~rT6x$ftz%<tukAID)}*e?^>$Qzs)ik8y+%XD{v#HtEbv@S=Xcc=D=36c-ZVO5B;
z@Oq4?eFtxIQ2&8;#qcr8rbGRgk8BT0Z3b*LBt|wK&uRv}*93p2eMsR>BMd`pPw{?K
zLpN{(Irp>B`Kb!R*Y#ki4Ku9uarnXrS~soiiSzcw9sER+)V+hU?fl-cbfK<y3@IxO
z3r2v+7m#0@a6-!ki`SOIiWnk4d_)@U{_T0IPIS4IZNT;+Y%TA@#)m?b@;h;(#IAL+
z8?4f2^Hl&Rn>J2Z8t}u3v{_9r!=MogT#Tei*Z%i8oY7nbOF^M%DW)p|GcXgUVfq7p
z;iRt7f*E};6`UdDW)6x@;IbHse+KF=J%JVc3d>ruohRs>+_p;+t1V)CS%`iDkqhYe
zvSqVph~84{1aZv5nWE1+`C2a@j0)YHG$7gefpTI8SsTRMm^Ncuduw__6~&cGVb|JD
zeY#ig1pLmIGku8c$F<=US%XeEv4bsV9Gr}Iv8N@4jkQiHA;Pw+akjwa801WCDyuMS
z<@$T7mADcoEem|YDZLuCE+$cP!G+mFWU*hVExV|Ok8DzHx~lA)#B)ONP_wNB+6qOO
zuGh7#q;8N{?dH8JbGRf$M0{b~ZGBqW`F{0t;MvJla(H}>EW(+1`-_S!FqqzxRby-W
za`fyH55iLyp0N!)*y&WUVbWy?=sP5>_?@9rq)J6ip_haNJs|yQ?~a1UQyID7x5_kq
zITy`xQ*s<EdcA~tD~3px%_?lm1UA}5mzS3}H-{e|i-GF)(*M3D98v2j{Vn}m+<W`3
za}>>8tGYr}omX6Pwqx_waSUjdq9dN2ZhOj-%N*9tFkDQnnt8qaY7@6TeX_-&ff%3C
z$qMX=V5dMF$GfI8({<mgUJ3R@z%kmx(765o*WQ_jL)nIZe~=U+Atjop71^JN5M@ba
zC&^Ao_HB@DW?GS?Qk3jcR2utWFrKWLY-Qg@_I=;CdC$9^-sg9`f4+bJ{^>aCsBZUt
zU-x;P%lG@a2EoNT-;UpaeM?z&W#RF48+evsBLzcoo-Cs<7lQHEUaEwHUP~Uqm6CyP
znpD`lh<PHZXniOLb4fLya5KdQ2u})$ptghr@yWY(_t_b6>)J4FeRNLCZ_6;1(Ypt~
z3PkKJ5QopzwtV^uu~A_{IqfP8<$Y>J6Lo%j>H81e{QRjDX3y~Gq>x7^*HtYTZ8+1-
z<rqbhV@WDxk(%xM&}z$Fx%{yqx24ar2^$RdoE~T5bfJq|$x@WUrC(vK2J|4wbUWWo
z<$QiHj^GqzHy>6p6&V^r&sj0fLp6|CIHB_cYkg~tG%}8x%khR7qhdMN?2Ks!`~eea
zXtO~C6{n~R-cm_Doy`idDI>(_3%YEDtG_mdO9l%rpz|SDC{LKAX;<9YxGx?#(qH=G
zs#$j6#|Lkm22Yl~a-7hCl^$1Y-bKR1H~Pgt>gMo=lJ7D_c`$`B@g9=x#huAk_zFhZ
z=<R8;?ZLd@=^hfTpNA|Pu>48ED)AV^6OHD_GC-#xVw6*%rT_IsmT`Q@lP-_U`uQLW
zYJi0+J?q4SJP-fU1qdc4z9X*JA5KiIJgaD-^OLu)RcATR1Ud+tC!>ZkJ=ek#H2T*u
z#2<Oo{mBgMe6mkFE)}rD7i9NOrzc21MdNhSLl_buc<QJ!g?@KhJ4vG2(9^NbtajKu
zBk&ToOm3J=x>St!HXR>ir+V9_1eT#N6@L;bFrEf@K3X6vomwZvv1B%a>t=G&*{}UP
z2@#os?%!i~J=+zfc6}CZL@5r8lscvlL!qM~@USX;Z_5o_LL6(mC9;d-PTNH7X_694
zIquMt>A_MgeTZNDvIfp@Z7g7(j-ip`S74&*Ag?C$uu@K&I*vTkq>i?em5>9<R-TlU
z1rK98RK*jXY0C5`Cpnlbkf!9(&XHDdeVPU6Z+bLgg|)X0As~qPU(kgPnshit5GnI!
zDXwfvoU)VL<SHlJGxh^`pN~QID!*+GT251+Vq^O^GhIO-R}!$YxQOwaJ->kni2D|o
z5UvI#i5Uu0-UH@;>0H<_+Zgv5u63NV2XzVUTP0JGuT&!Xy?7v>A~h`<CmLEa=RIlI
z=<p`tv9YYOq7|xh>gK=@NQJ8PUGi3qk2ek2`;U`-L)gBlSz&g;5Hx$EWSw*5b1f|^
z97z%G51B@*aQAa=`LA!9&vWJ8yG2Hb9O%cdydE#JsAu17P#*Y<%dhxpbV1-V-!)w+
z9poArM`bc(KU*RIEV!G|sd)k;&z&YcnCyiJnnWGU0mVm$Yt>|Mf|)y*HqWHzrt*E*
zyOQO1$j7MGz+F7S(qSKH$<q-xDLKlNF4$^xsUZr9+_q60&~tFem)G3qsa}jr)<mxM
ze!BemZ^05=tzvX`o=^z_@Y<=wChV^(mL#v6FuzH&I8($W@}5p$GIR4UZkXigV75UV
zIwVx^5dx2Cr`s=KP7iW9T~pkt{+6VeBD)IJ;Z;cK8l5qZL;TTD@vbe4uDznW75x7M
zpdQj#_BPfAY!35gC`4O69nve@Nk}d17}cKz2pR}PMU<2u0L;h#{&489cm&xcc?+l}
zrK`t4=-@zhSAd*cqvkYXTk4+=<_$XHYkH9E^J>_ISTf&ohkJfzPh!9?9K=S8eW&9r
zq2`)_PKWf;dvMoCN=p8F(Zeuv@?GBw216TW&(}R*RT@&41r4xFMQMTI;lQp6zkHDk
z2|#R_ib`bg$tB(}{id&85Onp+2RxD(T&%v-uIyIV<r)CRR!V`sf$?u);+lMS;pOFZ
z8GbCPfbrw+ry9MK+6Nz$+}vGT(DR$$&0#n>nHA_29PwUA2}jZMn{^RPFR*nW_g|%G
z;meJUvISiq9OLSPD(lU5AAaSZ80cO0MR#v!y2QZO-Aw_<HOwKS1UlU3a!!<B=R+oA
z{0*;Fl`u4l!8?o@($7qN^`QRKXobc<z);wLyeB<`4N<zNb<q%{XN_sjMu|U#hx}_=
zD4Wb^n&9;~qW+z$|KFI;!8cPFj^FHfT63Zl%7a=a*0Pe=KfwxOqMEk0N0g`m0rsf)
zgamONW_KfGLJVU`X;AP)J#01pR~I0h8K_hJ#PSPU!mOR|8Y)w;B#O}Ok*V!U&)3}i
z@*Uiq)0Jorq=WE}ILYA*@3|<FsZRS|t`hh5F%~qXn)Q6wpL7ecS!-Nlq!It8k+&G-
z3mc=gGj>E-z=By3*f=MCQHeBGy!uvEGf0#j$TP2K3*j`?_qYsKHCnuzm|OS6ol}HE
z56|zsPrsWP_aj+;qKNYvEA@Jjrs7Rhr_ZhBcBt^B>{|9Cmdf#@0n_s=v<)21ZTj8=
zALue)lAF_ahg}A43QV4&)H*$+f6(^S4EU9P?bndOD55-R3t8?OBOYW}6uJ>lF6Zil
zoQ9Lim(+2qyL)YMhT;l^L8zqwSu$|^N@U<SRBqC8+@62c5OC?Qv$Vqa8NA(Gvw_Z6
zgI}G7svq!N4X-m!L3*R%?t|DuByk$K)Ne0a`dKd4-KJh#h@#_<$rU+nS8`0`&i>u;
zV;?$hjH-^T8B^P0^vcS5fB;bp9)s&2>VU!pBf)wCiUk~TFYK8LK)T7AZXm|^F_XCy
z239pMmge!~6WT`KlTXc&pqi>TD;6WtGr!95=!ZFgiAj`4B$B6{ed5I_;voyK&)Bl@
zi8<d@Q`{9rXH+Rb-2MbKgFL-2X4{7On*1rKlsGV6XM&*P)0b1X#0|E=uQf!68(0w5
ziu$P!$yul+Br&}6z765x?oTBz4eXb#KXBYbBx_A2gZQ`sokhIyCa@h9XawLdcvJ`i
zGRBegnBNvpKRH3!GuQ-c|3DXLu0%Ba)GEX)XtmT^=xPf_eyF@}qP`<@b4biG-l`eK
z$r>G$cx0}Arb^_{4f9;q7}%FN&X&f6@6=MhLn~ObhzW<fmr-qDuzY7(EqrW{QL7s`
zE3UT5Ww|N^jWYRL(w}o|7W8|9C&7~7fb@bFDnSq);N&qfHgHYI3+wz*-(QEdiP}4W
zM>f|oZLG06U?5uVi#>ke+G@3k_x+P@#+$K4mjmpx|1EZc_uyZG7GEuEYy7l}{HjqT
zPcSW+RI0?#T@v6sO&^XdzYE>eR4e2H4ySHQ5-L33atqKLjl~pyix_Wq>`EJ6d=O!?
z|14ZbZNnR%*BwcWV2v9Qsx7`r3K6VoVz~=9)n^I~Ola^T7H!6sR%E4rU(-rN`wc!r
zHqe+~x0M9Kj9<)zhXiM#aJER$w*ey<X+_&8==ybh$MOJKfloRSdJRNP0;>mpuJDML
zLMg7L*<Dg}-qq=d4C8^-QV9K5-iM3mjDH7Y;pC%nQE;0hred&y#|t4WQFkfOXlyo(
zHK^*bUg~jLJI2NdI)gisjQuEIvXNUycF}0;&{MjXbv+7atqt+-Tn`EhO#}zmOK*xM
zrp#|viO$9so0EF>Wt~F`l~;7Dzo~Z?-akkfh>nJwOts1BHUc{FFYK9-h0bRXEXEiS
zpWQ3$Sdw}|PE%wBhec-Cv}xJv0ulRlyOJVSss?vBjz?wf%fW;+xedM4?L+@vJun^J
zufEl52vK5tdj+s&-w@<X(XH8;l?5DbHB%&}Xfe!kQyK2c0`BU%(dS;1g2F0&+B8<^
zHe}Xdvjlfzl(-Lmhi3c0zSynrMFUZ^smAbsP}QwaBaAwuRS0*X^m>-vTSlC}a-F+z
zJe(kvgkorAt2<0WnprP=>~8d6JGBT24A{YE>IDq@J3%YOWCU#i_()ZE7gJLE;B6ol
zk=^nl-c59nnvO+Bu7p@B#CZ%Vo;=R*WGZz@sLGsC6h3l31_CZWmx@dhyU1w3J174a
zgHm&aQx+oV$V;uJ^R{%z7Ck^(ZgBxc9Aq2M$<}tDCoAt}76vWuopj$$!ssT#gi%mN
zQBP%b+jSc@N>Uyk@I~owbP*s;$}ITPtRjhG^9RB22&}QEpWev*fQ$`c1(o6hkopbm
z$O-kb*&yiloFD=v*yok0HRHP|)&$I7gO%(^5MpP?zl7y>beJ(W=i%5X2&vYM8%jF8
z*pU><k(}->zfPz{8k7HvKgXulAM-B~<jfdFJH;8EHXeaGr(#`yw&DaDsRid{N;#G{
zZ&nv@N9in12zm9oE-A~P!c@>R;TF!tcgmXluzG&jnDF(8TR8*60k60FM45|+iJdT|
z$V}VFeEZvz@k4Fu`SSwBY|J=lah8^r<^$=bVAAB-S;XZG6CvK`2sF_9wUim)P5^PJ
zIQ!2*zpJRur%^l#*o<d#v(VOEDfY@{i!6})%Qh;&d0V$`nWWA)`PS~pd=mZDr9XWs
zdDd+gVe;fcLs8K#U}bcJq3bFJdU_P!Ruf*E-!K186C=oVO={nA99yQza9ZT>oy3#U
zGL-CMWZq_+&6RuXkmU>eUfsoW>aksEX5<Z$eO_7}Qk4WV>lz}wns)5cYo7y~UDrOG
zWSi{OvSfTkA&&}zcg(xLy9G%yQ?nq!usi(RL!6!ov@XzIlx3#S(HD@6h?-hv4v)ji
zpo1AM*<6=kAahRWgPZD;nJx2d?HY9HM9sxDS-LAl|2J7mPxR5rZX=)j6cJWtwq3ae
ztH6QiN^^n6j7C6g!C=a1?>5#Tj&^bDVD(!Tu<ZhFbov8~pTAfhJXMtXE@(^?l66*W
zhsz>m8s5|t?LY?1k+DXU>F$~F3pce+h;A9e-27|xZrz4IS<~39IWG-h+b!a7?NeZY
zEo4|3Z5o)XMw9&(-YM82rO}-kGv@gGMTP;39m{OufXXyG%#HWiD=O^#IU#;#*EvX9
zlDy9KUt0~c451f^$xMqbGA|UpLP#rxo@sKf6ZWIZ?tn@HGAF;cBW>iD*O^c>CTFtg
z90w5AsQ?bdE3LFW7lC+v8+aO_Yum>LpVx)R<&7#gCx@38L%|jU@fN&2H7fLtim5M2
zUxwGtSy|`ksW$Nw8@tx)8`9szb1i-M3^Fv5zP;ILowT4(UAJ-tn^K;z+M=qkUy?<=
zCYVVV5%m9Rs`~&Ucm?Si2~g7FCOdM}<HeDpxl9RU(}xw;@IS@qLAV|Z=!}RR=hdA>
z%=edCrLl>Byg0lXv)z}`A$au4Z~>#Qqx-k-N0!aGM72!<@8D|qA{q-1hkl&F)#Z5a
zGh!1Kj1fQ|I)Copa?n85`;{#LmqW4Z3r1sa_bS96b*I5KeY=ach<AJP%J-e&oro-p
zqLAM^ic`u_jTgf5?JwH<YU3t8RSdE3394KW%Tw_F5vf;*{pYm#vFsr9W|5P+*T{QZ
z&hf6?8j_cidfZd_evptrgO1DKKTc?VdczLct%8a7M?D6F6+lqk$V{0^1XWbqqGr!S
z3b~8{1QQbw4x7S)$E+Ja3zjtb`R?46XK$U~{&SU{2Au(jC~D|@Z_fQCGB-Q+X1T%k
zb|Z>yY%z79za2C#A;Bt&68I~C5bOncxQJm}SVG2v@jJ?ChHbU|28SkEGv2JaVX>tO
z#${W;jL*2`Sjrz;+ABOSt<}(MU$Du7i3(4pRMjn&L=@)5y+S4mGF0KCq;iJ)syyMQ
zb6TA714p(CR=w?ex;0+(-Z!L+N-FJtqdwFzn{PeP#zwi$FN$VAFUei#Q0cJ+1`#L!
zclD=N(rJ5Jn%}KFnC>mvh`?>yD6WKQBg}MP0B|Q8uD|DU=zPaEz?pXn#*2ki)L~8d
zK48?vYuQH-x@jH>d3o?RPo5sboDvldZaEKcd|G<)C8h=pKe5ad)Aq_vMV08>uw>mt
ztBejML>S9e%wM)?3OME&Sb3I1>-s`vT>SlF4n$*tW<7u6zxLL6>XB?9(JL+8uDM7_
z0ta^MM<YU{fKKgcMg^eCHohA)eh~jtSlY$F4YWTP_PU|rhVueN&DW!2R+Q!_4}FI`
zN56KdR}HgsyX5?6lL&JYT+bBDO|CFuo=T&V@?Z8v_f)jK+V2-e;0ZU{?D(v(+xIiJ
zVa{gjG}q#v2zrif+r_>$Vrr~yh9{Y@Jla&QZ-zDli~u@lu?cC6#Ng}Ul&($;a0_5c
z1jUQiNV5*os+IIPn1j2cEg*NjxBfK@4f^L9iW2F13!lx)40ufiXxNhktSe^?D+#sL
zDBa6{;;0<|W<W!rRMx+~8bd=v^464_nr~1Uuk+f#*w`2XD&eDx19e7s_-ItJ^-efW
zKuRaB0^~Rl@Bm;?NJ7cAq*fI{Y~Ax_f`|BpCgy8{eR}*8^RCwelCF1%K}xqj1&5sG
zTYhBevL)oVOlN17o(}dD#C3zvU0sX9&L+1xEO`1UBsbtp?>&HKmMOOBaFqD=m<}!`
zHuhVLeLHN@1&5kM;aCT#e{aRSKa6NMT{NeGcKtPo(dtn0^OE8t($3Ry?Em<*a+~Na
zgOuxj{Q{p1_2Z8=g+qu+rLI>(FCKILAb?h1Y9BI=x9zKJxG_?BB!uBv@M1<)RTa=)
z!cJ|VX&xhD#$O_^iz&FZ=$UTmb=qbE?HF#!WkyukhMCtHR0ZAlhFq^BRLeyjQni4m
z)?}FJkOh5a7WAm=y8gDC<fD#0DBpK&M=#aosew#Sp`O=`X#?8nAJWcAZL4{4XBFET
z5w%Zahoaf`gS%pU9fr3&<)nJkCK7bC4+<{Ru4tty#Gi>h7jQ;N2{%lnCrd&yKj)NX
z{=#C8abAP&5bzRL@~vD7fvt*I@tGZ~5hw4ix`AG!^KJ4$KL_AFpA_3RY&RpXrut4t
z#co!`b)CfFj(!&Oj-o!QL&sz)2s&P{`gT_!I>r)9Q#EedFzxqpk-l6L!Z9M?HlpeG
zh^{hl`~7^sEfMyEJw#@c7t5g9M-vi6BX&s8BcM~QI-<Ic@eF^K9D+eV>KGkzN{o38
z40fJ0RC_rri4yLvGwAD*Zj=}4!de44I2Txq_w5?)7f&9+a*$?)vbI2X(eLEL-*G$Z
zr+k}?zGuQu=FAhYVe-jW6Op|sov!&u;DCnSF+}WfnL$kR(A%-sMRhPkNG`(R=1wQt
z0|L|>4h-=Cv@~EF$CAy_dS8N)TOVo~i#rx}pGEn5Zcb(`v(BG|hmu)X`iQ#R@f=il
z^1bH54J?V7fSdj;_%MFiYH1+t{i<^W%@e-0pd|T-ekY)tJ1Ib=JR#eqlzQpMC6a7u
z)*B$(gT%-kHF0HsrO58a+zTy#V_RnJ7s>|@(!E##JPE91pp!w34#x|89*bQ^{OA4c
zr4R%?$#eri6Xl=1VG-(d$rT7QT0&FV@aFcO*o0}sARC=THYAT#;#fJ4ydwVg>{a6&
zS-vvZ9s%wMP539N%Va5=%y)s~M{ZODRh;TvEV6Pyf(?R`<kuYL9_58ePZbgHcr;<x
zM#cLsl^XAwiiQG?L6rM%5E#}xpd&qTelbE{<{Kwpp!MP#5aSSQ&(@DpaO~Fbb!be+
zck7ku3YML(2FwH;_^}L!NGKL+8tA(4>J|1>4{HG!eT4y^(r%h1>Kso+dUxwd<pL)K
z{tBpOnHNFbs;jG_AcRG!v;5lA1keNjE93L!kl=-%(i#RJTQ|QrK^VOUt~$pL;YT|O
z?}aS6+*|`0W#4+A2s_Hd290l14ycJK^rx?>6GvRFQ3wq-CG@vUZ_YvF0O%SD(16+H
z1Li)aPwYAiKeYdHsZWI_)aIg;4&%%H{8St#zn=%SO!H$)X$7C3AuuGYy~pG(F6#BX
zcHc%9P+x!BbpRsgNklOqXSrO)^ief$*V#<|t*0yfGLMt<X%VE37VvF5knuK7r9r}?
z;<DD2?pj8e;^P^+d1ZWe$&}Jh-|HDU{_>9sPeKYl44SHtmVbJq^l^mWqUAC5V)Z>&
zsGFig{OgwucxF)EjL-77f5$pa=zO)=*?0kdq~)sRss6K<1>Up5lMCG^=F2E?5=gL#
z;FAWb4S=6pn%kI=_BYVX;4F+FNrLM30Ti?QB%9_2qnt)t!{rYR7?8<t%&m8LvZKI#
z#1Ll%B^{hMaoCy>(J6hH461+=U?^AnI7>AHP1hJuf-0E4!AUAqlCs}}k_6n)ub(O@
zIBrXHiZ_N9d1`<EGpBpV082J5Wx4+)9C!hp!E?u`1v9%ZUl?NwN|p(6d#>;?6M%u6
ztthS^fg7#*3lxy-8MMlON$tz6)2R~SUEXMmgdu?1;UjqM@6uk+9TRY7n<2;`4tL55
zIjrS0|7xG6mF69Qx%hs2$JWCq3<xn0c1+tuKY}@2=%x>fYZ->*9$uUKt{vEDZG<r_
zWY<LlJ$ZnuZ^vtm9G3<_`fmjt?tqm${GXA)@~)?PVt_p=Z*)(K5jtUNg)PlJQ6W@b
zK(y!#^3GKP4cP_2I0ap&@%v$BH1CYP-rSFeB&U%}W6OSi`5El#0(LZ{T>%1*%-5hL
zovB!v08}>!;b7cX+Hu=>`*>L&=>9SSSHC_u%a{taO>*}mkPmHCe)49woB|Y$uxGg2
z;1!C`GdLVWNp9=)rVZHbKy3pP9t{&2%^wm=vCUdK7fZ(zN(LAoZqb{6_1D%{xhBxf
zHvy_3FI$2a6Y(n=bpRAEnP>G}*#q5ALPPf1;u@(RN7mx9fn7oGh3%ResxcoKRlnld
zOBB(iMc;)NdN3NjSJ4g89~Q2ligZWnFL(YH4XkC*1tUj&mZ4xVn&?W2WAY=?qLXbM
z9p7s|&ph>)^JnVh9!_!T)~{PTt=M6Go;ul0i+vk$a+ycXBv5!w>)(GaA$BG<$WLXw
zps+OAas0#12upbpN9(>9LpF%sYK0qLat`37{|Vq%==m6tBT4UZoH%XcjcGTd77Y26
zDz(PRn{%(e8Q!B+yIlv(joK}S#6xgZw(yy@q_dYH?n8sdc2=B^9r?#9MeXiB>1#tz
zqtAgt=mfc`h4UViaR-Sa5}5o3${~GWr68SYPLw%<eTX#J`KUe+&^FtCWp|L%P9-uS
z5=>{bpE=@t71<o^%3XTytARdRQUTcxzi>>fjLaEZmI-z$38ep;juTdItO?LngTNV8
z#t|Y`1SMkA{bR#S$Hjt9xjvvlX&kfek1I6rH)1?K6p_2wD;=;66Nf9?S5Xh=r()k)
z;51#l{>SH(0v{);PZvIkA!xW6ag*izI$pue-0-@qe{J&=^tI=ly7Khp?6%@4HT@Hj
zEEgXsIr-aN>hHI-o4IWb!e<bW4>bVu!Z#uE>c-{(v>K7GiOlORi=6Mu*D;q<t%~jq
zw5kMIo`F#eNTMUwOPov(%2yi&)@e)_RJ4Uv32#q=K6prmzVAHfYD*Hj33jmT{%ln9
zVJLOgUuQpFefv5R1IHV!Qs0C}54kZs0ShiTss%s5UpH2^^s*J#3mtMpvB3ll`Qy63
z1;#+gxkkeeAh(G^eQ+wBI}w(U1PKHD?thP<{oiNs|G5szsI+YwLkIrohi{>J706n;
NtfsB{S=s9D{{iZvAZh>r

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide6.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide6.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..cc5e63a173f3e372d11e115e567ccd2b47b19037
GIT binary patch
literal 21277
zcmeHvc|6qX`}asC36)AhPNz_WN|v!lCA&hhCnWnan3yqCmQ#w3HG3-iR%ADbvTtQy
z#+b5iGuE-pnCJd*IN$U8eV^y`{9e!V=l73_W<K}ly080M-`92B&$Ts`nGc>i2!p|x
zRc>9s1B2}YKkjKium}93z@thB{-HzMQC5WIv~o>?U-nvG)3^qMeF|Zsm_uJ>biQSP
zfWcU5pg(kVPHC1f*quC;>(_K2o6h2x!}UhJCuoAdiTqT{v+ch+w3PGNCT6Jjt3XZW
zg1Dk#>{!3wA*tU)=;`12dc9ilN?%Ug`{K#lJsgKeZqVJhecj?^+7Yocw(SZpiA5JL
z&HKMSdepLEps|tmC13gsrz}StVkIs=vCWzq*_8iH)Y>V)V514X(MOv_0@4oC(H#Pt
zIJWl^_~rKDD)9I9YcP87!M^{|2kZ!ID!+hFXb<3$!tFi9Z(c`9=TJ6OPeZG;EAA;i
z3H|Vz^FKfQ|75}{xzJ%?g;c|;l_*Ee$jmu)^-=qU)VPJ`tVl{(h{xpGyoCsH<NGB%
z$9zvLpEqB`TL<YqCt<L+{(DKpLPwm}Y(Y-~dZB?^+Czrm$F@1JA)TYMVG-;|{{Abt
zVrfYp^<lmv{e(l<5lAqEj=94E=cv7Pa@{3}am{!s-d6)xW(T<4?dv>zT-5`$tBr&=
z>;+Wwxwa4e&;6trF*{AE$qyF6aJ~!utQm+vk2-f-c?Xh8x^Rm<By$Q(H`uFU#T}$Z
z!*zmXS9;HxtDUhVJvWiQn7h%W8~^aqHn~3!4o8bxd(8A+Zs@+l#6b~BTq8~EqlLTD
zthl#dc4Qyvk0WQLkaKaJww`WN?Q=exHc=OI51!bj{059TSdxz<4vnISWNnD(2R$rW
z>Z=2~>5(?ugpA7jDp4_%Up!9u3hz5<(UOH0W%hGn=X`;F8=j}p4VEE^EZkD(%wv++
z$~+zuLc?^hh6~%A315Uc+S^|wyH^rZz02NfQhP6tnzE@uiu74Kf&MX}+CN^(UDfNe
zkQ=<zq{fCnCA8r4+3l5!=*n$6QVO&2Pji%uHdfea${B`glH2=s<*cILf9Te!&)M<q
zCEjat2ZXktNu(=(*b-A%x<q!aLpst5MCBt7_S|^|+Yf_1N<!}^@ypFW{~#w=;}YNR
zR5FvF?jh80;64oId!o@CmTQooq825r;&nHmmAhZ88-1=ThnHiUi||9kTHaGSD#*K`
z?Ny`2NAGR#U}(?d>F(@^l@+HijPU~^(9u*BV*X;Aknea~BL0&_-8m`2=GS5Jag9CQ
zA={Lc?{g;4;lc*9Wa^N%iQE65w%0$gp@AOu_Rgq)FL|a<2ac4N8U)MY@uc*zEkb=&
zwO6C#WMD7>y_`LdTVlja$~-h&^A|olAv+`J)zC`!=?-Z#9+VFYV}L0nH|oL6P_qTx
zQf{|g^M9u{S*xNsKEn3@<2@o3t3S_O71Se#ARQcgLSO!ukK3fCv%tq)Ds?=#!7XO|
zWBEka#tSNs?q&!p7uo%rr?q6}ze0<rvdk75uk?+f9jFKH|6>oUtG`0D6!Y)6&-Tvw
zt7$s6ZgJ=z6v|U4wfK_^(xpC2XOx(zUD<Q;?a^ZxK7Ll1uWCdl?DVDk3P)oiroQHI
z#|#vl1EM2vjbQD%0E4CJmK+?et*wRVx293v{5uw|FHJ)9i)_vY{=26m1ob^;`^9Wp
zUbwj=%Ky797CGYjlhvX=hzYg@mVGwUqv^NI8ioSdk9bQGg?qwax*5j%oJr5M(7!jd
zG<K9ImG$JV8Y9XgE%#nh;5|=K*E|h_MF=?2$yhapLrV0`z<fVt)TM=BnDLKQ%J-3e
zg~*#mN73=%OQ=Py>yHt|k+%P35iTwt%Djr(-B3;L;Qp(VqW^6buKgta%Jr_+cxir|
z0SCm=mE2&bobIE+|2X(dko+MAUFvuz4LE|~rMR=N5*`1|x1Ispa~jcnS;f1sK)aH*
z5X~yVZ)c+nqt|^}-8M2p2h&bp;HW~a%&fHs%7yQJWHddPgky?|8;r|J2IlEsWXT%F
zVAq{B`6F5F$6JTgrth<-?Bsq7X7cj<<690e2Bkn2hk>q-{YBDKBl6P=>&FxA=DI?G
zTK$XAz_-X>Xy8Xqw4a&$-VB)txTCk*#6P4jfnKr~cnM{>g+D(i=XVuYlo4O9PLJIY
z>iEakTInSpQ|9aPX}`IRv&^P2%caA0nk4=qg)|t8BE5CVZ+BByl)b03LL?+?4gSG|
zcL#>=13Td}&+hX*T_^S2ou;v3%~|~}njYe9B<dfB7ztzXm>cx0<**JX{%Vhx;#y9v
zJ5BxSSLu+`ap>m1a`yLw!%r{hr1<R@_T0~da4XB-9P<;3@-BpH<VPQ_5{ho~p6t(f
zGa|wGwtq)oj+?+31_~WGrQ9B+%dy}_Cp>VPQ|q^WrN+L$+QjG*+rN;&*1yn9!&X!S
z`nxhoHR4+$e3HKSen*OiENVq&%tWQ>^m8CzQ~jTC{Im~<NPWs_xovjz^0C6w6l2YB
z@f6OI6HYL8dtB*qzg_NZceYWgW<2D^J~-mdqTUbquMzp(%FBt$15+k=rEOlZ4gi7C
z8=Xb67=I#_<fqEa2h9@C#n$Kc(iiuEgNhV3k@8xa(9wjCh9y!g&md`^?E+b5gAdY<
zz919j`bV=W;pY9rVd5I=>WF)VYul%29xDTT+pli4c}J3)TW&f1a7_cBt166rak*dK
zhx(#T%-V9_w(#>50^^&}QbAuCe|2Nh`KH%oxpj3Vzfr?kI%9RF#XRPJCG$~Tj(1I}
z+;L6SR3V=Web-Mj4nNrI9C_CF$&}ve);pyfFz#f7DESkuYWf&jge6<aS>K=|RR-W?
z8Ty1khXt5#M!{xu(ZSRdf?24>NDxf;Cd?YX=VY3x|28RoQ_5ycsVUYbn{PPJj)wlX
z&j;O+_AJ(^?)na^qrun)z6C`X`&nNo$k5HL-sYww{t4bxiWu$uR0`A07o<4;^1rQu
zA0Tp_zj)BQE0i5j@pR|&M@k~th5SHq%y<ezPMMoIb))4eDm549ydi3osJfY`bPNNI
zdsAC5k;Aue0i0Bv?C5jnn9_-%s4xv(c{<~luvXW-mtH4R8CGQHDK!XUsqIXbkK_hk
zi(A%i_>B`Qmn`k&ViYsF(x$kA_$KAa``k?CO3cX1{^%mJ&6>n)9M)R@aUrtXgeQfo
z?T`ZJ`sA1yyW>kYTEJ`rUq;*xq4f%BjRl;Jl@A&w&X{F#h^BfpKdoRNavl4YDev<^
z)MYf<!f8UnV=meqKOt6PQ;$4dwpLJND%>GTL=R+3!up$K<@-;x>MHFIJgF&uyJfHQ
z<vo&3N2(MIuuF|?4~H<GRG&Dx6r17IWcoXOeKFA=(Zls6*6206)Odt@+jkA?51V#&
zd@=|n9U3BaIl6G+-76YVxgh}-o+u6DwjDAts&1I-(@CgrdDniKws~+zP_B4BoYo18
z4S1O#<~+kp6)ru?9nBk8dxJVv*@dX@jB!y@<>)o977d9uk#@Z~%ho0zH*;LUvz*%c
zBr~#nSh)Ct{zbHKTyAj1IzM6cTx`Z&g6wQ;_oj>|7sn3!wOx%0v5&j*<^gl#Kcepo
zTxgPD(@!w)(P%^D%(-ml9rgK6YPJ~b40uyyd_wzfM`Y!()|SD2WbGtWU0eS#7Pa#P
z+sey70@yKn#UJ6}&UqL3piCuh^uYbMejda=F;q*$S!<T_Yq6h%;$9DqH%}PIXO4du
z*q^+qBgij<xV3(GC5L%`<GH;<ANsTcJ4BIRWIIKWGudUfN|G0QaG&6Px&?Fik67pZ
zryof69b?|5TmX{4TD!JWNF@*AdIjE(k`sGZbR__jKAW9LYJUH)TW+;8zCn&YyffHj
z1CZ?8R#~1&8kz)Jqb9DG4RK=TmS3Vfd)2oC!xw?_W*5<_D!!3hxPbh9jo`7|H|E+r
zqW7$=EC)#~@(S=#p{Cx|d|X|da!vmt$83O#0^Cnn;A=wJ@M&9It;E6c9Qb3R+gKy*
zko3_3YuvsaIT4&iZ)B_G@JKmca}YQtXkIJAtu2Aw*e9q;>#r7-FhzAoI*9No)4g@;
zmS4O}kP*@~%X^NeI)2BB1X{SRg!5xQQR6Lj-20KkMV^L(nf|*HXy><w3(+EPS<>6}
zXY1Dp>jcsF*6<t^mNps3R{70|K87$0I&CSzN}8&qmh|*OU4l<vpazDR!7Jida|J!6
zBkeH@IFUN`4a?VpkqVv0?$BJtuu`v8DLoc*jn~XEV?@-mYk1v2mNSPK+ds6<*e2sK
z_qI#kIF##%?~jtX;1+_ekQVZb19r^4+O_@FCljL3jwO?Ovyo7aogaAC<n;{yqxFvx
z@0()r3H|p%IV61ALpQ8>>-N39t~hK0nV`ZGT?|&Vv@yrHLQ1lwPS^Uk6=Z9|!u8%L
zPo#s2_-NFQMz6vYsll;Pb5)IrhpVh?hZ&h@*XZea*-MKKqacy*`+8>DFPpL{*pQO3
z+|g3m+OntT+phgRtCzfNbdvdyZdBUPIr|csY_{(5D))*W8MYlcP#*a+5mPijz+3T2
zyY~H)CsTO@f3xG719NNN3oY5!4EeJ>%`l~FJq;;?M|PeOB2UE{DSKxYzR20|u1rtV
zi*!$_c^NNVU)rwF;J?J2>Tu?@t26uHQGqPB$Pmsc@m$Zlvl5<O^1IsNzrxciu~{l&
z{J1GPeI&Csy4=<ow=~D5ABiyoV$yu7+}pZUWOPrS1Tja5&@5E<WQp!-E}J-}JB4dd
z{k#NNOzZE(9%ijit$r&=AIcnp3=&+Qe9PftgP(Q6;MH?*TUavUTlIK#gVBL=L5dPh
z9a82k5}u99ZK(%#j8EiaQo1x}tN;~LsVp^GcJeXh=`{&iXA{>MnCiK&Llc=kM_cj2
z@nrSQJ(Kj%0)Z>&R!q>Y0P#e`s`;=cA|5Wu`1DnE)`m7uOxq^*t?Rn(rBfY=l3qFS
zfLJOm&L8_28UL}m#n1Jl=DkIy`Q`{(w0+~gT_rb85T?qa`EcTCj$UL^Tjxr**5lU8
zJ(QZ(BD`wwS%v3c5#zmQf0S$9V>xjLvpH(wkN*iNcAqYFS4&*rWlLld9FPI1QQzW1
zg`2szD;J&W_$KoJ6Tj;Te>vSnDRcO}W2I|D`UJ96pYG`H<43e)XrwoA&m4U9%Qqd?
zz<H#<;*GBLw8jvhM>Krjdp`aKxq?e`q_vdtsU>$#Q~b|dpavd9!QD-K_3JY$zkg$s
z{_zioX94Qj7wT@b5~7vSh~Z6oczM#k_OS);o;X%M>uJBy$z1r(o2xb&MjP9R5_kRf
zXJ{TMpU)Tf?<i5pPb@s|mTxxKYlQ3yaK5+jNwfF!sEEHeBT9ERar(rLJQ8haPrT1o
zz&~<jD(Q_GyVZ>cLNDG7us)4(m(Mh(8*S%8oe2xo;L7_$m!+F=02r63Rutb%6FAyC
zUH`h$WX^I^ZEF)PFp)JrV0t6?%>aMFr*j>}&Q%mt3a|2sw;y6GoJPrJD&q5H`0Rnz
zn*?JdRaB2g4q00@F4|Gr$67V1<X7Xf_-ZaR+v+rqFZWd3mN7A=@uU)m@&)0=Fy7;*
za-B`tGer8XR|vnQf}jNop*&K(@Kd`+Ixxs`e!RwdWKfQ|C&~&w@oCYPXjpj>UI9xk
zT-eC<TP!_{YO9<f_G3h}rv3b1%A*?D%#mZb0^UAdy<OH4giY@GZ00>l_X{%T%aMWH
zDUeb1$76qr+J^mN(%xTUAe8RsTBz9zc{Xg1IMa?|gbNLra={S_8Tg;k>vmqjMt34{
zJff*9&&PQ*gRdy0`&ThVv$xJEE{y)RqG|=AKz~+Nvw{(SOC7N?pPH3wIJ(PZ;Np&b
zyc0PjZ#T0{)tNuhc2#&vg3><(Nf)WP<F08qmzD38dV`sW#xZ>)`;erG8?mV&#nz*v
zKQYidy)>lkZ<+wr{5dkB$AJRfK%zI*%%p`Vq6OGxwuJmSPV5R#+WK_6uWDp|uV6%6
zHk&wN)F;3j4j5C}yQ!I%{&UwhDnTL5&9)cwQ+Oro)dDOh`?4BPn3wUWF;<^Kad<au
z^wX9tTm&vzw(t5}H)GiY0>T`b@yJ_bN8X*&>g8$1CljCU&%kX(sGet5w#+5p=Jcx$
zbQZ9;75W9n3%gE&<@nloBGWIOb_Et=WZjr^3KBTIiwAKR{eC7dCNKtCEp^FzZp70!
zbZ2y-8_36*5`ovcwDY=l`;FF~027nQvf;b9IN$-J9ruc@m-bs1%0+Ow_~tc09IRz8
z^d8N4hU~yD4a6p(S0XN<$olhN)_{vSD8Wzxd{{$@D(}vz#6STl2zpYU3bk@vF4C%5
zwpuIm4M^F<d(MoF{6QbCGchFK75LHLbYL@9u{Z}`?7o>D!v3%}RJoS2F?%*;P<$yb
zqQJuGz<AM(vCm08<iyn~oAHac9DLdswNCQMc)sZq?3VEP<}N>E{FsBVd@eWZ6~%2a
zm#HZM_jzT!A?3xnEm|}m|JmTe@UCchXwoX9A|uvCZGUImR5b9PjBRKTwKW`JIG;a;
zJTSf`-pmP@$>sJ)WG+hh!^8U4!-+|Xc2i^GNE>ra?knRjzpIro8y!n|xcXy=IrAW5
zy|`@M#Mw+RmFS0xO}wPu@{!kHq_fC;(k2iaO6ort&vi$pe2mmsGtG+nb4DwAOS8XL
zght73J4*iY`jKM#@3EIQg4nZ2ETk{MVUeOLJPd6uRQO~T+taaOXIZo{EyTk)n56<s
z9bU$W;<>t5rK_fJ@+N^cUZ^{|a>_;NQS0fmk(EYCMeIL5uY~e&d5I4Mu$ih5c>Q}$
zn3_M1Cv*k$P#DTCW!ZcUA4m~?e166v?I(Lt_^6zEOH+cwV}gD|%E9f-X`5R=3qceI
zB7K!E!)6_c?6GR2#=O(Lo}%m6^|5Y1aX22t^LHiZ?FOX!%V-MctG;DQG~Zje^{84&
zK64P!`}Emlud`j^r%YD9QGdHwt-G1XYSUx=#p7v4EDzCC${3Yew-^;%;o3EGd|E%-
zyHwZdSX-H$+Uh%NEzbvwnPp)b5!bg8tvJ%}VZT_u3Y{}JkguEZO=cNC1;<SKLeB6_
zUjZi#9}@bD<@NJ1DJaqj)fnwDE~k05kXxmEY?7#&nX#O<AIIxsQ4(1Py?cFr(g!BQ
zjA0aylVxdT{zXBkxT0cb_3F6AE+fE|v<O)=NYJXIPJ$Hnpz*rVza`-8*b+zn+l%_w
zLu^hJd(_MV*{W{@>cljTNeXT(J?}PAM7}um8hwVcUmY{i&UTwEh4z&}lLx8acDsmc
zs0K*Z+#e_Upg=V2p*(j2f|8{XQMmM%jVIjTtK<G-0OU_4N$EZUT@Moe_7-djo?oyX
zfE%wPw6DL<-BgjJmbndyCWVJ)Cu{#PG*83F<EF;^5KS?Nu?s`2GL=MvdC<YjGHEW~
zBM?L<^FtCMRFeh#=-s0`4sKthujVWNw@)Z*z)d8VI@k+%X=TS6iK1=5(OGQVLKvh1
zaaxLZcFGMf3OOGT>l3w=iR;qsI01!gxAG1@OxF$a6ux)cNCJ!67F^)0)q|xK3KM)h
z_04ZjMBsgR`#K+>$UOJRzgo^ztp^vIE<Kb9tRt}t7rU&ySSFTaEkOnv6UC0(aLr54
z-nFJr)!4feEA8(?tavWi#&xCOV$*{_>(lV};_x|Gv|WY6%)0fEs`&Bcu~r=5t>eN{
zS(*1JPD`>!WCr(9b*YfR^^&kyi6Z5wDaBS3Z&m)uloN-dmGU?ZmC?vOR7W8F{$}X5
zkV4+r{_q`7d;V}$I!I3&gN%mmxwO^#V8>8OB1UVYK~60Ex6>~S4m<?LD$s0YBtMty
z`@#(lBj8sEGNh~~9zoZpxOS|$@=f)U?2kxXfY6kvM}eJ}pu@Tr`mKIVw%E9^<SI&6
z!Ws}tG$jw_;289uI95>gm}n{Qv$pxlC~^fpr!c4(>VHWs*>D80QkfXje<y>nEshKM
zHcyuNWBSY^W$HjibwrMjyrm|vI#|xT$5XU^#WOe=Zp?9cImPM(a#6mSS2%xp6{PQ!
z)_l~l!m1UdaG@GRps-cm@@N+d4Qc(DYOZV>AR_BMXi~r~?=`{rA-#&@(pr+7g+BRY
z#Z^j<zf+Ik_z(?j^D0fFZc&7&ig`Bh$2OO?Wi=Kihu(YX>&%>cn?IV_S``-{?k?}t
z_owOPsTY@ZU4YPc#)*o^s7y4F`~oqG1ue7EqyZ=ni;}Wg>F^(YYBsn~+qsW*IYWLm
zz2Ed~RQ%=5IVEzvZj&DoEB6Yy!i6E-AzXVERIBbIV=n_7t^*2jf7@|Y2LUNNTtw*D
z&@6S(&gl3ZAel${3sA(LU7YK$WLBU20Cq8IM75`aZH`&Xo#b7T+sx>{Hr7g}cDmFm
z9y#t9jUZ(4Z%*R+bK%Hqk5Zqo%c3WyCG@1I{IaWasqe=FiB%-$r6nHz%6W%tJIK$m
z#|*2aBUOhZ+LJ{iWqxEkw%WKOP{F0iBkX0}rVGzH>Tq2Sk92a)H3aAS-k`h0m-|=n
zl*GR2{pwzCm8s0K%0paU7nWl*<Wh)!MCLcERq5`F`^AJ0;!zF@DtmXFvi5JjtYAa)
z_bOfNfYd6+uW*!=X7!W>@Dsb7?Ih_o7I>QyfwvlIvKB^9s#y<ln%*ozjB%HLC5Q5S
zFima1A9_R41_$8ADOIHJ-2qIk=;k>9KJjVM86OYhLjL9mkP-EJj>jkPZ2(PRUh^v2
z4&VSoff;&a2Nd?rr;$IMZ+244t)hjNJAHmUlWjHPs+a%<rL9=0oB9}GZa#jx=EO2E
zsOjtvUF8X}<;a>@F{Rj2@6lmJ10z%1pJ?mwXz)@A@KUxFUP!P9_g@fW!|iAM{c!6h
zZLOM<hN+JOOpBQEJ2duk9*N}aznbQ9Tl`2K=2y3fNx94irxsE&RoFPz>oQ*<BYx38
zmR}<QgWS&)u^YZU6^wj&y3$PnvJ*^B@O(!@gFTWmi@gT4*Fa}xvx==Pib~<K0YbJl
zc4l@Wn5*T!Lp3p`g9|qo@*01W@g$hwUspJe*s;V{jQbP~lrcrL0e?oA`(pL`PDFEX
zY{<Rq%|36P^b?CO&7|IL8p-N#s$C&O=ia*_1-8W*?KAz0@SzeYxdUd+?lGNgfC%%2
z3b2QfZxOAr(nvG$>EVJ3v1O(W5Hn1vaq7GQde&23CNlu!wej=y$Sp2$VY+p{orAo<
z6wgJQH$5T|wwK$$Yn7WP!?uDJ8FEf0aC)EpkCW{Wr7C!zBC1;{7->8jHa)BFx4(iI
zD8eW5smo(_NE-ER-O(%>VQ3O2y&fYk<v-d7f|yiV{XSlboIRg!P=gLEwc?Y2^WIC^
zq_2ytA<;Uqss(SXzi#CPqeAmL(ONNmbQubVAW%|~H5r{Mw_~+JjxdO{GmD6LCS_~i
z>s;?FEx~d+7FFlltQ{(!AXJkHYi5=CpPTakw&PnNe}`CAW#V5EKmC=5np=Py5JaP-
zLuKApnm|@RwEiJrsTdiE$mdV==JhO5b|h}n=j(IHe3`zCw)rEvbCI(nbP`GEuV&6v
zTfO^n*I@SoBoim$m3&8NEIoFJ`t-P!ry6>sna-#%Vza}lif4IAEjMs|_VNSIDYvwD
zDI273&+3HM?>8Os=cc4GIm%=^I>uhKby&E{5M_-bXAfV<?}@Z@xwxZ@?e+jP1rA8P
z%P<b{ojZ~Cqeqj8rtaWG^&8LHxS0zOSY8J1`bKcf<Wt?97OOuxOSF?)s)A1Ez^#X@
z^5-;6kwHSH*fS&E8dT<PPx0g!V3cOXKc~~u_GPc>Dn8lS$s_T_m8QC(``lU25Foo6
zZ$962w4D$wVsbB$%jGwvGZ(Zn?J{04J=jQ~;hY1y7lNLU=j1QB^y5-lqC3h6E1fzV
z9gXl}$?ke90;g+Ne`=UhG3$0P1yJj&8?6=+5?7bC_8G;AdTnyNukiNT_SqV&#!=t}
zg3}U?%;Pd=F6!_r9+#F?#@*48AQ#M5s(1`ju_Xv{l&~qr7B`4t&(V;5emOa2h`A_G
z{u;1@q38i__%6GerS7#@?7LVWhTUMLboeF9ww7IXKcU;8gKA(c?GZ_<L|f&t5`TRY
zayn#dtc#f2!-EnZO%Sc+p@e@KnMC@0RC~IMgCqrL&dLJ#OgI0Aq}h7Ejg=G&z?lnB
zZ8;;(^wIaqy=w5R(K_zWp#YB{<A4j1;4gD|v&>nvlI1idhjy9$7Lu}QX3=;Vm=Gh;
z&y!727o8fc+3xHFF;z(^+2z3C%tX<Z!A>rmi%Umts1}%V4cLunyw5d`m}%e<7;H<g
z`(xB`(Jn8JE{tdDGMWHFm-$<GhF!EEaGlWB9URZOq1e71BRT=TK2-0h0>HUeu-NP8
z&pYOFfR~B5`@mc*i@m7)&zWW)Mx)X$y>)7wy6O0^;^=rQ&$3hI%AEXbv~r>EjzCCu
zZqjc|ReaCnDx#&o<5*aB40GfR2c`gyG)#<ls*K>WdrKJ_EGZk%ABzLdvvW=BxJC$p
ziMAX=<lR`z=_#b8;OqARYqUI+Ym+%CKl!C&YV7p)meL&tD5#DkK3!R}t`FNXDU*8K
zFMw5<xhugRjy4YgmIUS8mAURe)$@6Fscg7HC$Kph`@owZy%@3^N?zDTcnZ=V2U(*x
zITl9L*hPr&g|W^CEq9FrI1rMG;W1-zxkGmla}?la<JPTeE#<9WR!V=k&g69O?>trl
z!0=xqE}Jr<zjHJ6_z2hWSJfc8z2El$CWI;-3de(x?7%o2UM%fM4A6Uairrt<5JYHz
zrb+qvC!HdfQnFJ(#J>?@+)iQi|9xkRRfIXiO8c4ellZ-kpcDc?)H}d-zhYu&uCz9E
zML$u8(iigjzI4$<+c+p(;I|i{QeGu2YhX_~YEbnp`kie>FrObj>Im<%3zK&zYVAY=
zT*CdyA0b4l!w6rKGgsz4Zr3osuh9wEE$U6-v>#<W$$VFDN}6-RM+-ZqVvN`A1RM_X
z_+b|M^h4!QM5Uc|%2U_K>UPGGzv;Og7h}&#lLcTK6#Uytd4G+Q1E7DOOik`Iw|H=Z
z&&}0+Y>}+$)DWV?ugF7b04N>Jtrj&aut&REK2f*#bXS&aXlP0GFY6Im9gNDIMQtK=
z241E6O<qP%0<rzHYVzkUUr_yPD@HZBbNWx5_5{vxUoL8Uoi66yQ5M+EOgn9e1W;X)
zd#p96omAJ}#>-Lfvzp;7>U+xk6<G#!v?5|ieuM_d3xfJYZ{F;;k@`K+P>fGF$1g#Q
zS~>ZNLY%IM?|-E3-|XAK8ZTw3Ip6hh_uU%?XneB^lfbCzFyymKPyiiAUFXqFa2t55
zU4JGZD4bM$gBoDgAq~JXQFE2q&RBllx?}Z?z+ZZe{M<8Ky~I(r9)1RKhx1i>+@#CU
ztM{7@dJA<#>@EgN&Ig(75j@w)smpq999oZX0E3@S>Ef~>pO$Ok<r-_3K#7$88t$qb
z?d>2n13nZj&K8AQ9JCzGGaz>#GK-O#O&@|J%enO-)mrCM2bn?)aQY5(if@hW@yQ)~
zr+s}um;tg;fJd?)h{e3T6{prHR`k;;NFIy{QE|zT!m%*!qhT7Ezb~u?ZdA!Mr{-`q
zMKQ&>W@@cSE`h=*vs*Ph{{X;_QO5G0N2hN^I4gz+EOCP>-hxS>3@FSfrF8o(F`K1v
znOXg86s>DMH8WJ%&R6Pf)ynaUs2RVayL{lj4xt0lDdf+2w~~LCzQL9FT(Ra;GLRX?
zML1RWWYF%m{fXGHBu!4YW6xH4ueZ}COc!#GCwMl`e1XbBBa(YKyc{;1syQ@U<t}lp
zoVQkSauufv7mJ{yDoPN@g>e$=u6viCQf`xz^l-eXzemg1a||x(l~8%LL;8}*$9F4n
znW%6N#tD~2hU`s3=*WRReFA$)KpK)YQ5jy<r{m@3ayoBX2w1Ocd0hfa7mvOgX(RCL
zS{9-2Zf!kvc<q(_awp&U6wBy64D%!Nl<D1JvaWvhET=sEGTcL9%$1`m^-%O#UkzTZ
zSzh@AK66lP`VSH#>xUl<&#N)`?R6NHYRitUR?e*(Y8NS>-m(Ue@=c%VW*o|cw~nRW
z_>3>Y<J$~~$@hLtCrx5n%k}IMH8;67&+KrsUC|1Ave@Nz<>5pFQyWSGf|PS3A61n1
zMA=M#qq#dZ^!@(N`{?Y|W7XRh9UoWvzuj?V3QZA{54L0-K=BLNE9eg$F>Q(=W|t;F
z0az`Vz5OO+{PSVL=;F`BHN1v{M9RCA9+Ll_)n|vds@lP~8btOdfT~6sUYs(rNLYMg
z_QPL-d0jq8nNRuPBON%?@FpA20TdLer_a{vcE(9rw&JD})HV<QO}_JooQhc=t$U%4
zm8>$$`Ch?hp_4IfLeY#(450coegu1gdx}gO>K)Fh)IQ_t!880(Bo*!HVf+5_aWRl4
zo@)HrU~Q)IuPc)1!ZV&Ii^ZSjEBK`N5EzR9vmWV!dFoB7V)9y4P~<X~yV}&}RIgMJ
z8>hzS|5ot*&5M}o@nePMrDxZaRs(k}0z=38emS4MFGvF&PC?t&vYZ=C57N!j6hM4&
zN2l_Vd%q4=2f!2L#mHpwQ#!X1wX0=H1~WLH-~PD@o0rp6OIWMf>zpr<TA%>r;M#<b
z+{$sH4Rt8b+SyKkMXy6pOy?Pdt;pK7speGGaR_8|ESB9O?({fLa=lTnA+f4v*Oqh_
zTbXq9mOPWInhmNI;?VZdxKT+pdIPlRsxe-o;FD)<3y<4uJWZdfxX;U_iIZ?H74L7S
zV=dPuNzW&LY~|b0_LNZT;Tx>%26TdlcR-lzrT5l6ToU6vBITn8y45Qtf}c6jPJgHU
zxGOqYc#V)8;B&gdd&#{Z)2H|x{Gh_<q#nN0dS<FKT1!k5c)(QjTIpbUtb;3@&}OKC
z8zgT~5WnrDEQU@^h3s`^WPcYX!*$Ow$eqMNm4^rX<^4x|R&DD<a=5A4qEJmKkEU?r
z6o@_+TyXXvi}YtJ1cG!62%u|99m!w66)V4%Pd42-C<T=n`HxQX-lEEFr&g8mJ=lrh
zHr8lE#40M|xO8B}FsbcwzMVCwpUugZg~~DrbiF`{j}2cL#;^~vw|0YEcS=@DKFBaO
zP6N+K4j}#@m(ktXjeyT5MVYSktxi{5@-TJyGxKTP#R=&LnqB39*8ywRFDWaNzfN$W
ze<6R-Hg4w?DOjNp<8zMghzzPuJ4mf5ZR|tEAn9^$jSNwX(P&M9AgnaAOP=ev2A*>A
znH}D!+5t0*%r^5FlGq5rVun!<ok4{<*i7oTf={5XXVl8IP~aFujGcUCmL|;L6ma$P
z1@{G=8*D$`4YZd4Q02geYhM-$b;xW&^}dcxN13@E;JyTMPMDH}`nw{@rFdSLwi$O@
z)UK|TN77}KD1XfRuk=9>W0BMtw?1iYk6$y)Ln%b`E=&bGXXs$3`4fRnB&{>q9NFy4
zthn&?=2Xwgj(`rJ(;&vzS(T3YxqA+4GoUa&z*Axl(i6aPP;6krNdtg%-AV5YE`%lV
zvIcz*km<h5Z@tt(%1kzA(ixZ6yA0IWQ1r9S)YHIF*)<&TZnn47*PFM3J9Dq5Xp9X9
zCbk(-M%5zO4Jaibb>D88sjsS~k>A@0ikw;~h(V;1Pv;MNJ7pR*u>GLE;C&xPrOT)b
zYpXcg>XG5XIGXU@aZ=v(+EiHba#U^>9JvV+Ax-K7i@ZdTq`GCsE`n@~A0K~X*MVGQ
zjvXSlYhBm>4%OMFE51Xg1Nu9<et~G6c-j88yNN943)#32*@hc#d8{<6o;j>p@o?>2
zjqxG-{>NTogQqHP^guR0d#!F)W1H>!R({q;>lISEl0^SXH)jrM%R+fSfCIYRm+fy2
zt~crPNs|JhI>hE#_N9{*y7aMyF#viB_yfZJMq=~!?XvqJW;MUSV*KvtK&8LUqbXW^
zouj75{xkuCx@_Dc6dQLw*S6A35>_nY&)vM_|A8pxdSlAxD&MmUgUd(&bT$-OH2qDT
zpVuMg=!`~znD2{i#gk8_L*wZ(S|$YbyW*rPSAHt%j4MWn1Iy7#jq`*OtnxYXCm`64
zX7JvsAqH^+$mPqq*T(^bE)jjU@XhW?si)PozfRA{ddGoMssx=r31H6-P=)|ff9y?R
zwRk4IPOdEc`kInqVU{BR4r*dpOoxo;^_&(igFK#fmZVJc7?a&SoaS9ZK`&7`sQc!L
zDjYY2!o22G06M03Bpr-y{{#{Y%FnNy$l0OT@}Ke#n=++=o@c<36{xe(d78j0PXK1D
z{z9x)+Y4=%W<Z<P6}G9%S1mIHG9L+-k2ZhG-#pXD@m8t@1mluxZ~#lDi@p<2JbW?<
z!w_P$1Yoptxdm2=LOo^`3!h6%x?KJ9(@Wv2`l91Pb7EOG)Y|NZBHW+5O$gD9%O_7=
z(TmW#8m7;DTB-mE=<bTa{v2M@A_ffT@ZWHdWWbwpe)>!%r3mHD-7+2hvlpPFAX}Gp
zsE>Qme<3Vn68XhjY`mFM#_{%_dVPkw-CvduSC=ZFJS?MeDkh&s?X<M^wJm(S9Dw93
z{@wAVE^7XL5(ZR`3a?zM)^yK|?Le&l&J!rk8~w~jWO1$qBeGf>1B``cMVs9Dov`9{
z^i~Zr>nG9|#H5Q#6?enH!Hv$Dl8q&qrXb{^^ZYLe`<Po^M6o(`1**k5fLNb8DyRiu
zod>xH@vbWb>3n7CQ*fZwiQyvyneKl}ZU*(#ur@J?5C(({|3#Xys|5;}=|#LG-JgcH
zEIHmg-um^XSxLsSQ!aM~Y11S3h}wW{rMjNhB4=eXwe{GJzlw&suWE1HeX?Jm9w_`3
zkT@45Aw(;@#6Q@y5rPykf%h~_RaGFAnuaF40v3x+wN@iky=i>O1=)4TAtOE3Xo0By
z_onxeHJyqpoFExc#0CavdF-zb0;oZA<??*{I{=s0KRh!EY2ygXlT%m3Jdao%UJ<;i
z4$#%qu9jI*4|gD0x3GsS0CEQa6*+tJThAc4Ub*y`r`q4DmbLjMG{94j!s47pZYMRg
z!q$$quC~$E;|b+yJ@r;r>m$wpCI~n@ds|)}9g)s-`rwlo^sFlcCcoqjDUT&DLU>E3
z+7&)^+H=;WQ=hJ)jX?cMPf-7X;NL<qmh=kGu{bMGv0t(i^=K8ygp6UO%k<mY8-$s(
zI#P0KNLi*XF@|LamtCKYDJ?mp+%&&p=t2<OoEF1N7n{|bC^7kHlkg<0XZcjg?`Vrd
zC6;S|(V^(F@W5RaZ&u5t*e_)Xund1fv{3-^z$Z;V(m6zBfTV?h8;0^>AThwZrf4P_
zN&-B)H+8;M@&2Kc1^4Htn9Dqf8WrkBk3p33esgv>&qy{<M9PAKOV_u17Xz@JS<$gV
zlE^c@Tr319;(s6gTVbJu%ADp>3%a^=%ngJPbD!m;@fg8Aj^BvD#Gt;(azc=9O~aiQ
z_@Zrq&?N5T6gc>0U>kpLQuWW7z&WEr79D5A-5ghWYF@692h`lpMu#5YFn%miJT|zN
z)MmoTHLwKmWbDn0@X@Axd7sHIA?tXkq>H>6y={}&Of6jp=^Asr*q<hV&YY2yGYyWg
z%YDFqf@BIa--U;h^CJ9Hm=RO2;ncdJUgT2fm{3cH&&X3!5PR;BcDGoL;UoaHE*Nv7
z4RM2qR~+AjZ=_>azeCXdd(5@f6U_KTec?@mUP(}-*kGZEuO*xp-4@xcqJzSk9?`0b
zd2s_r(bmc1RQL+Ez5$|QgMyVw?Qi~!w9|#h_&Qw|rZc|e;85d(<%H%So*a(C4?qY2
z=rO=qy%r0J>*WV)=}Ke3AtFWxH_U(sViutY*n^(yI;df7#3+fftH6<<6Ht%*$~3(&
z>jeiKQ5ET6V9dMkb+X#HESu&5&dQP`=f1_QLPU892s>G2O|EXmot+>uX(?gQ=cbJ_
zIT-``s(B}zDiLqzGMEYMS)i<3O;!-!eM>E$1atJq^2VXPP!bLm0;FND*|*x?I1j-;
zTl7?rNPuw$0CXF6Vd(iXXIl9d0RD2POx$%=(-65w806#dsK*_kqLVBaamis#jSvNO
zrj5AV{req`n6z|rOoYjPD5Dj1s+8%Y^EF<DTGc@RkF$8Oqs4l6u9n<JBPcWErb`zG
z>jT5G{?Y^oBA~zog&5UXs8HkFFpqQ^=Nx<Ll;zo05MM)dU;PHk$3LT_^GshG@4zaS
z_TvLi%$2418^{l5^F!;#6Q9KOf#T~gYlnlWPet`Nt@9=?0d>FvfE%hDALJdxVy=OD
z=Y|$~xLs7GqY#R_0Q-0<KDJ}o&8YDp8^nN!bEJsGI@?<y^1*4fa;uXgd^bOQjr|MC
zIUapn8Xz>qwtcH3$@C-D8g@RXTj{5B`bjI&451n~vNz_(fcM7xbbb-@vkMSk=JH%*
zsF2OGp+d0)2%;_+;|EapbXK!VfF91_Xv!GD+<U9Wt3J7g3yKSpMqt)0<G^sF`f702
z3l&gtz3Wjy?M{m>&~s554w_q~ys_Kwxxro^Hf}bgH<6a;xappk{taqz;7~igR=tnG
zDTFJf=Z|Ma$}F;XA1mrqFfzcormDYT&G&~G9mAZ3x;#cC#zIQQl?STBu9^|!WkH|z
ztvX%jyZwIpZ_4k&<X)F`i*QQkitEK5m;4J(TX+XN3~2G@_FmRfWHJHmD3%=z3&42~
zV&_W0cIDINhu;}IDCP=#G31<9L&q8@%S9O}O0vR~SU3^9h|6u|5%!H+&Byd3Ljd=?
z3EI4%cGsu*@5bIq%gIxwFw%I({EmITp-ebV#q(F4&l4<lGPzIJV3jY<0n`z=iz&j{
z7$_BH5aKFzh8b~S17kqPIcelZ4!Ly;TBZLwjMB=ok3&ApfD78z3Z<C<FluCUk8UNz
zrUOa&!ag%=QAplc)$GGHApE)l(n0^O{OTppa1nB(2_-Rj+N&7k7FWp$5Tq(vi^l=f
z1*^>Gv>3n`D9lxL_oh>RbxTfwmUo7K4=U<8Y!mA+P$)|w)pVq4RZeAO?q@#b)_5Tm
z{j-LfHuv7d3q$E`+Q3|GM&6_>r>G>V@ID#<5z_hRf=;$^&KCX^j56kHRiFvBl%_CX
zEVxhGHOXfxr~k0wL`s3YmUsj{QXiQG1PYiyxiP#GQ7kL{@MGD47NX10PgCY1P(747
z=W-6o$ui30AwB}3r(!ixD-O3D4UVO3E=1_C%Y3$-?GSNP!G~r<h*y@DkqG0g&RC=k
z<1pS~d&od`Z6r+EW9D%+BR{2xj5JPNspG=Uy|3f;o}Bbq0W%IdLj8PTJV^BcIl<&r
zPlA|>)i*3F1Udy6JS!rig*=mBT~1JY@0qK>u<}w!i+_3nU4m*oOU@J6@A^5mnm;dD
zzclR!6~{9ykk38*LeG1gK>8VLR0I%s@0&vK*~+0qtc%G(QD&r9VuauPetNGb_k=0w
z>~HBoL0u$y>9g0lt3VdE@+ri9Zm>irRa50xq56i0GWWnITq0M;nQ&5MlU(eJA5h0~
zcw?rMBlmK4ziOEf>8}RMg`+@%OJ6zp9hUmb6kF<8^%kI<ZvA}y)F1hrHd)ynCEN^5
z_macCeuv(7&v=eH-g5`-K#KPQAU`tJZyyE~(yIj-MJ|+oBrQ}2&B(9RlAf}9Q)ySt
zZF<u|^O6WC2hBWG0W&zRL37^l^3EToMCy|pr>ek|4t(-MSTKC-&4)+wX<$%}yC<Fc
z`3rw?>D7ENvBbMHtb-I6*K21;suMZs)<NIHE5h=29ay7WaAsX0ck3984(f67xmg_e
z1@w*Q$^t+hJsuf}YY`Vu1T4&mJy5yTj?BGk8WDqGt6*Qjf@I@h!#ZZGd_+5$0#F);
z#Fg@I1T{aA@bw@59A1+`gi(KX4$~=+$Dc4=<G~6#l~_4r$w!_G`${<ULt@N@n)#$s
z9jRD93<ebMnUbT@rCRQTu^SsP+~V@59#(b#N3qSIomPLS^G*3hU*lrX^p@lgwZASD
zH$u^(fBhl0r=phD=9{2!;eVJJF?76$8EbF#26L7@&ZG5db)FG!D(rX$Z34aaGnR>l
zBJ>^P(HB-n-~$>~X9v&b5v!a6wY&vp+#`EX?Ckomwm5jcMP0B;VWeUGu+Lg^pjqWl
zMEXskX)v`C%ET;#!AL2KCNN4IO7L1d$7Dw5=?cak&iH|T>93%BUPJuGS-0U;gV^dh
z{puSZ4Kn<uu{vNL=*ZcdZgYe1^sa@8&LkX-(jjbMdi<g<T_Yi(PXA@R=%a;<FJQp*
zzAy;ke~o!>1`ti+%a8dQANX!`Z?B{!)ZKsF{fPZT@p4k;$DRleEb<s#IbytZ2#bm$
zV)04^V9pSXpvqh(IO!xCrhwzNS{Mlm365P1;mBp;4lQ$O9Vem|Rz_~XM!rG4eL`-~
zSV`SwPpEa5CF^z2{ruaXdQjv4+lkUfYk&&1B~txCE8K@G1?o_l@l_9ID5gzeGYhnQ
zpYc;<@jtIBe&hWfk*6)SP-`qUr)nCz$Ww(yk>}czl#ZrwNPV&kWEY0RySC@Y?t&O_
zI>YK0z=I^sJvsDiV~Q7=F9U4N<Q@%lb%%)@dd&`XN)l??rV~8{`;LW8&-HSxVBZNp
zZBgI_j^%$h!)iI4td9`~GbOzx)|#c+UnNvxK#yFnXTe$PPC3wG*uYL-|L54Z#D_VQ
zHcK6KfyLi@U75af{*ah?Y-LJ}ABUM1JDY>SsL<06OeVpOJ<k2l(=CN&fe18HNY-{x
z2Qiq70reIQ_ZfyaP#MA;Q@L2tM*|(Zp(3!?(h}KJ&jU#WokXvVc<~BOhm~q!{T_`F
zwPYavfHk}!g6fX+kHO@YgHh!^lMqMptjK1FZb2ldsw+f-e)J;HB*ym8$ECN+n=@wJ
z$M>jaDe`6|$aD!~fiWJ$=Z~%(`%T>cTVgfTycni&IlVXNev?2argvJkVo17(PUe1<
z?}uvx{a{EniukM@Oo3?tz^ET;5i9Wnn5&#)gD*(S4|$^=&1o8YKJ1kQ@DGXU)EKn_
zWirh)V!HTK>l1FR@IIKYsII8<==e1@jg<hxXQ?yVV7)bvn++Po2j}Ald}f<NlB)}J
zfd%iso8+gTQ??0?6dWPX`pS%te1Yf9R>|VeUB}gK2AX|bXHat!IUG{`*Xp24Hv)ac
zKH3jI0s5<1P$mQ*f}+0i^Ut-;ErS8a=}t_ORiZ|LoeCJRgK68uu=~82YAtHX0pUz9
zwb#*W5RFCp(;}2v=(mP$OWjm{<X>@{YtQ*Kc6MikSU~l!%?u~-I52IPHn4w*tN~p~
z0_fF}g<Nv0Im~BDCsDODKNL*t)kRc+IcDD%ckWRx6iVYyP+aSV4mf#nVcsj7QNup=
zKsL8e_shbGDx)ID?9W9$Oqvse7*Cv8%0j_1G#;JC1&-xG*`h;!M)85d5vc!p#=XFp
z7+(ZSkoUpV1#^c@WL;9y+;kQ&Gm?kib?|%ns*XLfLg?NaRlmAIfAhx2WN3Oqm=aM3
z`rG(p1&oIF6ayVnalD)GOS(X7)NUHMAei=<gF*7&JjMr}zV_*IJz~#@1p`PRcF36U
zA{y@_CCH)##^;m+rIQdF8=P_x0=;)ZU;qO%x+i>I%(fQo0~v}3=p!^IXa4-X_h?MX
zf_#msRt(i`q4?B@fNxKhK}UgI&p8P}E#L!9;2*zA?1d)NKbHW#c~=Vd2O8G!7G;wz
z5kO-yVHGR=cG7MWuSXu<SdVH8U%dh*o?!~Hf)_WB4a7S!K}~&0^qHf-_lEX<vNIz*
z)F)FO8Rey`f>-^^sI$g#dqHS#?S(sg^Jc&pxq}nCzP$Asx#_I@(Dp1{er_MPjQd7l
zDi7u>-U%3}i>z?)z;VacROt8I_?V3ZR@pqa4NS3!ZAK{P>{$)h(NuxvnZYpo&U1x7
z{$hMDsX7-h9tWP&odpso;QL<&q;<a*K7?k+9g;t;c97DwAWmEATm?!!37&C-<`GS`
zz$D{BaX;8ki2UX%G+o^jJZoA!GP3*dyN|0nZ_?(_0~e3o?<E82je9nYMI5h6++0cx
z1A77<d*_sXL$&;Mw$<Fe8B4$Svd*v%;4h$8J!_aMRmJs^zybTx(Qk3WWA^o5XZsuK
zFWI~H$I$BvbMpJrNf$>Dy6V9s#Krl*sJ7h3c-@DCg%OaT*53naaU1Bk0f9xK3w30M
zVUf~QrnthIxUO3OR0_<o|H}~eoqYq1z^lM4(>>=?d*{$b&Y=0x-#d3I=)U|%FciUg
zMbcBd!1Z{QpnlFwe}UbPnB$W_<$NM{C{x>^OiuXeD`T)Wn7TCq>S~z3fG$?s&|CJ|
z-MT1b>qpFhMmU_vlb^xFNyxUo`Zu%MJIB~O1DXS&Q4mPZ91S-XwJYJp2gF@HnTEj>
zt{a#%zR-G5bXX)r?JvbqC;=0oNZ?dlhL$9?fGUFn19pEu8a=*(mzw^SzBLXHgPqR=
z!Y92xmFTzrYB8k_9##Zn35L#AN>C-0McE-{DSGEpfpEQvA?-?NJs;PTRc6#FeQ9v7
zgPXP0MEhJ?f2EY%&V>b|i$JqkmN{oNLKgW0>B_C!5`G0s3#+s>IyZlxJ#9mA%h(#F
zhrt>x!H6geIE(|sl1^T6YT%-S`(w?~bzBJbJBbcLoy1D?SWUXKP#1Jf!uhahL7Hf%
z6U3`baGL^9ez~P2R$jXPZYks=ay2i1>b8l7z2$;pi0Z(^U*(f7rlX?1&&2Gy-6ntV
z+O!vTfVtyWVmvB(c8rh`Dl|aBAxQbp_(ER}9u@a}mZTI|IBR<kx{hKY(cKfw=_Y3s
z_(AgYb|eK1+;$Q{bfI!OjNOxE7)&L{L05HvYYx1wBKX}L&4P`1xTS!~!c?ch0FV+j
z0gR?{m#y%|)(6QjIDZ70P6&GBtV+nkOuIS{=n9Z@Qz0-e4EC3F7<<NB2^M$=jQN~}
zvkni>*YUJ%K5LJc{v0d5`aw4<(gIS@@OEI>Gs<jhww!E36otY0@F~freV?QX<fn|R
zy=C+=tWH4}M<g$UWV-qsYkk!DJ(9hp?$c(JZ)R>&ofn>W=E1>@B9K5fb}syNKgB)p
z9dLmNhjrZTiV!8_lyhg|Y19%M+ot%xUgzQbCH#VpB7Kz1vk)+uZ{UNdh7yK@ovwe~
zEinOB0prpj6jEDunE+QaOncnvQ9dOfui9Y``8RO!#!EWmkKk$zaD1m1b3RywTnBOn
zUBM9M^8?-_h=E!=q04FJ_d9<9quaUIm3~KDsryvtWUobt#?%TGGH-8a7{xk}6MoJf
zT9$R&zVhPR>8cj+>OyeP-`^`kw?iaSM);XZHd0Nb-5w!zSEIDceaPR1t*I@H9s3hM
z+x)P;bG$zQI#-zCt-pt|lVkGVZIk*M^c{{^2}lY$Y`+1**uViOI70C0e}4EsGC|Sw
z?&i7N<KT9acMz-oUv4q_pWl1*pC|r5TpslwCjN(s{||SM{YQNMBR>BTA1LJcA70w`
opC|sKIRAHw^8`<wqcM2Y|9Jh!A@MgL$bzZd(7c|bc>n4D0-n1;LjV8(

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide7.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide7.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..93e61c87f489d494bafe9d2d77897306a21d1be7
GIT binary patch
literal 32317
zcmeFZXH=9~vo_p_f=UulK;j4@l0kA-Nh1hI&RKHKIf;URfPo}ANR|vurcrXvS%T!8
zbLw~BjdPy!eCM3?y?@`e&W~9=vzXrZy=zxpb=6h1JLr|HB+hNp+b|dm=f&S*3NY9W
z@UPc;Z(ai*PpPEPzz3Rxg5+~pemD6F_>b!*&t#s#U`63r=lan9ddv22bq5#>w*mSG
zt;HtmH4Ns}|3d7UlB>?fB$k@;Qo=IQ>_cKVCr73UzmUo0MyF4(%j^yRU-GYBSjNG#
z#jpRW$H2e-?2bXB8hf)Qr^Jm1Q#T0N&0htXzWTsZ(|o6iWG?SiGLOK&@$+Me#`%3@
z|3%aT)~vrgN=C-l&fL7i!<p<#hQ~Mdj=MC@w2|AY;H_XVY+o1#^ht6J4IBDoyMB!X
z`qaC5oelb|zjIR$`t*Hprylx*iT>paZ59^r6ow9cV*CB?0{?4+|Mi3ajfMZkga2nS
zkuAP@h-ZEaMsf?41=1>%?RH9V_JUj7wqn`Dtgq=kL}df=Uz4k3#PFr{ulpU>z9`Bo
z9ibOZHrdowd1_p;!x~VTVJ2pE)6>fh+1%jJMeTysP)#1RW~cvrD^4*9U-9nH99VY#
z2#pOJm1e>+X6F*p$J*X4o-)>Aw(JYWjps7dCeeB^E5^>$2-P9ZtziMrAjhhIrj&59
zEhc*kU!F?!aB1O9B1P^Ww6TO%m+P~nAK*{Z)ahQzz?xB6>FG3S6r4atlYDZ|HPxZ1
zqszr+E*sfxM7SUgW5Yl_x$3q0xcNwkzxWZ6xRVI!^q!Q4Zn{0ylhiNe;zyV8e3d$K
z7&>Z0$#hZadAQx!_>;LzRzo@2u4CdgKZj^E7uK$^!BA_R{$+6RTUxWs*|~+6O8X9@
zE`=`|DqHrJ=5SC^>T(y+LVs<P4vNpO-YYwQ0VAhC*_pWt*m<DlYo7Lajxx>jDz+Mn
zlg+YS$!oTbD0_3xcZcR$sgLKBDH2TV&2-*5DM91qM?Le<BriQrL%>AQBPi){j^-vN
zEKmYvC+oVr+%hIBS_Y*L%29Chq4!C&?q{CA2`WWl<fN#l3KUNEtK=on*a<n;@~UkT
z{rP6{g<2GRoN`*%u^<$DaYHNS71eKDZDD2?eBU{0DCz4txL9im=~jH3g#Lb4{NUKK
zL9ns_Mox@cS2MeM&k4N$2BK8iB9rrnn}?@Q<-k9GjpNm=5OXYuIn)K)aiGr&AMF9g
zY1hV_Nk%b@(MIB=i1#^6Nu!e4GFGw|n4hC%BC2ugpW)X7Xe|t~Su-zhsic&1R8dfj
zqd9cYbhIhzvEM@DrAFBUZck0;b(TdbWp3`kpyN!>#-}Y*th~t&lN3XlMa5d@F`=>>
zy$qGhv#l^Y;i|t>{`JRz;4hwG%|}XLWe4-~fp?3WPA+v(;Nibf$5@!c5%(k{)!o$X
z)^b3l@L2lQIoXe(@sgt+&h3~xYv9-RA}$<h#ZOx#gx8Jz>0jT@`%KSD9z4>b(F>iH
z7Q$gJ=a&;K>>b3Le66%c6BS9O#ED&=g(TdqcCWJy`cHqH%&U{vv*%Cfmj*WJ{!~F(
ztwYY_Lg8!p;OaaAmtv(24vL8O-`%Pe&-ZaUDFt7~eG68ux6W7A=670Ao6G*5|Ir77
zH>%@|Cpf1@cuX8Jue?%t4|nGm@REMe7E6<gMJT$**hN;llJyHJgkLK}zwJ~<&YsqG
z;CgE?As;0>Ob-tc#a1W1>~B#}OTX1UF?ZGh&h9C6c9yxPhtY9(^Rk*8!vy6i1#VwU
zyjtdQ!*n^sIx7x(GAu9Cjhbyv(?<zP`bErlU<01418eAwM{!HII`EA$3+3qf4zw!e
z*n}DZ2`V&C`aC(mw(Ql;K>Lv<<fq)+ql>Qobm`2C$#_18#wh9NV+36?85AFI>V6)R
zmh<El%4ZycKRG-lSdT2?tC_!kb4eofVRQ^DfBbu%NAhm8_4i5-={o*o9$ilb>YUxx
zZ*@{u<_^JW9-BHb#oiaE$H&LI=d;AfZ7(^YgTZj0^EtXCPX+G_JY;XCTvAq67M{rY
zmN7>1Gw7kiBOmK%jNQdp_s9@0m?ws1e~!Z(qPE`OPfRJnT&JI4@WD=I1je!scMwX<
z@(-!k?3z5i5JA4CpTMQySM%Qcn?k_(F5>chNhFEG1ULE}niUjWg&9?6M`EX8F`M{;
z9Rmz26RwT9;>?t6=gQ?l9SxQAr0$M?9Nfy>eD{rj$9YS*Sc~vUBi^M-S$KoOPsB-=
zO{?Bh3+O<;XMh7?WyZpa)lbN+!;5Wn>IvxDyA``x<h}KzZ(`R3-Q9NYxw>E5ih>3e
ztw0%@Kw9<D`3jzRh47Ali`;Z`qUY`l`JV$BGSwbO`GVfCLkL^+D3%Jm%@YN9_qkZ{
zv!sg04S>PP*c&+X`?lQd3HH-n<6X?&EN;F@6xv^s6W&3$NVhSR6WnR+@wwasn{J)M
z)Ki5_s8X0|Y0>K#UU);e?L;(Ep=|PD?U~lH)>hx1k2t!hWVdBIQ*eK;cJBh=*=7Dx
zxKZACz=C2Uc%3hqFy9N5!zy=&ogDNqwjU~AE2EIuhnvoPD#47>6`?!A=!I+EBgT3u
zOZ;eG<@q2B1f9qWyG&fG;|iPH#+H_j5?H+~WTx?lHmZl)s*?t)CyZv})-<G}oYT=5
z0bvm@-G2Waf5Cb!>F?D8#HD<rjlfVrZ;EK(^u?x&=R=Z08W8oqE#;Y(YlW&>hAQKE
z-pm}{1FQ|iz4g7JSQ8Gn8>*Gkj}(JTg)%)2rjh0?76=?*(l%dT7%531G3wintNlgE
zE1U8*xUcL7Sq#l)T1NtvrY{vXsgjjla&7L$#CK;LDx{W9F7CClP$GBI;DXCp%wp)7
z-@dgvpe1`juqYS?V&}OD-p!tE`R9XSU*bhLU89!vxo+7Lqu~e$j%X1~D6aH7$4!e|
z%pD>Qb6vM;kLD|3y<yPqHS&Oyq%9%M=ht-L9-~`xGUu@m8g%TKStcL5-0jT)ctI54
zQVRg%#ej26%}jUkvyIj4%)~0kmCu3{uyHH^`cHH7eqFk6GxZ#L&dPVtt#+Z&WO(T0
z&QcBcmIA=W>Y9`b_0h;y)a+UMqJ4Rglo)|yt_obi(ZOM^q5i1=4e)$ZioU341Cwdd
zC-e*1pY9`8qG^^+Yf72EZ(a&jw)j4ZG!C7hnEQS`Cy%acYkI<NXi0Jz9EOBM0Q43&
zpto4-)QWp}Td_@6MrO9kBoMI*KWG)O4@Z+Zo&E9|5S6Bm4Hx45pl>>d8po^8yiW&`
z?D0uV<se_odv%!E0CpX!DWqm5n-|Jg?JaJ~a!s@Yq|r&rC3nH13zln0;DlvF^>V@|
z^Q%L-3-{6VJW$_dV-aBUOvzwHYCKw*ToTKoIxkh*<Vk)(XWt)^HetrRyc^^tVE*0v
z9f<t>)gkFbTo_yuL{hIpQ#u)V&8Uh^^>pqsKI*|1#C?G4z2D-9+i$;=jqX2vTx1^N
zw)YO%uY;5lB)DbD13i$Li;qr25%1FL_~CaPlEdJ&K8~~bDfCz+!R~6#6rs&GhsWVm
zZNCO%=n`FLT@Yspz(STEKsHDg&$02LcYop>iRk~Tk+^+=uA>=EjR))fc9q${(+>FJ
z#x)l+9svYY+;xN%vO``<RP+le=>3KCs|Mw#HR^H3T*i5J*ti2K+pYY3J0(@D;bM$;
zQE9y81Lw7yeJe-<vkXE#Z{G=7V)S`0Fc>za2v?>=PkuqGsKZNOG7(arc8ZJ}9`>>R
zZrBngc!$K43ffwDM$1_eGlzZtR<Tt<TLX-9<>z7i&utdvm5T`TvRUl&CdFbdt{2z0
z3-uf3S73V5s4qe1Tvd%6d$MRb4WPai`kwVc=(R5cV}m+UWg2#uJ&F`Y7Un~8-n)cu
z{IF31;4bxPT|G&!HD=OF(^-sPZq9&&H~g!XHvD?;H^5nGi{S2mZ3B0GZDbxNNDTIa
z%tV82o4&BY0BJ1}KG>utjM3TfG1)Hq$c71f%iXIuq(5KfE8&aU*9QOgjg%5kN!BC9
zaRpXFZW}#-SG#}1WKQAv@ZMh+gU6YX&SmY(w7$8yxeH)F9H{;1siQGkAL!9C9X$E{
zRakvX<_jjesj}H{#Uy*TRJDfDK0$pni}vTGYrb6&3JU%_$xRA|$+{o$L0_TvHJJ~E
zzA*%-6aWwg9rz^*MB5t&xv!C!Vxpb_z;_LRUPR7dUrdQ-=6rLjKU(SPMAi6)U49Fp
zl|8=NqYOAc2u_Ioy1uQ^9s}|CSUEc_NJCJsKj*O<n{Plu>kHN}c$-G=6ZSHh#2YYp
zBXR%+{^tXaXmO_AwcBb2Vz3sb<1#^#rwxF3F#<LwwU#z-F|p)hamiPoAm9>Zg67MN
z!mu|tXL|Xjq?A`pQ=?Bp6OoMI0q}d_dn_`A;O!rHArFG6&Y!|F9$l=eo+sTk#k<<Y
z$y4iAOA1WePAR0k4R$2h^~6`1uj2AT5uR3OH77rPgi?NAUbOg}w?stTb-VrN=xIQP
z=HKPgO=oc8OpC%Lk2?{79ygg&p{8eF#g<a8P-qqZc{CCDTDE8)u5>g@BuF{`@uI+0
z7=R^i_3VGGd-RMFJK62Zt1m+8MOOFUcFM?9d0$k4gb9aO=k`s>p7jA2SP$9962m%R
zi#-<09(wF8g~%y6`cdxgf-0n(l{a<;hy&5U-3K?46Rr+5Te?LI7C!wgAYJFt>!_n)
znk{fdoLm2L>>qGDKmj^t2NQ>KRXhAK^$bxqd8MxX;2N)Y*Bx3WCw7ooo4TVfRtmv$
z@o;;%q8l}ke)U2=r=mDI*Ea{V2PUgs7Cxb0#kpiiqArqCJDu5y{gwr~7)&>}>LG{+
z|9|BG@V;Rftk*(Lad2=V199pPp%g;NT%xzy%Hw$Q@LNpeCAIH@K8%0IO4}->@+^Bt
zj`7CCVc)p89)3NO&y7+#2u~y@+~*U17gVk67}o1!)j-bI0NYRFwD`42VnVAsA<T&G
z^I$*`o!@bp&7_;$t^Nw&>(dhI>^(;r^9}bL57}qt&r^J27WujN-U%6LT5L{z#zbHK
z@gFX~tKDnhc{!cQ?_7u!*ONyePm+*rnLd8b5Qo`jOC@t~?!-%2I14npfO5{e>>U7d
z<nUgT<6P_6JB@x&5~d_cw)%&|nS#Rw#2q=GlsNb^v=^k%L_rPM21MUy{0<|mxOyna
zIU7X3*^EBl>a9$a(l_^fxmVlxe)_}ozO>!VA0@S7cqT+InGxQyln;{iQMn-xK~x%(
zw`dJg>NntKJ3BkOylAB;T4Ah}Tg3@rem<42gNM&2%e1v($iFEX91H2L!W@mWfNiV=
z@cCvkL)DY;_%#~bhbAqa*1-YIlQpkYfDcCuOSY-zzKoqli0Efp)PrP8&hMBd_&01L
z1ibOnGJu#M+sYJpFjAf3B+lt;yz+UbOia_Sdc81xD)N5qFILy6Ot;M$VA=)tKdI%%
zsmn9XN8h#(&Zb9XDFRNb^NWkkl0Il2P#JP;!{uTlqsie6><2<hDOJ8R=GSRR{fjN(
zOv{lOV1<sEu}!=-C;a9`OI<r&pH7M{O#c*h6K*3S@q<@Chk^;WP7p>h@Q^SHflCmu
zl1@%#f`YJ(Fz9{xeGU}&e*FMeq%0jT$!br&Z=FgOxj`DgY~<}Xpw*J+qcWEN5@dZ*
z-91u1kEdfvxPaf8w&ewf+<26UMd;4>+q~&3|Gdc)Ijir$)YjJ4f(OKftdKXlV~Q~2
z#%|)*^=6KF1dufoM3EwHGC8x_x+94Ks<KYEJ5S|qEN!&sSNq5%;y&e&%BZn@*d0LZ
z*9759&{MB|<D^?eD~#SUFJ62Kyk74O#l)qCR+ao+J?UWi`Bc-6_b;i8sxu_W|L(=+
zDavg>415XB4xd9Xv*1OtUkcX}R`lHfmu#u6Or>0>rDZjf0;Feqf9~*5wqmBw<&w{0
zkfg5!1PW=I2$!!})5aqE6YFZix5%qkJ#rZ>ObCWC-cOR%)(nL+5&3+=2*CJLFvg8b
zzfjTaY_HcnUlS1;`ymzmM*K8bHxP%+d8MD$jSaRQ3XDrKzKC8)O&n=*L&*Lohu5{t
zqmjkX-juzXH%qRM$uxO6*-BStIzN6^ixa$4u@WJ-9yfdD<2|@?t1n+paFetzfgcny
z65`?sLLMrV=rAiNVKh7KcM8iR`Z_ecib}n?zN>Fib46}AC$(Bm*PG?zX*A~2ilU49
zvusVPkxU(&z9yzsvEoURb5_599P%RpsqScrB|^qzKDL`MJEDJSt#xpt6e1KPN+$;n
zfLq3tP&&LY!T%ag+yf^o;xLGAp2t;zmphy4naWHi6#OBaZG+A+CE2wmco%L@qkUF(
z2B<Wim#fP;xj*;p8GIlzaewl{)U#*yAywrXj57dQUjAT%Rk8;;)XGz%S*8&2L_P%b
zx~xN}=H&29Y5V)hI|mcHa!*7&p~w8m#tCpFImKGd&@Y+$3y15{W?q1|DVqob$>OOr
z@B`o9@3TD2)YY~v3j#%UvS<~N`ZH#Sx_21uCQJMV&O#F8;pIl8q-u#=_4y3KQ}j8-
zvkQmSSXIm=UP|GFd@Dd)Fonc8ws=}ilN(*zclxp(PS1t%O&qaD^%c*4=prNFY+!aw
zRMpHk<`I~U1Z+|VI9?)ey*I>Mi6+lnQ$Yn&{<_gSxc1%|(-EbwG2}>wWA5`5eYwN6
zHhiEG_u$%IUV!3A2J5>mSM?ZvAh$u4iILMgK`-BncqqwRXO<;T@T$@s<RNq}c4Q%g
z`9lV4g_2bExWNL0G`i**O$GDUoj7l^?i|E1lXvN^qzd_tqD1#W!e=s?p1<@&fMM<e
z@g*-@A#OY#a2X)^K^w$?HW&;F-_KLnw>9;;@Y=aQWc7;F%$^3C<)=6Z5e{QK_Ax(r
z0r<_1KX6Z*v)G<+ay#B%H*rEAQ!Z)8CWlui%RS;Fj;$V<JkbDx4$225icX$x0*dqf
zpl`aFhQJOaj<o2rAQrB!qZ|ao*&;00hkE46tPz0!8(p?oi5u1-hGxzZ%oMx^RhT@_
z8#aX}*a{x!1b&Vkk6-(u4H5FAJ>4*v6$Ezyue{ErKiQ^AN0yefU+>wuGq7z0V!D3i
zLOfaH=u0!s3Zssoc~RcAvuA|dJtniewr<>L0+q^k@=6OprHLFLT#RTu6y)U)`{#87
zw#Ta?@Hr-}@I%3C;&?PES8niudIE%GqjO(wYSn~mw#o}#vOJfAZTvt;5$M9SU$}K+
z{dLbbUCN+=7V((@`dYF4>AboJKpVn<j=4qE`y$>{Uei-c<!%0Z1qXOR@#GU~Qt^vm
z+j}BHj`=3wl~u}ekJnN?n2(eu*uBV_?P(lA5G<J8SpV($`|}^_Jw7<|9+IW>J;H#p
zL<*^^<kJ>w3KHFAZR4ucUgFcj#nR#)fhL2xck&UB1UrNr6Qgg=^>tj#j$q^as4A5n
z%QU?2<uW-5w#h{<-nRy{wkTIawVOK@_{)944i1pL+%(WUDm1;)>gOG~%c3C{5#8wq
z#Ofo6C3Gm3{7W_#rp@wPvpyeRf@~;09h`9whv_n(HYiSZS*u508WY}N-k(QdSEzM8
zV3H#Kz5+pRIOa9om|o5KQ;-ROuZ@xcU$eD^*PTy9_1vRd!Yk7{7}8tg769KM(wagP
zvDK^rIT;8;wmJQ0_rwo)OXwBB33)&zYS?#BDnA<#Dz!Zs(GLizFRZQ-&;?EhH0V<3
zb<=i)LcVy;=c5<$fV?u(3a50VdmWX@OyomZ=BP);id=PUd=S{VYrtnZs`>{%fKCZ?
zYoJqty!{9qn>Iv-C8PF@d;OzKt>>=yIG6V+zAF=%t#2_d&c|EN21<D<1EbFH55RBW
z;j4DkblTwksv+?RckO&gE|AOZd4cjHK>v2ytvss=yd}vic+VB1!`cXON~<CHRQyOd
zF)KWz-arH>sli@R9(bsf<=dbE-{2x+_EAQB*wd0TN=tcUnb`9t^vgmJ#`<qDS$Dfq
z;o&!R{Mj}@Fba{Ag7!%T?Xxg-3Hijl=wM4hwwsESx65hYzJb|i$kjmQG(2Wg-36H_
zIP$bQmwzuMs~rGj1aSTJFgunmSx8!ilBKU<vEU^Sc;ob6<G?M>n;1i|Rp4ai&3Q{K
zrlZqg*Teg-d;PlM^`xz`_g6x7GFmn2G|bjMxx6uUeLQ9y`TNqO<894i7rLK5G0<#T
zrfcC3FOF_xbWc!MJ<iVVg49~zR&GcMH*DBCB#iQAh%AQT_7t{Z`uZDJjLmkNPXbhR
zIgk6X>xd3NveMoF+I~38`+KzNAVKxUP*TzD-$k${WdW#>o=p@M9AU!4@I^6oOeGhj
z4$!dx_mwIs^8(2j+oQ?JGVx&R`|#w4HVP(LaVVl*U|KE$vxTw7RWMhTw@!K$b!RHN
zE6c8!WVeGblbm(u6w8Q`T!)V$p)QO(S|KK6U0)_D7`(|aIF$(RX8e@ym}w`Zjew*j
zR3ew?SU(QBE~lV(0EE~OYbMH}x!0ff4%$}zeWWE^7_ry1B3V4tI%u4WfnUGd;M?7X
zjiN&l@bpS~$haP-++12O{h_$85UNh6@-5x^(HKwmny~IW6r(>k4#y3f63%HU5?jjs
za>~Fy#uoi(SVVu>(wx<(ntSP0jl+x6tp&Sle_qwzxs#`MGjA|f`H^YVSynoFWN1Lb
z`%bihk%z5ca7(+oa&nYy@=rw@9i!{9N0K`)A9U2bnW{*f(lu+%*sH7T@Kb}%S^xQk
z_OEgVmUcso-c86q&|~8Xt-5!W&GR;Zxa9~{0QQ)S6F35W9p{}6fnaUHDB#)CZQ8_y
zs?)UwHPln2B&sAI-SupDDR&lzBc>T~#vE~5oTdpD2e-5JnB|n+5vlS*J(leen0U#P
zP|RW8<?nFWw--qt9v9xR2zZYwr;V!hTkM~!VPb_Z`Tsm3Z0>TVx(v9fjW=+6di0HN
z#6?e?YnHyCB#Sf!mk!v1BKz2a(3b(I=$R<vlbrJ)qC06-gk3#AQV~Q#R&mA9FhZ8B
zwln-*A*oz{bx)9@*J?kz+o~TCuewHT6nBqJqeZI6q$tsum<@S+O{U!u_lLg?%70@O
zocAe$XOsvPuTAw$Muv{ZU$G4y)`+%&KT4paBB=dxnDniD6)dc%9N;#~f-3c=Q|}eK
zp7)vtg_!hcf`-`SgvjF#M)S$!Z`x8j3>>e|g-3qhU;8Za7}!V(dH8fDIC_XHwTy2<
z@__f#7nU9@-$QW-@VlT~A2~KFv9OIz#ojLAFW1Vn1=R=Hn6VY=%4H87j2ik?5*G!L
zZz-gfg<lIc*i%#Ms9Y;vQIfTD>lr+0<TZi3#x-^C&~*fy*8<-)lm_XW#^-LFbW<;V
z;QJFmJ3GgQbv30&2)RdGE~boj2hn|8mSnr8qu{X*s#HH1nK9B@Yd)E^2|{@57nk>c
zl*!~RUJGl@Z<ny@9(r*QZVl8v0=Hrqs*aj54VLB!=-gG}?_wYnoDc$J$0+Rwoq?Y&
zB*YA&%S6FR*^EPqOO<mU{dM5Gw%xgYd^7z)TzBSN5|OeS?S`Km_VC~zxXQ$SDgfbZ
zm@#;-A0<}1FD600H&i)gB*PJ=_AQx;dn6}>z5d#4rSVq5$@N;6su7{rMK+y_n9No~
z<2z3fLY@0L1RuM)qqe(S2oCIs_uC#WJ$G=~e1kDg1w|z%9BW1&uCApqA_^n>)l{;y
z^s=Xr$!dY<S1Q!y^EM1=3`zQXPnEQIWj$;Y&YT)jEFb6KRV)n<dG|7OSMZuCE06%{
z`$jo^=SHyQxk){5>KW$bRKOoPENuNW4(9aw04=mhL&W?cL;)-lFSo&*Q|xH~$`kI)
zhnp)v`PzN^&nkE7H~6j>rhtB6@P|Cz@zt-~pz}Gt6RdG$*=iy6lw084Z|$(ggIC<<
z4zij11eM<9rvZYMSGBq==XF~TTr|r&6Wms1v!KyPfowF`alotFkB5%zwPX3u!fr^_
z+aM(``T`bWwe0znN%FKBsvg3O<;I3Ts$aJiz(HFd4WkgwVY1=^ewr>jVIn9x6;H~c
z0wTe|%;17B0FuYy77n;g0@2@)I*AsrjW-aylTHX>s?}JoGpP7oPQ~a@6(%mh1mZyi
z^u$Mv6_!T2>|^AzLc0m-KEg_L5voOi*IM#&h(<3ut-w0{`Sui#c{-=!agP9O0+nqR
z2h$BxmA*`Q$-XFq%k?rH0df-&ijnlG;KW*JH%#!Lx6(p2pHpYNopF~Ujs>)c@5k3j
zt`CI`PD6op1wKlA;-K+J9J=aeteD}64jZLJ9ed=^J!)Tw<z^2b-tz9Ex$;uU_*MTT
zq7FqoFY~I>NZs=@DIfFyKd^DR=DUGn5Q+{*QCUwkJ=J&s6XWYp9W$MW__JM(oHjr3
z(tYCrfjENZu1%_nr{4-pMsu8XKtFblOml|YQ(t=X%@pOB)7J7P6jSM?!k~l(iuAnZ
zwfbfh9N$h@00?$^#1Lq5`qt$5F<Gb}N-t7XQElDBtz{SQaec%#5C*#4=8dvdxIJpo
zoX;V6jO|raGHU8zX?M2jK9m=do&VNRe%&_|1;0S-7$=H0?;5KEj6X6|A`zrhQ0W0|
zr)bwcg-h!Nu&j<eNz~_KU*x3yD^7OdZuHU11i23mD&X|P==g2IT`r)0J>!-1%$pof
zu_xJGc9}j74(eLw*lRDdgfNidAtv6RBI>d^xR0l`Oje@4^0KSDzT&z3I`Jb9C`|;@
zRBygt&#@d_)Jjl7E@!X`(|_8!4bwz1EK}f^K)4E>{)VAR)?(jMz#KKwzw)++s?c)6
zV)vV9l0i%NE_vlku&C_QRkV@cUl|ErN{MnO$hC+t=j?*%0uX2>E*~7soQplBuhdA}
zrwb!JW2s8w@|d!7^7Tj|8e`X8?1!}zaI;*n?MbS=iCq956y%4&{{wiSp;Efj(bl_e
zO1%)LLQuHJl&IgV@@~%aaR%8EH0&7U`1e=36G)Si-Guv2q)g16&zp&Y3`9OH7I3~E
z7^9f9(i9^$O5<KSeH{&S^Mwnth3aed#wd2?Zlx_XdTq@In^$mv4!BcsUkrzP<@EcN
zMDWNq4%|tp7YS=b@#&5kQ7vALsqgIMp&jpK*NQ7g$*reZj(}1}$jdcfy^a=U3Dw!W
z8F`XZMNKWf06XzKnz(?zRXpiSSl4SJW&J^LvbV(FLrUu!$2nn0WcT2@t9TyPO*ffa
zt;U&bH65V)4SD5Cgbww{mdf-;a5NJr%*ZaTdLCg8npaZrGCQ%;;Ao030%~K{^}o^U
z!x~v#!?L8(Ik*H;hs2AhJXiPt#>=)XTfYOH(f%{~2`*j(IxwW7Gnz;<O#p*KvnJ%j
z%O{9ncL(3kR#r(fe3ZYx4kIcxbUC<fT0y(7qaf+a5-C5AgGPpCFLgRcnIPqZkj&-}
z$^cZgV%3muQ-Jaiv2j(ytRSLoZe6`5E~AaAbF0OaRArf$m=b1}y(%}U6{tB$PEPN*
zrKb2~w#!yrH$Dz!R^V4GnIlZtA{N|oclUu(68C=rr|EkzZd5H6wZO7mF{Dzb&F3w1
z(In?Df&;p<Ngx6if&xU3wMUuc91{D##wPq|^AIcA`uidj0y|r+BW8WsNX30SV#mI$
zL?FbqgJ{Ho>PAU!K01Et0SuRCJ<dG44IgJ!@F+o_sVql&m#N{w>u$}{mwLv)ziGL+
z1QKbYL#ZNDffrSBXK#5t%2b?7Ld9hB5tHohC$Moc)cLH>WHvSEI+dKJb^+S<##qfK
z$gK9IShMen1;IRhKw5?KLa?-6GYCcKoa%~hST_*E=UJAi9Y$)?su{`3_&`4lv@<{>
z;R`r)i>r!k^Ro7RrN3AK!>y5K0C1J{W*FL78sul&E1N)d|B^|9`_OR^JG_t_T2!=o
z%Yh=XlEv&Tl!G`^GPBx)qqU(^+DLN;ItO%0sC99zq1wX4P2<8m&&)5{tPc46O~dX(
zAo40e^$f!HuS)4@??-H+Z__`=#!~bY#*;Y25`abRgFAo;2<PG>!K!mXU_e3}k(V0@
zUN{#8Al(wgegNr7Z0WzBb|4C>c8+x=ot9)}koE7tmZ7-BXufwyRk!`~wGHH>cCSY$
zK50X2FZatT)#QE?9qq0*nw~P0=6A-nskg5gYF}U6EodO45b1E3e*$e2EgMrS&?H|L
zTOK&0=jVB*{|Y>6_aMmdq#l(3h;iT)7wU{3(}9pEv>H%}uKxhsBw`Q*UhLkR1|9ip
z#baXxQzvSv!kce|zs_r!Zgy+;`Z7a)B037XdX1={8!WSXiB<D_>HY$BAzw%bB5dow
zK^q~nWMFv{RmXBGeC-IZseg18n+RZ`HAq;IVE2-F-fC&HHg(gg@}{7BWu%DS_a8;%
zW_I*QY8OxXYzM%N856P>?5K)N^|c}(Pkw36Djx}^1Rev{O=~Adm1q$KPV#Cwd7#q(
z<roMKQsZTIs48kP7BpOqs{OlKD}~KEPL?VjkpEJgC5ZW*r1sr0!?*=-;{s6=dbF{1
zpvnppjEIh9b@3%AB&~sle5=b1-$-cd=zPa!I{*mgR5bTaYUg+^p16idE`z$cnnq^-
z2?BpWs2zd`{IKr-4tu)m`X+BT@T-FH0Jw!ML*_msXdwc11>^)Zl-?J0a`FW@;G-ox
zWzquXFNNHH*~YFx71c%<d(c`U1(>(6QH2Zov-y4<cry8hh#pvE=cc+Kjq?Y{x6jj|
zr2StYh!L`IvjX4-Zku=ZQqiOq0aRWDV>s8`OZl=F7*MmA0RPoFt{PyUv;G$3p6O8l
zZCnB%V|f4Ai}1B{hJURle?Qcs-Mgf_&TN07(j<P84=aRb>GXJweE0>|?~RNqUrMc)
z**#x!$LQH9JRJzw;GI=_x#wj_bY2DYligoG*u06kI|kx2fvrtIfhNB$px~1R@PqH}
zh?jJX;=)~uUpKTPr)+X7p<doDGnyU_<nHgEU}c2G{5G4!f&&ZS#4nPTjkb?OZScyE
zp-a9znomPCNATIKaQJ-%0f6$~Fs(xc$Ip<5jDM@!`Wi0_rVz&RyZA__OL*HPK{omh
zO|6s2PhNO{5>>6%4>W2CaV#gr;p<!gqXKGY-GPkl)~?4klCNcVtCNI8H4AbO9(cXg
z@z1rDi^u>=e(}Y84@nnQK!V%phJ8#=7T-MqHZkRZYyzn9%=SwvOw*ZTE>W=yYvGTK
zazxnl{{@lQPw=3W8Qy$&M_GNRIHXuR?vDTrcHe6Z?Y_cv#@OW7>8Vi?v1L4y9N`BB
zmz9a0>&EX3!Ms|@-2Es*Ls`XtYOQ%kgZL7#hyzofn@7!Bxl<Vcjb3hOonBKtBe4r`
z(n5A_ydA=G-3hUmepp_HSP>xAVt-&<7xDc4u5Ai>o!QoW0qSZ7fV@EXWoF~^+IbP=
zsR}E_0ocLDPg<ehc>m6UfVbV>U6Y!->-D1&Ej|CveR2s!BY=gH9`c3c)O;RR1-3uY
zbL^Sxcmy)d(+Mby&Rx*M+scd!R&)JY7IE-R^rWn@4?-|N)!+o|4H)zbnXa%*K33lK
zdw}&NF#k>nr(K<6Sq2CTRQ^gV(!$%o(7x`eexUr~UcaXbG|qcT3l?U2YtCiNkDMdH
z4TniE<?u71$8`Pw9+6bTEdgfxazmNx*v6wB9u_b37m&li92-N>@hA<cnxoTT9I2OZ
zVb(>~x}O0a=q5&+pqW0Tb;|c>HGufQe=4$GY|#EwWSQuG-}_x@d=dpSMTcH|t5U^j
zCEZ9LWX2E;!FrA8H)+vBB2>IUn+4(kz0d1u`~(sE`~p+&#G?>QzKYC7S~kPp3Owkv
z05LjqT4sXm@^iZz2uP2?Q2|ZqUTifa0t0psj%xNPyCA1Piw6v*UGF(89-i!g!Y6Tq
zlDF1poxx`ze4PvGfwMO_%<gWtuikc>f9h}!D5Rv#D~N+6sWAm#%t-gp18#g6x~W1k
zTnT1#i?#g~`O<auT}O5&EfzGd&DcRKp0fZ%_}(5JP_MZx{p27rCM4WEIBj{p2-|s+
z;vm$rsQ&u~l}F?6x7|CG?30F|koH_A4(D{xFh(Xb{h?7NjbTvcia-ozUTMLQCqJI5
z1d<*IhwTbLN@0{xJqCJ+%sEc;T2#4OH8N^scldjxYm)go$}hIwQ<h(jz3)7|nN;`4
z4+~+MqG&{iVwBF06SBvyioVS2uS-h0el66KK%&!S(-a_-zbY6Yv!S}WeE`5qzT;Dz
z#)jV<f$JjfF;jqE3+TJLX2{L*=Q39ue8m4v@ku_ZjvXs;i4b9$H~p?_mvad~@fcM5
z-94MSg-`(-ArD@wqwUPL560bmlP|SZKJ;s&Gvg(20aqpvkuYXuud@y?G7;M4`Rw)U
zR?Dp{9?Zsq`@5r9Mzl+KTqZaq!VNMt<De%cMxeQ~BaK7jEd_M@?HhHonICk?pP$a|
z07kY-8gc)l`%@@fmRgNi#p$5)%9mL^OKe+##!LVm0s6E{I__tiTfQyN0m1DxAIV`^
zSdS5Lzbp__4rcE<!0Fiz($Jc2Fo>-!Qh%FuoW@@1-&Fq2<n%ql;G{bHKH}o9HM|=k
zt7+~~licaM!q3@zjcaO~&O{t07a^G}6kj~>N~g*gRe`AijbVUH9m>2-%8tCEu(F>j
z|Iw~?gAS-P>fBOZf`UyW9H8*bj<Sj>xloA>aEX=NmJjjF7JI)5g^<y+aZS4PX9K@u
ztIO+h^Y0+@^K`b!cFtz|L!)}gLT)9A)W>fyQo|Q`{-*nm^Im<$rh6ABP8F%8%dTi|
z*k}=~8dZ%|`m8nUYodN7#IFKAOfT^ZIvDO!kz&3s+3Opng!FW0c~C|IJ$=V)3OX56
zArPD}i=4WP1)zGxQtU|X#*v!c)H^wtPSsFa1H(zRrA(duMF!7}-M$i^26pcSYX~Oq
z#4LJ<vycm(51pQ6G1Cz6gB0DmYwSb5J!&6ackxnB;%U1#iZko;8Pd>k=C>ICJ7F|L
z5k?SB$p?^=uV?3<8}5XUYgQKRu0873y(nFm3mn<feq`KzN<5$%ZYF3DGFo@?H2<dM
zPQvaMjN5ieHq>I~`Sj>DkhXb28U&|M^^V1XtZuuVN{9^XSCM?oN%FNSROx{MSiljk
z%3EI*h@BQ@N@B1jQ>4wAxW%YEF3Zl|A=8{U>t3UZSRsh{#JSoO+U8^0*5WJj!X6UG
z@(Ms;?)_L!YymUcz^neOQ6)k_giNh1dnE6TV>$|#(G*w1m7Pj^ojGX&fU72TxH7br
zpxPujrkJ1o8!fRnS<ipyVL^7?VCou-T(rB~cxnOVhfzf<^^{{`E14ghCtS(;plo-=
z?Ai;O%G%GYmH1XMdO!SQQ>up#1wW&J@a+CKzfjGhg%ZZ93HW;!s7wUC9I(e+N%+4O
zBt*mZdI1Vn%3`%!ea?wM#RbHhk|sK2^8hQ6hsL>6lJy3sY-7M!?C#m1{G95;n23#k
zr<i9333xqRK)U$ziUX507EK~1Tb4<z2#nH=nM8t53I5Gq$zo;#kX+fbtyP4sL$jDE
zPPh5@h0Gk2WsE<o5D%B}_v!Z|O)iI2pDsL?nHI=mQ7As=0VUjjLvd!**sa|3y7r?>
zK1T0IUG&%`V!;*xcYlRU?u2G(&Ct;J?d0@4*>5MTA41%@=ec&QHXn=u*5?1^=$MWu
zgq6xMsA6ZGY%-7SyMwa3h*330ro9y6=XU-=^$3)TLm==T1(uoo%Ew&d^D3%83myX&
zz7bdp4r%`9>#b%A)B8p8=Y3Q!2Ue?H8mS9fT_Ugs6Cc>rQSxY<X#T@E_gD|1+V77V
z2N;$HG<R5^Xns2*$fsk9dGav=gL6x1?AE8)%Uc4Zk{S0T!37L`Eu|7^O>Ma?1dEy2
z2R<3kxV7z<pkxJ_SG%-&Gadglt83!!NL+%|A`7@40tt>{Ss|zqdWer;13uy_q<*ae
z{{59V&hif%84fX9hJ>-^BA}h}*?+yM(U+@e*evX6#t3xuCKva>C2W!sTA=stnuki>
zEj#Y6CG^g#k^P?uXzvpUs1KFCAd8R7EjKcob}gRbJz0ERjbS7#D4@w38V&B5O#DpG
z6O>@VFB1e(<-!mpXKHQ}BO~kRKPc=dg=(6Bkppd3Tu`9+MnP=6FcU=gELE}70#!!=
znh;b4XT8GDmokByGf8V!-Y|L}O-jb6uigJXiq$H;W{^g5uz2MCNmpInlmCHvqCwE*
zLG7EO_&}Zmi13l!mTf?nG>=&s3S_mStxKu%h|>f&P5vnVWM7cU=MfoKKv*Dm1%#4E
z=bSbfpc*NKp^WQQ`%vke@$hnKX@104cAmxS>}&l85IR!k!ltCGRtG>3#?u&I1t6nY
z0j_<ta4?xz&R<_>;Yw(yXMvdaADxaC=yY4GKw@ErbPGMXA*@xA&0}q+cju(NK@+-@
zjyOP*(@b{EcrvXyZr-@YtTo=wA$ar&x;O){ESQ+sFYo+U283cJN<9As#UxP{1v6%+
zT>RcqfoAXezedS}mWL7JDcr7-3yQ`qe49a+tyZHv4d}9sbDNXv;%eg^Bh1)tx)<H4
z4C0y$6J>l<jl5L4-15MGDIqFQpU8xKdw~FAR{`3^V^5coEES^{IUqioVg@LSONZ9%
z(wL<+-QF^J*Ff^R&SJR?zG8%snKS&Yi*OXwG6Xs$3#SY6N}!|+{%6fdU09Vpz-*UZ
z!KCHTRA#VBWw_eIH;3RJQQqHd>QQievQI<J(4y4cO+8kJB`xji!J8_tH&Y>fGHG4G
zr}nf}1=Jg^kD*hqJ^?Qo<b$|0s>%cwjf)tpypVkD5l}<_b9e0Hh)y_}7j+lHE>S(>
zYkFQY5xvv6Fb#o_fbjFN+6;$BWzkvfU{o2>LqKy9Tq=QhT?#4@Emp7I-U%OBy@B|o
zi|$_HvxG_6pqs?WunCAEmuztWFf!=4nNFg%FTa+Emyo^=k`2k!iK!=OC37!1N)<eJ
zID#6R)z@Zzpw<k41)yJzZ+L6LQFn<TtkQPxl^P5+nO?R|_4^~uMJk@&3vgkcu2q2!
z1O+J8n?ByM=W_>Zv?;q5qit<_Nj~ReLjm*%XT?i2jTz|DTB^I6R3~WcJUs(q{P+&G
zo8tO6qd=-B3#?GPJ;A4_{NHJ%9u(rvimI1I$I)wfztm2+EDLg~jJG6-mlmSlzEq>5
z-gvOLc044%axyCrZGF70#m5sZ+zaLkEMC?=F?_yJ1{`1Vb*1ihP4mHV@|f|az3m4}
zkI?X&PM_AmjVY*}z=Z0FLb7WWlXo|Tgt$DGO4u2h?bh(&KY``|lrPXZi!$>}QSe+L
zM%J}H#J9KB&lP+g{T{KawVoqxfoEq|q0Snk<MpUF5hbpBFi-UY?3&~SC4i{iWcgb@
zW!XM382}n9@Yw2vu?Tt)YIvwn<0&(GH>;b^b;8<Fpl?;?bhH;D7wnuDKDAB0wTzS6
zdrjexQGCBO+ow7F{qK)WvrgV|+r#5lM)x9hJCghmms}n?aEn(NUaIBhL#|<3QA7!&
z&{HkGoYtA3t|na-)(+q@9ppA>s3VrZu!ZvW_azl@Z)s5h*}9PbOdWIxExZVn_e)&4
zMrJJKe})WG5{QUj3K~`e>;votaeNdxkp3^`h-`+Gn6K5Fp)2{o>m%g8BYStdu#BLR
z7IpEex>B=B{=DR%dw6WuLQT|a@*<r~c%PejlPZ4WURk$92;Nz<B~bzEH37u&xAPy(
zL<C$7jI+!OOlHoFNCqI4S-!j!odfEIFzJM-(gNTNs0PJyV$xz4#A_B&M5K$yBRa^Z
zBv^N0#Iz*1y+Be{i?x4{jvvWx?PNXleuUU>gCK@okZWuHq-TMWM0qN>Jn2?v=>cE)
ztn0fxfZ-{2Lfit8ids_~d<i6(!b>l%%5zU2h;QOTFgtmtez@!?Eo3UAX*1)V|7U0V
zGQ;ZooSW!9@5w*?P$arHGCUbwZ6H{Vs9GVdu&s|_Za?XRigw3;CwNO))>qw7FL3__
zy6psN8kwt2By8+8A=I}h)fo}Za|8=~tmR@#S*r5_&g#K!te0LRGwaENf;2u1u8!wS
zpv}F|X$ov4dnG5Dk%Qjle~Xf_D8G)2ePcFD^&L3cu7zM!z(gV@r~y6=By#Y&g`c&i
zBV=4V3bDk9>#E0S*%~CMs^CqoU#D@yj}7PCJ$o&0ZFw@ugYH-o$r6wlFRds#O3=C>
zkr#@W4KXWWvXbHi8xQZ^>)+OoZfdeWwm+%dC9_9<m<MJ37Sz@tiyTND1C;e)@CC#K
z(Tm4U^pwh}9scQSoURR_h+zE@itcoQAm70o26Qz_H!ae7Xb>Y~KIn9)dtBN;uBI!u
zRwe{5Qv0)q*+Z1(82xxjtYc2tK;GF?4DyxkZ-)fR>cU0g07XhzhflV(N<g~wg!;I?
zldUioq?&i=a8_P5)seX$*!-}2b3qx@Vs@F=D92aFu$;LN5V5`0+52<om?#V!Ku=(|
zxmGw{EL6$=I|t;RwYphPbM?w2jGeQ(xkt$J>~qpQcDrZt$eHzLbHcAaOVG?wR5lA(
z7NLpY39`9{f&S)#rKvM$p@<jx>?=HlUKv#Hk4P*2KgIj^d`#3RU%hn>@wvpg#Q{M2
z6D|$QzHw0ytDxl^k=WOL{Tt;r8#7OO9?y8mC*zmeGoafOkHZ}9@%JIfeNbJohm={?
z26q&Z1~#C>_0MDoC}i+adFg>KRZp1z5~vH;s}~)ljbHV%$gk|#haS8d29<eZ^;}Y+
zC?8oYMbfe-0>6MKzXh`kv^)F_QIZDrpm_dtrJXV;6)&M_;Ck!JClhk#iU-K4T3Unz
z*5L5uQMlHoS-0X7i8v0uP9k}K50R?JM^hx;hOsO6PQvUNI0Zm74Iw1|V5PrdmDMeW
zocL7#>ST~)thihRz?>JzE`Qc%LMlb-CL0mu3GH`oi}SmWeh0^v=0(D5C<{Ok>U=CH
znL0y*A0e`qS>R<OOHuk_=04Pm26sq*9n9*1us#+*?Y`(-K-t@mC@Ux5j_lI{=sDgZ
zVGdttsBd$?-Y;DOZDXkA(}XC9zugsftQ#@%o+lqnIYR4lgXRZ2YS;rLW-x&W4STpO
z6!_GF`-*jlw$|2|LzmejgOFpI6B*w$RJ4R|S^Lmme$u!&E!njw^=<w8=u?N2qgme*
z>Ko#B5*c}V4b7fYXL^X&q{m}+b${GBYa}ZL`Xee_*y;ka+JPpIYKim-g>XFju2ah|
zEPP3fIBpX0oJO2?AkL3lM7(#~=oUth+jx|el(PaN=i8zuEZ4%`LuVt5D$<`rMfwRB
z<-%M~{_0wl!Q9SGP5rEp`m3%9$6rPWdrD3L9sS&930YIE@UE+Nt}I<MA0z0+on)|h
zAup(yjV3DWsq(?Sh}n_u1c9Ef9DyA_ZU075@tw&HR-1+9;dulTqB-qtU1>LmXIVe9
z`~5WS>vb1HG9&fx(3>sgyZwzN*K;Pe->F_MXTJuGPW&+^6gtYt=-4hLid>{DkJ5Q>
zmuGn~MG;W48(hDSjVTJ4DU2d3>%8i6w;+J({H&FITzsZ)r$q22W5k!_v0Hx&%>V(x
z$>2S~7nw}Ucn2_j0cz=XXXeiC<xQd2r|=XnGB;o$pZzqduCE-Db=kg`Q20K@wC9el
z!Z2Ce96*G&m=gP8O;0j~sY<Rnfz!B0xg$wSA14bYlTae3`6kP9wy5``&qEq36QhK3
zf_>@I(OzeVZRJG0khTn4uD2nW>hCD=p?Wg47^zs|(k^Q6_^<jn{(4s!kniQ!GwE=Y
zeu6#_TaM<Z%8=r|Fh${_0>pqw`9i>wBmoYScUNoc$1Rp*O~|gjRl62E;$M&H)?WpS
zyMX>9Qfr;TO$p=UXZ{f(#~^zGPN=)|OOmNQEd13jST$<b5!}*64UlMOJcKVHPh04c
zO3g00cXvCOG+|zlOZpBoPj07Zwi?i??(?Sp7H?Bk@@B^C?Aw?=@#CF(_M$psFL`Vy
z*mLQmS6xaU8Y=-x_>mGfYP9-1O;&Dn7QTeJ0&e?>iM{poN27qW3g6+2N57fnA4%md
z+pVf<+-_(1<)J|tgNBWrI7d<Yj1MmsnmCbPr>ec1>}WhNMmQ<MX$#f99v<CegHy=M
z9*T*z{w*Jm*&w<ATndJ1y{9%1XGu0IH#Zmmvh04b)I4|plVN-~8Wu+ei!D~Sg6x~9
z7NLh{2Niu%=G+NLo9A1gQ+`71!8kEtC7iV0KbI-8-g|j+#zVpH7{hH9>lu!w4HYE5
z83~Mtl!6IUDq)E;lI`qb!r6_1zNPXevU#3ar*!VPguLI{xM3?PmX_rw6FI7f+0z)`
zv2cV?+sxSDb(Osvp+J91k>-W1b{uM!5w@vE6;Uc(^UQ<pT;o=EhJ)?Pa#SaG-IjPi
zj`cj+^3@rzRs*>0_GN9PpribL?=8q<Of>)p!y8aYSJ+3L;`3thp^gR)-w_j|N(`gb
z<@?>Bmzyq5$Iu0o!*2&(J@bzI>CwfL!@y&<z`^1^5#7iE(dg~%y=#Tx2+i5RV!16&
zD8skOY2_V-67{987W_|$hS?Qn(`j6QZ=)r*RS;n^0BCU)Peyfd-!S3Y<$$xtvb=Q7
z(Uiw-0>(b-Cx~K+1guIj`Q%>Q)ecu}YzIUxBG3feZ}Gkkz774-vl6Iim0yQyR<PI!
zTb(2bK5%`&sCol<U*@{36?7@IK&N(o0(%)6i-2)%oE#59BM+&|?w-^RCd|_`cs(Rn
za-1=c;~*ff`KbaVl;lQ}f$aF%T_BjECe=2KXIK-9kNQ=+W~kX5p`lS&EYY2k{vX#y
z1?)pp65l@j^*zTPM~fF^9B7rs{3;w>*M;yzDIe|x(6n_4Tg|X6?TeS$U=C4~uDFTq
zb>$4=FTF>=8UyjsYy-Bcolw_=E$@}=J7c=53ACZ02U2gi6N>G+A?XT?T~1iBD+Zo-
z>VO-c55b*O)+;xq_N7?osGsU1>r|YI3^XB$n;M%31QgO9K;j1@khjB+(>^E5WHQuF
zqXf1I&^Fmb9*TaABOL%2OR};apU9$*qkb+zCaxm6wU6jWu!c!Jtj-~`+3T1L;!`Cl
z|Kd|p@3<52B%BzZxX5#Bc$LD>BWF!S&c1qPqFJGWJVT~yYK1Z?w=pQcMPPU11y(um
zt<RtW-As{=01H8|iK0($2~uU2N+L1=H%0Yq^90Pik@;ib-Y1K8)x&ACrwvyd+_RS!
zq#<ABtJl%*2t6cu7&0j?pbdm&+6@KS_n=F=n=t*hTj_b}7`RL4$yt|K+*h@Cmg&PJ
zs`v$594mA`FbDf1z1kmP-m;lSegt?muE)8yTV`&q7#y~ZQ03Ept8X2^-JPb(K4~<m
z5bgw_zDY53RAW}tj`IvnFNErHJJ1uGqzWpdU@{YDed72JfUx%1E_LAk6YnJ6_%BN$
z(SasxikVChN^}QBwa`3VgHR%bI}s#P*++a^Bty_xJLoaACxP3j*>i-&Tws_YngzrW
zCsfeFVx4m{RPLsQQ9d}^F49u^o|FFxEJ-IV-Sj);OL9C>7sTk0Gb52i_dwre8jQc3
z8LTDwl`>Zy?HCcd$dK4So>H-R7A?2V$f($Gc`=RDuIus*K&73vYl^$%#t%$^EWB0N
zv|Eq85}m?POFb<w8)9q%F|l<3BvmQmJsF1tDViix%HOv7-r0c`7&K5tt_pt3!yxVE
zT)<Z8?D~gI7DmMhCL8u-qw}7-08%Dveh~n9Ch|$&ifceYSd#o-c%2`0k|Gzb+7Wyl
zU^wEFbbl&g2i+m)yE^1n#w}0zbG#AX9G*NGw-+OVF1dpddfAoq(p+VUbFHs2<Bxi7
zB<o)-pU~im%?sGVC)N7mV=$FGJGmd?qszxP9uLySub(Pz*giDVgHE96>>*~U1V5=E
z`1ujI<+v46)~4Mn*}~t^I0&zH*=lz{GleD|V5Qyq!Wc}rKdi)}c*))LB<`6e{=T;p
zWpB^P1%^ZRLYZyxv=ydTX>_EM1h@7=Octg@M!_IBaMcmI>c=PJhkTy~rNPD6oyU5n
z5Y>k{>diMEB#z5wL}4F^XnuEh7=rrG^k<QKhDX5T*#I{56e@#(-ay2sC*Fsb2~@2K
z)_!346}qGJq2y9rulBkvbkmL4SLKQ$Fu6R#BR%p{VB#m;^s%-r$!X`ZP03(cIcq{}
z;TJ7_t;a4h1PZ;4R*BB01tWzn_l|#`P^4#Bdu!uhVP>McRN{?45ro&Exaa7FA9I#O
zMn=KY&`|oEPkck<GN?)C(^ka|X8T*6H-q96tM@m6@D@>htX6mY?Hlk?Ck0*T4(7NI
zt}XKNVnH2t<)Np^M<Ml;`i!5K?za20*MwROhqySj9(%Q`n<w3mDD40$Ei<q5dIPJd
z35!nQ{&*y#eb2Zh=sp*Ttxew$DxE;I_6~jnm-_+&0@@W}rX(oyi6V&oFh7h@YpiLV
z;J7^u0HeO4XWKlUsXPpH?HP*TH)J}ED@&EP<L@gH$1?~#aA|cH?ux7QQo@7Yu_W-a
zrh!o8V`a9G$oX%)Oc^kNi<Yd9Vlp@hI@fXk9%D_l(6J5@#6Fj4KOWnB5EAKN6oS8m
z&<9FjvKy%EG$GmM>e3_w_|X~Yf@?F?jbGqr9za@>OT6~SmUMpfZj`_Iq<))vf(`~w
zu$GRpHcdF1Pf4|DW**W(VlF*zmeAPB-Q>&p7&n~JskhISk@eld$?NDaic*go6^fa%
zg4h3Gh!CIKr2{Aep{%rO>MS;MfY+{3o%fOwM!Iv$eG=&U_RZtM;OB<4{GW*H1&V9x
zo8=Dkm(YXY(F6>J)?k8H2YE6`DAH?yjtocch@_^b&j0uW<q7!g^go+L^aDQ&jg#9r
zY;F$<R<0e_Wh)Zx<FzYvj~e!Xxo_&sg?<$p6XrH8vnC+fEr8z_0=WjdCV69`xeaTR
zDMZo3d42Q&9%V1MF%Gz<+;05y@{di28ln{oH101OM}gwjanXE4A>diMfn*kCvAS2R
zRx|i<W{e}#=yGL1ZUL__3H5_F<(C@0f36!}w#j)@<B|0WqL%E*iydDwP&GSms84Md
z`@8^d?h1r1UVqzp2wJF_`Lu@W<L$s@<6n35t59u-{H3de-tJPsjJ;)WP@TLaf=trT
zyj=y%N>5L}TX>}u__C-AV_SyEebg{IpPeSwhnsFZsH`U4YilZ)VuAN;{0P3HIvC$5
zU&+ph8tZB4gNgN9bH+fEP_K)M`UQSx$8^&hkRIs%cMn}PbZwG;K#>d(sU#&;)s6-X
zPN-K2bC1;ZEWTV#IwTzxDH`Q^Y+-&VTbV{$Qoi22XC~R`P04?d{_C{}aZe(#O{(d5
zGhN}>j@)QpVX5-*@>e&;sX7PA0W1qJr^FIewyV)=uu66FdwO+pG>bL^9!W1W7yw&V
z`#m*qFycou0$&9Q6Aa2EUj7){;ka$Sw%JiT`1zo^K;zJ%Ey}*KZR%PVr1O)>CO-j<
zh*PT=p=h{y9%y<I#l+_WLH{;9bPqV9iT9zLhU-IxOxXnYgH_saqAOOoJdsi?4HT!T
z#UTw-J;0gq2Kti+Yc#pa?2)Ap9R|cZysZU1wXOSMeWhUD1gI}IyrBjdbX#U>;|zbz
zm+=j7up|nxBHze*YN~1i^F3)hmDW!HznP+xNyl$GGf+?EeVihw@stjg);*TmZI0hE
zAKuhV5mOQfzX`v5rVe(+sfIY(^x<)9GAsJhlS$DB<~2emQS=J*1O9TB16`uow<BN_
zoB<&h2;HF(Xg)<_lE4ZVb3a*#23Jr{QNIr+`AYuepqD-Omn2JB$DW@!+)c8MIhNVP
zA^X<@=1R9UonLPV4MYNE_}q75i9~qF@HqyUUk6DxlG@=;2KLC2@OVsQ<n8wCuY;R9
zpxg#?FUq?^4A9Nw4JVumaB(;#C1v-oKeXPzLRpbaB<<<l2$M!lzz8^TmEyt{;7iZs
zJIYRKrbSe}35C~UMtkHP!hP%%LwJm_w&wE*ZM1+7%<_;zKY#^?e#)w}O?(FOwVX4g
zB@zQFtpjl>+>bW3-QN1<L#kbU3OPeHr(-9>NHP-;NHl>u2XI$Me(f~8ciF1vCmk_@
zlzSm-bm*HnosCaglXxYQ*#6Y_oX@(vQF436)>K2HW~z+cS(nI14W-(StAp8cNy36h
zH&3B!9WY5nO&1Yhpz{CI-k1MVwf+BZ(^VuFMHGruW`!tYDJnOj3>k~cRHkH(6QwdX
z7#cXlmCR(ytRza}B2LB=jwI7L$vk{tYoG4z-uL74CwzY_d+)RNT5G@7^Ywf^U(fB3
zZ{9Evt7`pa_XSvN)oS;4I=Qyfin`woadtcisTXYepb|K_;^i*c)d}n?#0%8S<ytiG
zhagS<VcOn~Ct9o;8XBW~WjN%jKV@VoujpsDY<EnHH_Y^hzLc{?UX&WtMahDt#ua0-
z^;z6sGn?!L*9!Re+&w#fOwdE2Z-JB9S=FU#%39<^V&SLd7ThCA@A2N_LfrKH)b$*t
z(^;l>>NCvosIK>rAs{$lb`;w$vhM_#00qN=Nnf3zTb=Wk1LimU^c#hfj&`Bz0~daJ
z!^adLzYMcG5>jjvDp$UX>*kupJ?jr#ED7u`aY<Aa(riXclspGqpq}6N;FXuBml@>`
zzf|X`AW2`0Sz`t@8mL-EZ!CDt+XxlJ)D`tFWU+x<(=so|nGco1meH~qN6iA@*Lq9c
zieM)_d(cJf^rhY4W<af8p&oTo6G8jHKqq)fee0Z;6?gp|?NqL=H{Z$3>&SSqpB~$6
zej-rX>*8H~k19xl6kMy5q;z}K?O)G&Yv)|qTxzit@Gf~E!|}N6iLB*%L4M=6(R}Oc
zy3a|EJcS(1rQ*x#1nEiDwZyrXwy~~hdOv1W!=@{dQO|ERxN&Qe54>b?g(E{BT93^a
zt-o!R@FM9*VP=En-71bw@6Gy}qt`-ot2o-G9|Lt8dVOZEbjR@h02WKtiH5wN!S`Ev
z(q?4i)93%WT1x3~$y89@v-6E<w^YD0s(~-W^&{VU)Txa%b(?$Ewg%iciRX0_D;cea
zAd_s;t0Pm=K{FAFsZ3FN@!NT^X7Tnet*B|*3(?v#===mn6eyC4o$NlVxBbT>=+<Yj
zTuEK5UT)zoYx2!#<7u9qON1@o-_B$9Z`->4WK%R}Q<|xEtK6#Dr_OzZ0NgQIB~dbZ
z00OA9=<`w8C9_3?8j&>1;DVcJ!$!b81U{&wz7VxkM%GA9-9<^}djKkef5A!ZPx*A7
z&DyeK`j|`2Aj;TkYI|0O18VQRIvS=oI$-D;Ks^fWYq0QQdy5y=>GchA72BK`qFtsa
z4Mi0S3wiy?2he$=kq`9c41$XGZ-cyV1Sb4yDG8<iPVGH|$<GS|#l63Va2{S>IOI}z
z?n%TBSiyF4jwR$Qt@AS$BDd-Vzczq$4-^qQ?{_(fXYKs}eR$*4Wl9bcZtjLjyTkjK
z87gD$u3Hcn#FwqMebftz0<p`5KiXx*)I`1wHjhnNGkf>k6#|^-<`u5aICspfAFPl*
zKIDX}MB*uYb`nMRrOJAuQ}aSmxYv*X<Ot*bvI<cut?3{R;WF|8K#1&#>i$wYoDXhq
zTaeijW~NiRA?q%tLWeTNi#?OUPXiq6zmLM6Uex`_RCVG=Cm!C~p)ncMdbpor<Z8sr
zCn%(m-4ra?ctw3lRbiZAbJM`y@>g{|*BDm~TvVM3?cY)HaecG)fSZo5*MI!{(yOhm
zp3CgHLSN0~-HKHkq-;#L=Pw1%7(o=16KlRY<j$2Go8Yht3at98Lxv~2d`tC<^9u^^
z!D#kE+)9>^P5oiuc=B<9%lkGRzSm$qxCJ*o?s`2D7qnAmmpjJ?-nEXOSpAv+C;<ho
zt3-1AYpamgkQ{Cg*bVk6(hou74)xsJ@{|}IE2c}x%!yF^d@z-)bv}+RGcye+y9zo{
z&9V0gDJ2vTl{&2Rme#;!Fx=N{)BxOH(vNfKQKAE&`>D|@H>Q032%b`84Gt7Nf>17<
z(4ig|<npn7K}Tjb)W4plRU}8`a#7<ee-|qJTh(&YRk+$2XCn(K@YR7HXXFj|a`#Wy
z!Jn6MSf0p?ZB8$)O(lzmh4vog?WU$wAJ-WRCyC|Ty7P5ISl^fNzs7`G=FO^1La0+3
z%NlEr>4&Kbuw7=^ENy(TbpDge#k&+%2fHO6&s&0$z`0F~406{s4n}VI2o0w2C?0Nn
zg`~IH!xP|Y(O}8yff##i?_aLs%HC!UUq4khWE;C@KJ`2?<UZL@#PJD4M$m$_S2`yn
zV9BW%0!{u;^NRm<J8ap9U=b;FT!y8&QQEDO)YKe-7!y(@zphtf?<FxIuXD_QJ0W_^
zzg8FNtS;;6cr1WPqv{ml(&=SGSEDNt=B}~Ok{3d5C!@mI+V{#P0a_CtVGt04UsMF9
z(D`rZQU3Mb5>aASGo0lo?)@R->T)n2>z_h-6-gKBMK9Jh-2O(<D_Iw%vM2IqQn%x)
zt#}qL<U-bn(|C42{#qg~s4=&9Ohd#8y`U%BY3m@zgLVffiVWpwY=HBVTE~WZcQ33V
zOFE#k9o<A?gcr?1S979d^~1cO=u}R-FA3E?--t`Wyt>4bBTzV|v2_VgHYf&*#C3M4
z4TLJel~Bvle_q=t*~h!UR!kZVU|~r@YD}zHWL@sSth*C?ykxb(>PX7m>}Ix;gSa9j
zDGLhCZ<*}Z!<Dk&nQ*(!UoSB)GLdLOHa-~XWs~`=l-E)ySGB|E)H7EtSBk>*r?qFv
zW2|c2w_$tgxg9?ZofW798~jqC1N#&ec}Eb`@d0%<yCVK#sp~j8?++-c08bF>Y(HFk
zUv**))H!ZvRDg*MTGUm$_2Vu?fJ@~FD3p*YjuIF2C95l~ys<i#db^J_eZK}y;jKHW
zt~2<9s!9I)<@{dMb4Bwe+8)jXMpbN(8Z}sKUb)%=#wit^<_}dj33*-D{=Z&E*~xx1
zT-(qy+6TFdo1R{Otlf5%d~sZ3arVF<l$71M9@GmSaQkI|21b%V;QF{;xb^o>CT-&Y
zw-w=>{;NjL%ipX6<vok+8$yWfy{0mZxjp{z^@2~*|L4Uszp6C>_LO+`8v7<L=ss`4
z66`8~4oaCXz(!;H=~)|9ph*3U|MIq}7oxl9Hd4u9<w<pV&Wc_d;H>)3`N@^%m4eb@
zLM{B&$B&FeCqUp(jQ&&44uB(ZxJ2}VAjM7R1j={*<16a7W7+VZOYGtbjjc=ew`GS;
zF~!KE36WkyUC*ZdeY`bA|NEM`7+HYd$p{;o9ED3=G*M?@{O+}X1vy7K^Mz>Za`7mR
zxpFX(`ag_Ef1Wiy?|oG^Fg4tSL_@33mT6@ItjItya{uUpDcd*G<jYiL7g#ws9@Z?^
zBd+#94Ot`K^zD&lL@=*bt8pF`+BW3{mLlc0(_A$stqYL-WPI>PGs%h1WM>w|5Ap&N
z0mDvPAEDO1sJW%>r~cOqln7lw@GPwOvt$%cR3-X61TfZm#}~b2kf;nDU<^d|05i&+
zgSQu6tns&r{|;uvHi?uzru}#h643<|A#~B8!Umyz5s{1aMu;uonJeen_Z8ZzPk_Vh
zNh)gFMRpEgOOb#_DDC4GCx)EO(%VyP!_yqTwsQCfaA!c{<sl<0uuW@!+M(S(Nj>xh
z%mV1fDfF{=do$9n(94QPb?P^PFb|X)v8tz%?hb5EOrZ)~-8uz|hO{>tl`3yepIU(X
zhMXV(dBnQ>Ovw(z=6@N~_CRLdi;UbITb|^zRqhcG4h6e1bP(a7KCWzI2aUG-m7k`K
z^j%8vMXVFDfS6$P#j9n~j~-RF^%frds;wh1+alJ(67dxFr40!_fPhmEj_<Kif<ED0
z>n24!)4}8m2({Bc*nH8hiAnqJy^uPXz1rjbf?a^qa1Z5%7InsgeF0z!veQT|3!6}r
zQ+wt=+XY@5jR{Qj?c~wroTU)xPYHarUQ9sH0VuOX)zwnf3*<@qJJGrVPF1nR5|KI8
zAAL+RR_9-dAqCDNvoh)y!0bPOuw^;8AjZa{!j04NInV???`!&D<JrbNjS%RGOA5#1
ziourIV1kY8is!BSK=(88IWxw>qXJ4*b|VjrXb-Mw<Bn=wo`An~&5aU1YkhU~r@y{k
zWs874e+&PVJd`}h;dNl}0C*kkDw~E{MXFi$9=Q5Fjg55UX#2*#D49?xX!h56d-&!S
z{WimnAnlWX8|^+HHZ*<LYMh*Zx{K{M+{F0O!#MK+uHgnwO-EJYa0Tc!K(Q~1FZ`_E
zYv5S0=FThJ7>@5eU9_82Ct$3^pNu?{m&<7plN&VVXg0nv3e$J!xm7~D$=WVnJz%dq
zxBEUX+;|3G!5m!TyMj)fd%_(WxgWU9I_kvotv%QNF$XmF;(CypO}L5E{vrNLpNt0m
zG14C5XUV4*CZp>)72JP<wgC_a!YqW3d}uYU2}Zb2Q!Cbat_5(J7?+@5MT*jt+;)X;
z!#AEoAr1PBkA2%K6kE(_hI=9laK{jV@YFtfs9vbC_Kt+aZC-^GXxo|2=B~7PBy;Hs
zbek`!(Dh)Lz`%L4F(|CV!9+qk6Tmj7l9;zAF$MPvCoh1@amI8&>;xELFMw6`Kfhq+
zhOeSGNp8ZsZjVe2Ub0G1O*=tkN#gpeB>75nHL0B1Y79ngF!J82__T&XHoU-UC+TK?
zof`3JOt9FL$U<HOL$Ix7ogg)<tIvx2tq?a)FSHp55<zvb^$~I?=jDFjmJ{FXn#?ca
z3zy5|#g}2-cWUO<<nJ-c3a`LTG9F?PaSZ?!e4uzB2`!_`0$x8k@J3$bW*P$%Asgq2
zc_98-Fi(+()Q<5Bobwit!wPZE$o?SYCbl_*^ZcNpI=tGSksgx<DoZB(-mm+ViV@Fd
zMoTGxoMViSp<-5)_>~RYi?6BUYoI+ER4B)I1sp40#R9;(`|C_**)|&_DPPN>2<r`P
z0s8^b;pLFU%x6no^hoS*P^*cs+q@^J05Mee8L>Q^IwbX@)r|oYo2&5T$puh=zdni}
z7jXkdD8g~T3RJ~?AH_-@u&;6Bc=ry#i#{7x(?#F6dH6Pz-pn<2Y#|J8%r|`e?M&}V
z$Y!uVXG8-?vTnG31$dBX0R^)6p{(0beb5iR;O|!83HcS|MIflvETLZ~VEBC1?#D53
z;H~>|yWC4v#k*XoW=(nGI(1cb(#mc)v4Q;r7+DbaA-VvY5OK2Q!Jty{k@!&H!)Wa!
za{#caiJphXNJAUbp!<k`o5K?lVqlZYa5>SM^F|yDT2NJJ<Id(l8T|;$ODg`o&rMvi
zs<$dJ5I*1!{9#D`jv=RlN7CJT>Fwt1><4ls$}WCN1V-=OaV#ki%~oCkth632UZo95
zFY6`!W3KWARe7zJfAaZKM}*%$XXlXG!-lQLKf2<%!?#)fqS4?Q1)BHaD%)e`viir%
z%a@y$*Xrus?%lMPG6DqAJ#&T{1Ds6y@TN&-`j%;%GGl})3`u1zSPZmPNkyq}FaECn
zja28P=q`yV|Bl9Y(9FoAbb8Esh;Qcm;*cWL8U=LVrmffygbi>IjU_y{I=@_N@CqG|
zFX-Tfp#hWF0bfWK7S%t$X)N?cU^n3`l<c?}{NZI&RT)o)>q&7g0{p<XpcI4Pmfe5#
zBQ3g<?2tNBw!t6DVAI6FSCJu%3UeSk`hDj)P+(HlKCeWuRe1T$0y|I`J4+Y|qNqGB
zR|qWI{lbxRTB|)3SM*m7_@zHvGMny$SK4&u8rL9K*}3uJJw8Dfvp&1&xbRuueT>!M
zol4|tSr*R6?u7}7{@<x~CWkRjz*leOw8Om)>8Yr${o1!Y*=HJfU`biXi@8P_t)UaS
z)Kv0YNxhQpR5EZq8yGImmxiBL73ui%eXoT6ol6Qp{P6M~WFpL<7)n^}ScwZ_pCy6_
zv9~24d=u~mP(q+Ff2!eAavJ1~|6OC3x0!g93gekwXu{`+k4q+wTe)=^+Kjv6w1C<l
zjnW#C`Qa`XuZ;__NQEl-ckHH~^zPxpn9!4J%6rBdn-Rlozv;fjE2ZuNT~S=N<vI*X
zsuT9F?wtfkOkhCkEy-v6vbZ-e)KRw-V-sJQzR`|i>^p7dP+FZ(FUTM#g=Sjyk3rp+
z-RrI6%(Hje(8{4eq);d<`i3`(F`#@dlu~E1(J~f>&KnKH)S3ucmYvf3=Yh!`5;m=!
zyx%!@=!>~kyu+sEywv%+os@l!k0G(VonIID3Y0X6_0SAMNu>IC!$PkKrDHEU^yzFi
zKl4`p>Dh`VNf6VXAg1hn0U2#$V8!o~0x&B0Clv~#rDa$ELIT%TM|<=~?cAZ^4`x+5
z#FTomK3SL;47ZL*y0SiD!H_woK_U}41%@%c)c~!z-LX`KFUQSZSi-@rQ!i40gZ=(O
z6w5<EOEa?<5jyq9y1m6snezZdH0ti=09-F4p=M1?=~#qS45NoVLgQRxE1|X+x7Ug-
zCnIMY;V~%GmVllR(?2Px2;=P8=kw2fkg4KvLFS7G+CUuojY&D5L}P<ne#1MbBNLaF
zt_Y-*qt6X7^LLaTSWQ1ZG1&q#EMceN9ix~CN;9Z}Op-FWhXc0h6Y=~$kIcehdTMI%
zkijNk%_gKcJNK}$WJp-G9fXg}uF#s7rxpJ&V>H&vpc;Y$47nhxJFsDM>S#_)k2}e-
zfGpaG@x>=7(CNlM1Z@xHq3<&|kmIp?G`hML9ugio2VS=RGdFcW$VZ&ST@&kaL7*84
z!ya;d(IKX@E<zH5#={H=ys(a)WhT=PQYk|2;&UfHi80O{5dH8TV*ES#9~eiJ$}uw!
zpjsH{g2kF0ho;qOhOWTGF;4|+$Md@d3_W!G(D!_EY`x&|MzwC*je4x60T0ny?^4fD
zBvRL?Mzx5=YMhUzt}|b_gAGHRnA(INU=GWURbNtCZIN%DHIV(xU7uJDn`kb(9Y9{I
zp6rh+{hCbi+dG+i=P|%av^x=ITILL=Hh+^fm0{tj;pCs3+Ni7KZ@<rvW1@G}*=*u8
zH4QzM5OT!Ax{kK#zGR-#db4wn$y#!Ee~*iI&tJ81oxfoUsEHCAAN|S7M8w5Y^G)X1
z8YlVBAHtf^dMwh0=vfUcCPv2zdBLbUV=TA+h`7TM^{3RiLsg23erXgBp#B<ZFH&Bx
z>_qReDt&|0Sr%)0_`_g!jJare({4~l>te$De~6)%AQR6EaIv_Y=gGlQTLEom*H`an
z;VJX{AHoBxI)MW{PUDxK<;uDl4qJF_tEX8&^I#-aBY_W_5<Jueh{FS&U)mzaq|3K8
z^I=Js{`nJh*J#LP1J}qv*oZGR=QZS;;cLjZ-cG)gXQB6YW2A^RXMpn*TiVnJ3+Buf
ze2aPvQE)p!SONZFihsytY#7vj+%=p6u==ua+Q-v&Eg6{!%-FvE*Lkp<){t3I7qKy(
z1C&F|euM?I{5oW_F182~Gze&rUJyRg?+XCYe=a_=osk4&R4WA6e|`}D5A8X}7I)jg
z6XxL`Ng>@&^x<dd8T)nlG{S=z<hB4uRfzQSD=JzOcVGlXH~?Bb0un1=OspL=w8NqA
z`0TJBO@V>lD2~l+*gbEFI-lTqZehenl_&po-l0LCLs<rJ%Y!@*b>XhB;CD^DAl(l#
zvJ{A>+RRfRmhgE5uMm!%{LIQeYjL)@r^2j@5bECEO<RlMo$k;Rv@~#E2$E>{9q$ln
zF#<zP89rh9^h^y6j3}cw+Sz_xZDFneEGBM-_oodB!@l<?Qn~#4s|0jLFi#=Nua(^c
zg=i=&@kva}7!@$^MOBR@+t&{JPI(^v^++EMedP$3JVk3JBL%mzG(t4ZTv&|&elK=E
zAr<V?h|5|-hOh^YUYIHVI*m!4d$m9J*_C4q0{+Np2FuAVm;KktKXDyf3mFM^XUP|R
zf-uY&wWVSz=7lqNbVzKe7>7EDkP^cz#q@v2YL3x=750}!K;SI)Q`E>lIrjnIf1Rju
z%JZ$5oG8%f+}E^)^e<bz>&d1u!|dVWEl=Zs@uzFZNG>hX(lE#~55`^z|NgVNmo;-9
zfWBkQ0zUrl_d%@K_3MWNRoP7WVSht;MFA6*$BjZPNjQqp_zl#5>t9}`jR9SX|Ikc*
z3So0IpmUjP{P>uy|NHI(q1S=7?u$W<JEBi9*i!D=v0`V$@TsU4WNii=9t#ww4O<@k
zJuHqFCtK2GDU;ypg)gX;4>A4ZqenQ9HnaPB$xwQJs^4t&<BEa8x!d2otYcu8enLM`
zO8+4A<d9x4ai{=nk1hqHw)}?>*b;$AoO7#vx~0-PqC1;`{>eXu&$DlxAR$wj(5#L0
zI`jdpRI7`F!`Xgr?0UBnHv%#Q4)QLbNc=W5c}&-oorTsE()6|(>z=HQy^;Yf&r0j-
zCMm|)&$RX+s15boXwW(M9%;Y0ALK)81Fgce8zAP(wqtdiq2{gwGfK1Q4E$?M_CZI~
zxdNn^)pj|=DRTn4X!veplxVT!NQzDwNc+ReZ2?f(5FCwvHOl_re9>bmX6ZK&1C%*S
z9J2!L+<&6;rCR`$)2*>Lcy5yosw`TBw7o^}Z(I5wGyzlEC~u>tqAMw#n^^kRSjhw+
z8yYc9$rZVP2$S}*kPQmMKhufkTQOM(o(}xPj(5-^OxjowJL#sBZGYr|ZDfg0>J`vo
zWsE$h#FPkD$pFeXpkIw`mgvdxZqM9)(}8ujzHe<l^0+d8CCIx8g<+}eXqc!P`qiMt
z20#w*MD|;*hX_c@FkFRf1QpJ5#YamJ{Dw!yw+j1PUL+eRLPk7pIWY(wLTkufWyR{m
zSs28fQv1ZLTw$R2az8d?V`se4l4`j#wC@ID{+thLMd*91fMEFEmjSoliYu}J+^LSA
z3nC4wu@+p3qkw=iw12!gZvTvnnz+X+Q<X1@685+#*{@DF;*p=ll9e+HwuGWKrZ5D7
z9+FewB-@nPBjo!&OM+B>d-LIMmKDZ4P=%||gF1UjwZpfo?HCycjY->}<)~Nt90!z5
z-FOAgF&c8-n(KF)idW^!o(6qJ$nkv=R(7<|wszT{8*Uf(LhF7AwgPdG{(_!`9$;?U
zN3<3M`lyfJ^1ka(YB$-tst31)w_RX`=RxUXHy8Tmy!{<0DhO{VwIJT;AOY0%xdG)D
zgbSK*31;!w@<{S-jo-!u=VxrQ2t@(2Mvc~i0CanpF>f4s+J~KmKl9$JL_{&WbOD;)
zN#{Q6ZAuJ38HemCSfIeLPQX-JcWa;Wg@SGFlXbtceaLoLq^BTX8IV;^^nBgn!ip*5
z3%Bs#@YAr^Gw{kz1Ajhb>|_r6K#Cd%32a5wl*@~A4reO$m#-!$uY$?p0)jiG8?h-o
zOR!Hsz=C&+5S05Pa@2@GdO-~Vk{yNVC$&+)&l;TkPN8bFkGo9ho4Oxc+6Z61u&5F%
zVZ_XEb#3+x+_aR*);iLNVIHg{cpYMe&t9j>Nn64__)Kx`(7OUNH?SC-7GqUmEQ(G<
zhFw^cgPj8flI6qUpP&K?VxXw0_z-oEF_h}U9U%A@97vFL?g(k#zoW`}Ng_ZNor_aY
zQIilw_Lgk`exp@!mn4*ID0PB<sGAe3GWwqPw9-3!hqqb!+%Ui*=fJk)C^jZ+16scJ
zlI9j{IgrmOJV6QJF%L2|(!vSy+o+0t5EkCKhrp^oyv$f*i!)IhNdtycE<AmE8fx!A
zWKtbss)3@g&FL!_%_crUjJ$8ofMw%=Bh$fO-23|qTj>wC+dIafOKb}0LL%l*TlzxF
ztJaS<Y%D>qDq0N|cxZn5i;bson&m>#!XN}%W2AD2pCOMfo?=s>@3VLwtA=QY;~3pJ
zgz!TV&w6e>Q!Xv~5$?)YsDEw(Z90U_Tz@&uhZI}uvSOw<?AQSid^O?+pVZn}Qe2SW
zip&Uf*TVjRGz?5D<sm2SvJcn6r)JUXckvmXLr?lt=)PPZ$mh|<y0sV93m|g0H?KdT
z$D{*{ZpXpYGxb&MW<J<GWfPv*c<ydczAV7o1`&nsVw9w&I+J>a6^_{wZGObFmg$aj
zUm0Hs`8r8PuN~raYy^f$;_0Re`Zb`2mw7x*SCT~|kkq0@LZX`Y=6qdK*VO6y*V@#~
zxk?HNejNc)g6EG-aY0aU_tD9@G@^ZzAL#Sg#Y9=F%>4NKXr}wh?Djcgba0nc36)$`
z#|f%@oXcAcZvwy{pzwftln$)ht-U2ud6%0#nsHjgGem00i$?MPq_y&*ARKW0a7uq^
zbT|%0gOvk3DAEKKK;c5Fe{2)lM?q^>9(8razl)1X0p3Q~9agk4&UBV@z1`W`$$g!(
zdV?wwvgd7#o%9H&?#*4V+wVNT47o$YAT!;j4;%Z#V^UO>1-8M@ptH8`VU2ZZTU^Iq
z-K^pmOqta!j&2NwL-AlbmMcbocHm>?g({;0ZSpBf*}fthup~gzXa|8Gl!|-hev6fH
z@j?H}rLLDiZ_r4PX5b}x0oonim+s<O51Kgm)7wx7Ak-61@;$T0>|OA{9?xpYMg$rd
zObjOZu2qV5uG4QS*usS+so{H1ShVg-So<P-jg2}-mPMu@nOlfzKy``6vy505UPdeY
z2BqEY()%5b)<bZF<a-$i!m<FNdtTEa2ktgH0f{HdOm|TQnJ?t+iTWxcqr1J$zE0RC
z&I;e9K%PCA(cXx9!?u)eM@g6*rn!g~3t_!N)rN~iw-$MSBQ5x%3rq#jf*Bc3;A)I@
z;hzcgNcn7tZc9KYH8TaqJHE(2N_7*aiCa4ph`<)i@K-)@Gj|LGC7&0)N-MM$qg&}}
z26pgTZ`chKLK0Xe*LPRAx4s&uVbGzE18d9HPWfx|)3TU7-T~lo^!1-0ZZc`#57ryX
z$*qSo!)&0&gXQrc2siQ(Y5}m=kfQr#dJ)jN?<v+;3(hZ&7(p*7=ox&pD^i)$o}%_W
zoe^%CRYn3YdTBN5X97-D3&f%-gi4V-VAbB>g3={+%nZMct6|5Qp;7Tk)5km@z($3m
zV+u-7O-N!jIrtKA0VPuzi=D$|MJbod&U1fqxa%;`zp#e&qPlvp(m`{fs~*6b_2^=R
zN*vR1ZEqI$;5>~C2B55>Xrpft{2U1N6BELiCbCG!$ZH@=>d0LqSY@`>z%N37$Q(4l
zh>u-n!t7BiKepvH2oez7cyrO%tZ=!0r*8_0Ctt(88B$>Uw(IjfYHz}Tpw>T;OU+c!
zTenZb7`Z2cN*#Y`Ztgj6kZ*ou2Jz;WUFz}x*XF(gB)aPdrZPA#R*r`Hhg&Z1i<?qH
zP}YDif9XT31qcgaf0`eGiNROkmWYw&&*?rBqH&=jMgku_dG7eu$cv!T3bf!*0sv>=
zd(Zl38D-jq&&nMRFDm3cvv&FJn3z`wU+=&Q@$DrtJ(>0mP11wd?gcCRzes&t`v64H
z<=I<|jjkL&;b;ZLTjWYyWaf0p(k~=;)BX!afF5aSj94LR#s_Qn;)wv*);g227;KAo
zjd#r;Npo%CDFx5Nlup_ghMywnNEhgfgs==*x@5q15)9smBS=h)D;cp$JXhZ~rtJMe
z!`%md5GZr8^X91W(C0?3JNt)spDZM-1+gUj`%7|&PzK&^RF>stHcem3_<6Ay&z<M}
zJAG|exqk2k%(uv5Bd94D3#WiG`fjdBu-YqQm^2~SaScnGB4N?>8#q}8Klgvd%QvAW
zZOx7)Nut;wmkc5lD8zWM!)F+_Zn-<iQ)#e0=Rt^&5O6^L;>CVCE>jB98~gV<I#wO&
zhk2rjfuI#H;|+2vmJgCkp>-py>XK(gEpJr7`N8C(o2Fm*u`T%O*=l6e2?d{zA&?1R
qbotQ#|7bV_o?jnMCseQeSY?$iBi@*mw1`B1I(k^|P|m?KLH`ezix{f_

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide8.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide8.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..954d34d32e3ce15086be66af38f56fde818c2e45
GIT binary patch
literal 22874
zcmd?R`9GBV{|9_YQ3>r%6gm}^aFR7Kq*4iG&EAG2dv<1~R7y!J*-4RI_I;dGM3!u0
zpRw=8HjFW2?$<T?ocr<p3+^B8AI_sQUGM9?yq4$d`Fvg1)ziy5n!K9@HzNqbd*S>U
zJp|bR|GTbn<2v|5P2?g6{K4U-r+FI5sNXvVKi1o-YpWy3mk=Jtb#&+^=kq3R2*Ov1
z{>M?}^!XNoT%}w%qi*13ImPsiGNfLouoTo~9ZoUjFLb@JdX0^26LYvDDEODnHJiU2
zgY4Wc9djjYIp9iob0%1B+ZWL@#}5l|6zx&FymM!%S}wle%uT&CF|qe_d~?F6`px8r
z2J1-qgoZ++>RX(Z|Kv#9r!)OjdqYmauSwmJkaA()%sku#0rnlYW#cv2V9f)B3w|8?
z|L8}}cL`Dt{h5O0QylE+tq8{@6B;Rml;d3;dpEM>EGH+T*0H_fmX!ek^PZNhd2Y3c
ziwnVLAHx4E{ETJU^(Xbx>9+jl^F+bF;4GntSZzd`lPvER1hLja-k6BDWuoI<#qd89
z^;`d3IFGzhuOqJg7&?c%k&{d&-Dck(gox>zS`vQC{sVjA=eXn%Wyd_%T!A2$#n%6`
z=XWtZIPpFQ61R8bwd^=as*pCiz2QyQ^fIUZ{K(&Qtx3}A&yUA&<rv}54<opevh~jo
zF}RZK{O8Bw#%oPUjemap*hs$Ry(})R9jUq`Kv3MvirBF;37cRCuBzV75hWUrE#ZCN
z98Y1WhgoFa*G+j(6_UI|o_doi2V4Fajpzh@#!7HWj>&t}+OG6|xUXx#pa<u(Quhg$
zPrgCV*rbDqf1RaIsZ7oClrI?c&;n(hU+DcY%^p#_mNdE?Iq{;|ni*hDsV{H9jE+Y)
z{n=y6CbHh0M0C_(%bFF(4&k+ibE!h<Q9_D#e@0#0L~go$oq0~8flJ3`&5AG2tt(t#
z{~4R`&WuyMXzf(z_i>otP5(@~J=9$BGjq)kO=l4S!_)}{Qzcnv({Y6e)P*~@ReW#$
zK@Sd{;?gnxGu`fWh3Dt=2*0yCE9}zNZq-vq1T=)<)<r?ZvOhbS+eCg)>&rY>tH-HH
zU11F7{)v!H;&`5kSk_ozE7S28Tz@9$H8Yg+Vz0_>mhX|=B@yByy1z^3vXV;p57hqW
zWjM^$;;|AdJqB^~;N5aFjlbRWJ1a2dOQFC1*x~6RJkO_%;F2=lsIs+S5e;AOtj=g&
z@(=*Ke)@7h$8^oE2VWxf1QEZH@EWaW#)RJ=xP;|aAJZbw=iend<>JcbPV_0MD%m?T
zqBsz|-i``yqu1;;X-{(ooDL<|#ADH*4|&d*cX(6mwE-OaS_2u;-vlS06S~R-c>44=
zem-#!_VSZA4GWWO-(Io&>LnQP7nczMz{_qE`P7r~f;*!@2V0zSKgwm%0k1CVAO)?N
z6Q1-PO1n%u4<+4il}3=4>(>=(Z-;nz-B>X$0EXDe(e1MdeD%rU8kGk@4hABQpA29U
zy|qcUpG6AJ1&u41&aUm`!U254)p7WJeC_w9gZPF^pGoycz-Qm>)~0xasB&Im8u3|#
zgO7JB%zPB;PkQUbUhNo1x85EIkVp1wL&FirS77km)3WR>Z`VO$Eq}o5*tOYzisN+;
zyn%CD@kV{7_n5MtW)1mcu+}|+2`sD)Bgu93l0+gHM3i&Avzv@L;df_m#i2}6V}TWg
z`U=AeSYjIG)H^Q4H2BZ4OcowtR(~$&j`%%l&iI1W^uM$6F$3L{o;DJzHw?x$6D)j4
zUFja2uQFuzu~vL-0}K5rzP&pPr|rs?dN8z~Bpm6TyNKiUO;uFoB2#dTTIVM68;Lh`
z-U&#-n)}a6sF8VY{~FG|aSDet+(^71yL@f5J?BG|!#fX(e*+~K=TuEDGxif3nNWxZ
zAK{N}s}R3pL^YG6<@0d$&a{7&inn3ejqRG<7Nv7xT;bbzVG3qTkI<YPUq}8Xs}W8@
zkE2IyNV$i5U=&W`61oEb8FTBa1Zm4w$py>L9lPb7FN}M7lR{WJD@(%?{*jxl=0b3k
z$tYjzWGlcOVGf2+O+Gdu@8f+fJwk#qOWi#6kxN?%<%y*-SvLWvgU%_oiEJe?PY-VM
z;}jrZmn>()rEvEgIV_h{iX<jf(FhbxUGIE}DNIThETmCa+D8er8P;u5Xrx7S@{Fz_
z^LLS%IIR?2Pn4tkJac&R8<!N8sj9Hd;AeO|A@#R9eeCfiohW2|nr()PC36NzCB4bC
zLszpZDXz5grxozH)TZu^z~Hb815iwEoh#y=A+RU{f&3BOfSA)e3PuKR{O~x(hRl7&
zIBabeW+hA-cXGs&r&xM*i1IEA8?ZRlMd{3wSUmzW$i$~{Dz;EK4>aKJEi8u-tvZrP
zk}6^<uXhB8n@Ef+bXHPFdN2#RLFNQBcwYZ-O7NKK^q&b3=`L`xV&p&;ezaQVn(-87
z)epal_4Z*{l`rQJuuS{tbJhj-%`6BZvJ8{iiKQ>UaoZGbC+9hlR{NhxzrdLz2!27L
z8ChLitWu^jhvWV=ez*`^ERo6^l>vdXIP3NdOG+I+wbbo4%g+lL>N@mS#TBN<9<1h2
z#fAeDnKjOw1;yBcki*){jD0gnE}mi<B_31t7#4#z0>(REl+8mf#7+=)(X&S0?%U%?
zZfSLDk7fP462vOfOmZjrFMY?T07yi<Ge#8BNw-ZL$g*utO_r;{UfTd`WD=gQqkRJ`
zDdW<)$$ZRg4a6zGs2_AW1V!kUTG9HMWXQ)*LB$l3G++3#w6Uj3Z>Xe+-Ncf3GiG0b
z&0LgQz0wJrUr%ta;h-%?pCOLr*aYi@`yoUMVb#BNHvjHVs6W@AmCsZiVoj243=9mu
zVJ^Xb?pR~+W`Oujy%l5Wq%74~TF?64)s$qS6;5UK71+D!$N&eUqRh<fDh%qXNX{S;
zGo~BMEj~;XSd@6A8fU)1rNEl^YFv?6sO?%eIg^U9mE|gPZ(MGRwJME7op9*PrEvqQ
zolpA3<vaoBC>B(N@}6HxsTyd!lmZTxla>93F@xQlv_66FLU+?t0RWZN;jijr^xEk@
zHKVw=m{!G}V+)sa7aNmlJrrk`&IyIPrgf}KG`Wg@y+@)2SPvn}#@^;3^-$8ZTsZ?N
z={Lw<x3oC5llerROPu&umG&?H4-9ZdSnDW$9>w5}3X>F4R!~SeCgazwXSjOyY3f<J
zR!_hmxm@*EnUDQvrE$_azt*o*YpQN%eP*xm3Xw90HFP*_GqIIY6ALR8f8C}`j*Rs5
zU|c7Aj?;)hF*YGO_kN|c@?uHeZQiJ#bl#He=}KGi6|t*Roq0mCw=Xj|xgn_a$2kiX
zq1@;4cGKw&v@@^aT)m{#-%3CE$lzMr&Ro0MhP@D%I%9m~8*ON1aDEJ90As4+!wsgo
z3Sv*#>M(Ay=L+HkJP?x1y`6+Dc%DMYzRkOL^?Q8>xOYS>RwU5ofeISty29vmcD!T&
z$=t2QXYI+Oaw7yGnFFa*p{WMdg>psN9b-2|R<1JMptBks<t#)Ai6XDL7*$9<eHfRo
zJ8oTJ8|%MV=rYhRqKY7*c})Tp7tq4uNS;e3Wu;SNpSYcPmG98Co5cW|CBMr?_^4f9
zC<>|L#Npf<$wq}-QmJ%a->Gpv3TCCu#Zl}RizdPvi1zG#idzpQw&kKeBA$*-*!AIl
zE&LIjO0Uf6jV$z3FR&*%{&>D;HbfK-VUxUxd<$&CkazZOb86R3#GSk!#(^8v5Sg2U
zYI{rL<T}Nl?D}P%PZe?i?9SI}dR74rhy-Y3=x>DMnQvaFxD3ae;x!+p3N3cF?o*xV
zn(C(?%ezb%iN%$N0h<;*@mjdT4ydhz9XY>+5*m+-xDWVD`+LUs_1xcCkr*Jap$AXm
zn+*xSO^FrhsU!u$BC)GCorKa4CBnOL>hd>ADB}hhVqQB^UUF4^J*58rB917i(H@Ds
z-Q#rs!qtXsW5RXcOxmw1jI+)r%0~ai4qIYt>vzl(h+3nAviJo0)XBMwD5FcYro9#8
zOCNoq2w$Xrsj!W9Z>v&W{1U(oA^wa#_&5ws7cTl@cR0~e5|}~PFQY}KSJcO|E#>|z
zISr~SMI$lKaJ#|OTP~O*wh7n-Ia0-VLHHxvrI~6%;hYm@M^LUs$=^K!CqGgstW^qz
zxzIe@@Hbdp^5<sqF$f(S!K=)eJAEAv`&OUsGiKKPwkqA{x}YIKzg$gL)Z-aW%c)QK
z{Mn01B@V=%g&=X;zO$kum2^Al4*5hniIaHG>+`R11^U131x}>pkyxuy;{B>%p^*cq
z!Mk2>uP_0$6n8+!a@S;I5A4?t<QEZ^p~AF5K5|x_1x~o>q7f;R?1QUekad$6&mguX
zkag`9l2+kFii<9WdC$SY;d5JQsh}TNGLRlE?+_uR3cd<I?2m+LFwdp#bHfFe&l8mL
z_lq+kV(E2X?1>rk4H(SP_QTc<1`9H-8X5r-P;leY-YVj2ld!&*d_LYJm}$0Ixu45K
z!VH;M+xo2LHi}DEzJrj0W4EU2+^d(_l0|3RW|wq~4GeyTpRgsB`_Bphp4OZ@!g-Y`
zPjF2ou_&XjXyOBx(5kclc4}&B10?2Q7u|)92CPS8oH~H(;%f-tw>hsjWYHb4;|1=$
z0CE<D?VXC)ouDxdIMT3@u_(=~F4?}GJp`y!+}5Cd;KOB<YT(x$xdUZ=I0<%-Y)<mx
zU<|&fI!P@nO2InlNZ2@mtI=U`20`@AGY>E1V7ff9pkQ2sj%(=?>QHWj95R@+$LI8*
zQNbRX3$v>m0X39AfpXQi=9|_jR$FqWl59{Ch(CB(8{sy&gN9lQN_Dp$;vtI)0b`B~
zM>qO9;VsABk7a17=5LEFuzrvJ<>8p^+|Dn1fkv+_c|o6!@HMsnZ%NVn+qYGy7Yi0r
zsmlEgH6O4OY!a<_VG~(Y5}moDgu3#R@+;k#072h$@Gw4LPywuA!)CkLw>CJx9Q<$Y
zO<{!yA7K8U<A^T5{;NX0Nw83ox^j)_ooW6FwbaF9@f>>(Cf(OydW?J4hPP2BaHQ)K
zGxMjJ)2uszQKXtC_Sts-<RxGG`px8!6Ztb%v(+jYs~d%$NAKWoj=gX5@3%Ug&_}ma
zd{6}+fO7diTZ-Pf-g(B6lca|#hq_6ta>kfsW&PUFj|cDpvpZ}B(*L{r;wLeDKtFKv
zKX+%!Zg9Rl$~CQeJ(8}{%HGJGBRu3dXWqT#HdR^;h1eylbo3N@qW&CvcYZ)^y`Kj)
zncCbZ8GHak{(n<dKhEWHewq27ZNgvmBYeQ#!8T{9E0Yf{pa!hQ-{W>}3`GYWOCjNk
zr)p=_IZ?zM=Ty9H@t@x(f1XCTZ9D(7>l>Nv73#JBg~=^Lk9CTFDNh$S|G%qeUD`0i
z@?5BXf1D`a{NIgTGP^3&Py81!GSc=Od!usyvxZXjHH2H~zsoge<T(`;dn;NGD5Q||
zwb6hbI*`M$_Y5z}P#<46{!H3|20uRg^$@~$nv+$Tn|U{d)k3gb4v%6ev(2X_W_=1D
z&{#nVHzVY&zCm#k0{f+52R}LP5v&BJlu0|bPEpYTmAfz5`S#g$6FDv%jDjFYxbZpm
z7HgBSEOyFRr<Oua*pcT_NS}<qCBcx(*cvH(zyL2p+RDyPi!HMba|6XxzfDfE;V*I;
zjdgUQ<!}-i?mEY9>`=3d>LqgcEeJTV26hQxSF&|eAWiWyS6*=J9rvty=P$g6dIa6&
zWmTA%?OX4onTXUIy8oCK`*!0!8=S*iD9@VfNv8&N(I`=S;kt=^vdhCYf3xpFI{4Xt
zq4shXVV~qo_FlFi#Sh2`t^H48_MQrY*T6eyh@h(=Ro4Hk-6~#ViloAx5S^p?9OgzK
z*=jPs>k7dpcUSA6n7B&oHS13TF0EF3>)_fwb+xDe+#^8+=mNXmetM6aZH>4`dViJ<
z)v{k3RDb>`ASF1|IT3gTdMLcnd6TtkrsCRvCdg|6;DeltlWf|Cl8@L+tc4HKj9JiZ
zY0Cp{n0eMpH0YfY68&8W6JzVisHQxuN@`^^fQBt!3ThZ+^z26U7P3=k^SZ(>SaEkr
zL$*lr#;s_+6Y<6EEvfUN`1Z4Csr;<PNmUVlRPvUavNIdm@NNMpO^JhE7<AIS*M!IK
zksSn~JBuoWn$8nV(>;QGLwn10>Wx)h&?Ce(Zr@k&DTM?~j6`V`*|_Y}HXC-&OoG07
z|Hn$=uVv#Gy6nxUyK(~A$CWaqgG>mS7CqPRjbJvmb}lYB9Hrl(#xgI{=t`^)Z18;X
zkjvSC5E9I+RUjCZ()}IhmSY*rrL#^JQV(DLi(BQU*wgjaLN4MU7phF{nc~<arZ0~%
z;nfV4gj@8jB<xx}%MPbYPkG3%TPSeOV4l3I8RM(h*ZvxcJ^b}(J#MS9or`~7A7#z8
zeTNR?LBh_jE{60AOFjZc@Ezrq_;p$rk+AV%5~Fu&`FXBAak|^Bp$8<K9}pg7r{3S|
z3J<bHKDsz*CFb=8XXiE(Tjr3}>X0&}bcYB!Yy(sQsS8Y0XNTa(QKw1sMTel=WBh0{
zv}17n@O!8lpwI6Dp%W9u;-*k8qYHQ9zaF+m!cy2OmK$4Q5f!tk5$Mh{>2gI?D4WRU
zgj*cuDRkZypUjZA3pJ!S^M8ZPQd55?hS$1Ez=W@kElzC@BLJzUv?S~MokX{+wrL&o
zCD^=ItgqKtoVw7gTV)H2sg>_i#%qDLVo@m3`$lZOMgd<#rC7)&*omLS9fY;Q!p`=&
z?1Q-*v{f<nb!Sep81n;fAm&6#K>`)=SdqV$jSr=<^-e_l5=PS!$E4CqVcF``zc)B<
zA6H15yGaSrIPt18&%VHiI!>!*U%Rp1`SiFaWml|S>|_|O3;0uG<JMR~xaMs*gXo-R
zK2|xgst+>ks)dDxpELaZpy72+MJ3Q)d8<2Iti{?<$rQVfeM1GJW59eYN&hM$zOt_|
zn5~nlfcUFf?mCKpV??-a+~!Z4^mr;wCLR4@*a7!ge-<rzR{@G9t(ihtQaQ4b?7y=3
zeFWIF+Ug(%|Hg!HeLue8;;L^PZlVAsF0{jt@zt@XxSelIC`{HA^B3qc%;Heevvwub
zkIVUw@VPRV$>AYz8;zu&W{7wxwlZ{y>17_iX|brmUrO~dEI)S?!x6=x^J=Y*AhDp4
zczJne!@;I6jlDOLp?SvE0ep2I?+`P37|ZhmHvuF>Gdl&@3?5C8@9E(Ns+{1ZIJM55
z6=R?Qp|27(sC#<NJ?L<fsYnH-jgq8qK)8OPd<zX^cKGp|=llv$TbfO9(nh6wdtIz!
z1Ep66S2I1T$9^58cy<yLr0NH~>0?O)mF!767j}=z$~!Y{F>fo_(xrYS7A&zjC0*g)
zix!-$TlAzi-g{>JmwE(ZkHE|u8ru*)Jz~W`%IZH?c)KBVRsHGTl3=H5*9Z6G1MbPq
zfW|5IFS%8=kpYSIU32Y{IQtUO*Z<&t>2;vXIs~lzT>L;aAUiMhu2VdJg_H0nmpnc>
zlA$kyTSyS`M~V3k2;{hHb9*Y(g9K?sKSv0x$?dp<dK<5a_`{XUd%gL{qGx$aU8=oZ
zo`zPBx26v00|khym=j^&F;S+f_j-vf%21)|4TrnzwSmpf7(HKby#u=S4okfs$^}lg
zP`(RTG3u2m-vYrjYJCt}z}1{Tlj<~92XPo|LzI$Rzr1AD%kn;)?XUOSVz<HWzw1>1
zW$Uh}noyt{QA#M~O9U1Z#Ey|c%A>92G)Q}))%Z~>^bGGe3stDg2q`9E4`7ym>aH&6
zvT`9RU9yG^Y1t#uo|ZK6ejGf@!1o{ZLbxy3li##C4U+#2fP~%n#Rud;)ljERpH{T*
zbdlTeFPGnF+1#`XDA}N_^9ibN_e3T^75W)3|8?-wymy{NPPJm*86Br-ajNz0Q!Hc!
zu3)pqo*GIN)Rp)QRnv=9q>J)z<{qENu!gz<A6n#nxIUG@6T3W{5rAQtFo=lgM9|I(
z_3eU%hhT2-v!I}$>OwsK*&{-74q9h$v;-06?a!l)O|xN~_;tS2I)5#Z!t1nd;VTo1
zCFQP-!4~5JE#~<jy9(vXex!|eWEC#bz77P87-`~MyBW^}x)pA&<SgyQj$CHsA)?O!
zYt-$8XwSzBl9OVwblCei3IJ4xnewGF3ItN{>WU|B$uqM2GQ${0H3kcKgSP4?Y(X}7
z%#2{V=sn6+Bv+U79qMkyFU(B^Ts_Cj1hl*E(X~6;Z70WLc%R+jCbh^hsH-O=j2$CA
z$>dekH2FK_Q10VLjv&-;YI%^+O6@}BCx~FQFIMM8jPpFp+9%9k+RLMMyo^3v?$k{k
z+M2;>W7sPFIlOT7K~%X)!lS?VW=~x+vZg&yAh}U{iA&#DTaa64zi=vU<Sq4gr5;F7
zHH2o37IU4NE1bBoQ&T{`d{DVy;=Q4Yn_hmfP+qC;!t^M?%Du=qfTQ~LcA((!l<bHp
zC$Jb98?aML!<6?JHCj_Dz1EI+$`Q^J;5aJd<uNrKixpoD<PaUYxT8YdMGzX8p?$G_
zLnUSdxLruo<R@cGT1``Yn0eP1ta{zsV1L)wa+F9P%Gt_ch4P)Hi@zVudLoFO5i~zG
z_dru|m4w3{TU|m#>tY}huqPzIX6|2$dzC(!Hd>UYO*c&(UB$Kzss6JGpDSUFaI@R=
zYN2<T)6u_q2ukjzQI7_NWs~fg-ImkOH>jn&Q^W_bTlb<`MvjpJ;{|cXj2I7n`Ncgl
znZtA5vGXVYLB<S0QO$J$i^Bg>^qCBsm$!5ubKjL<e6Qh1Q@^zAOq`NMl9b@q@>JRi
zC*I(~8H9ToG#rnb&BJ=eiecCigj;unrIMs+igu^7It}Z16(hO2n?uwP6uNpB`k>(D
zp<0pEAD@WsN9QCnbxLOKeJy+}zSlU+_fxhbO$XFD_U3*<pMIER-&|_b%9v8LF>Huh
zz4Y~5=BRLa!BhkNxj&*OW&*=9)`l%nRhTwLZPLf;RASAtXUbX|riN}@Lz)h`0Nb`h
z-IhH#qZ}v|yHxbAf_t=&lcR^2?Hfw@3>Z*;(p(VNjJ)K!u!S5qgmU=m_j~MrEU^M>
zF27f*3(vL4|7gYvf2X=@6MpA~+Xy$C^sD_iH9uehj&xyke(YW>3fC@M`oNKs(#^{Q
zqZ|?LI*^qcM1931CzS_>rIeDx+k?$biB2{{sZP_+2)@Yq{4?Cn&MBxY`1l=Sl;KR`
z|H`+wTLg1&^Y2>1g_Z7vVTIc|#83R)TQM9k{38s%fga`cFdW5$wV*8IA;l}_)f)7;
z)^U7?UVgZ?InC2e-tP!wS=_39wMjzyjMYI4Y1TRWtLA2y0^<cHr`p#~qWFMB^pNMD
zY%NpvaN_+g^i{5=bHAHCS5NyV>~-zxuU}GqUzjydd=x6mm15j)EY>K$BE<Zzo@1YR
zRsIx1v)FQ#SBQ{rOv?{ZIE5TEfsWY|L9dA!42aUVZTACK3b0G{sm3p!ce>9g6w4k<
z*_V@In@Z{0`$6I0?57hFdP{Rox%YEY{Dik7$uA>NrG=pn^+LGmdCxtXvNAa-dxiqB
zvH8o%JKjr2gjNPyu1)L~RP^gG=w>W`L-RcQGtTcE^75M9=4;5|r<IhwoeSSmUd>Bw
z4i+r-{FM@;PpJ#8kk7d8;y^Hob`!o-`y*p|kIryz0c{=r^^OX}`Qm<mg-P?1w~e$3
z6A}{wZI>WWbykM#IUPTzSzUfJf6Y{$+WU-dWt9g3NNYhvlh^-)M1A{xlh6&o#@&#?
zJJyv|C{K}ImbVxkM<dk3y&<YfKR4dZ-kfWT+T5mVNHz6Wh$jRhnmkk<`#JY_PA5wI
zOe{!y_VIMKKkA){obVI;2Z2Xs9AqLN!EH3O&WbSD`J!^zSYVpBK}qP|&9Jw<tYzBc
z9z7M?%CLTo_4tq5E0C>S%w`GzkP;x=D^hT`=eepYwIw5`-e<@iOJM@K%qp!@(}NA%
zF702+PtOswqLqCr-bnFwEFN7)cCpu6<~U!a;A+=1G_NWUD~23=hj7Jp?_(@ClxN?&
z4fd!?JNmD~=MTDwRA*N@PcBRzXuD11@Y=w%vV|I+U?<-g3pp2myY2TN)A#Rt#>jT-
z6<FsrGucVI#KAV!kbTv3I^-l@d9wYcl@lLt!m+^<J%YQ5F9<_CvtQQsSYz;+1m?R+
z$P!B$3Mzf2H9@Z%kD9+WZbXP1jxbE<f3tyo=(h=<u|m6v!w=18iCLdL8oN`Xxb+MA
z=Ix<y9Z)q#&L8F2qSkRihxz!C)6I&g8unEMce9nzuQb0uJF*7gK9~dzc&pGy1#0}|
z@ir5ll;<ITA+IkYoR`m`8O5~m(1AIEEq~KEPAl;z+W*$vgMmDsJ;KR0Y0Y=YULn~M
zk9Nj*6jowNX6E?$giuckCTDK+i`5TqUgW|dJjY=kZro~4?TEkQjJ8EmZ)BI#r=D-r
zVCmS<as4Or+1E*FF-10~{Td;H*ZqahcBa<6;<WL+(8-hDx3Z_~gXTu?lFtys=)WLj
zcKG%OaN*ZO=-whFiW{M5`Ov-0tv}r$(r@Wem~^U;t6SL1^6NN-gl224;vq(O?lB*p
zB}m4BoIm$<Yh{);I@wsC-t5*gHyto+bhbxmo#HL$Cw<oo43d@-@Q?=2QO}rgs%1m;
zWKY70z*YSKm~}t%|CTmztnBZ^5{(~D?~^C^ECfw!lb-ME*?5u>J`nQJTAxyOyCUsq
zEV4z-bR*n-cPQVRnUNWl8-J_d{bmz0LRKQ>`HOgPVrxvUtH!y<n13L<qpIEzP@gxX
zoubXhey`$1ev08as?*9cDPsybZclQf?n=}r-ksbLq;Qc`8`dg^TN>LxU;J?B)ccvU
z(yWYbd1+Ng<oqdczD-@RTMfCWwJyx|zDLJvM}J$7%>aB|ubr;JdH285noH03v0U9C
z>JF*gu>iSg;p$KYST}sjEMCEdxdBY_C?ehJYV(|~MFd#2FZ1pA_i>|NwE)ns5H8V3
z6v`(fTa`MDM&-(9loM`lB)|MuiQwM(=-RZW|5;PZ^^5Qf!R{a&AkG$BIsvsM$eWV-
z)6?p&vYD>&2D2c`^^R8!BImu<36!_q&nfm=7QnY`1D`g&J<AO2HZS$P7;M#*N}}$9
zaB%$2L(qmawYW`w(|hEk63P-ZoYG>PCj|HjZimrhZ-QgDO|Fi4N)_b~pz%<Xt_BEm
zGw*S)<MJ*-lIiW*txW(uuEM%4g1cL(u_n~<$d&#9F8sn)ShWt57HJ-piqpvCEBSQt
zn};x|U~kz*^IRWHbFgWz$|({YLPFIZIll|mNh}4jb@}#9qg{;QhInk?uZit9+#v{*
z2(6)yq?Jd1NHJ~Pb;v^T;u!=ooZOtwqviu)Bx4OUmit%AuOkJe6ffrP{+pqwKi7Fd
zlFnofGK@9}jC+!KEiz5yj;h{9&hJOpdLBijd7oCA4JCc2D(5r<Wp%|$SViXcJ(!_B
z^A^b*UYfQjKbj;ZJo>{J`4|DtQN|WhxS<*!Z!y&&40&E*?xxH1%p9ZFkG_8>bOZ94
z345~12#<rprwmJY94mDC5ht{HCB=^UAmT&M$D?m;v6QUVASluT<$=ywnjuUXRcqwD
zBszFYj0M0wzUZ&kYul!?-2EWs`O8$`%!i7^jhJ5}X8fR<2*yeyAERI|2NuiA#am4K
zGHQ8ygvznE$x&52Wt%2^mLgk=z_*Wl?}$4xIxPc6Ja`*C0VQXb*QwD_Q{4bODqScJ
zTGylG-B0Hc%&fM!93kPFba(OuP>B7S(q>JPBL#EiXhiuJ&bBWL5Wp!kXAaRF5Yu~H
zD(Q9sPU8`8+`LT&Z0GCKH`jWXRzK^-qdR@%%kM!Gr`J!mQ0PJN3Z7#CSrz{>XZzNv
zi9x=~d~g;(0;Rj4m|PH&s`f6WL5U~6^bVHZ!2e!f2~Aa!1NEyu&zSNxkCB;okj6S4
z?c%bm3Zdfqj+syY+k}ZTkbiqW;`BI!AS)Pg!f)tQn@yL(bqmV%c_GrLZYQIWtr_X%
z7^96nkj}itU%q@^sh^ac3>|+uf!ntXBonYyX763UZ^%n5fa5W*v3CNezYj5uezFl}
zvR=bcYEn}o&+(1C`Br~1eS@$2pK8>_EY632B?%TJ8Uschj<5WRd)THj+d1}QL^bse
zAkpkLPPU1&uR#r1C@Gf_@9Qidt!p^6yz#?s$C@9iH;tFJkVfK!O3obb_OwZwAyo;`
z^N)KvB@5rvACM^fefl5%aN?o$ZnOy{tE{neQt4<P$mDNG^#|{|hIyCI#Zd5ab{YfG
zu^Q`#PsI#E9!vOD#(O2mN%qauLJ5uiZF>Zzhyb9DJ~s^gYT&&A^n8VjW;X+S^rD-c
z9J*gdmTY;i@7?bkP<yA(0Oi=RGHW5llVSS1G3#(_0bV#KW#kai<5U7s>qL?By1u=B
zq5x;2=tjQGvy3$?F`RZ4yzhRNE*Z`$I2*Nswn5919SW-tY8~$|&csH!+EYJ-;dFg9
zhh$73`mZRkw9#T@gYIWzSwvOH_&%A;XJNPyX&-@wKQ*KUn8z)3VNNelAAPAs^d%aa
z#(aa}B9x#8K}1gr;e5Hldkgvcd@A>^pDfMINj|_EOI*83X=xmyc@Uo_{{Y>xT;~Nk
zkQ)8_cu3d|+nZFLI~iBykD(xgCfaUu5B=EE_0F`v_v6J${@3dVtji%x0J!$V${`;^
zA-Q<w+yzF}6+HIHG)!f>DqtN3{ImF08Md=7oN6c0vPIU?;v!0?{rs7KqGzx~dP}&1
z&)ME-vFPx#;`5dMm?t4ZNYe#$v!D&#yD){f2b+q^Pn^^#Q&@80DS_<US5oi;3egyf
z;}3Y1!}uba<&rw;@9mH0?ipn^h?5$HN*-yv_-`H{(PeO3l#OyTHM%1iNN0V(ZU<VO
zJI{3FHW3Bn;~R9_mOev|P@ZU4xk0I;ZBF|tQQ7a&3_&TxOcoB#g_cY*6x_u~la5f`
z(|1Y{FPo$&?kU=1lAZ*94YSVqd5^u!@7`F9RS)J@yJJ)7n7q3JeeVz;2h^N`m!rE*
zex6@|<K%Wh#SwL7)jt!<^#2sdw}N7aa_P5bJVZbMD7vn~xdTF~B(pe<^C~d%9IE|O
zr8#4u&g_q=N*fQKFWg50l8OpN)cWW#ltIfm%xI@C%aTWTH|`)7tco>-lNyWURJs4c
z=c4pPb3tkp@ag>yr=%Xh?asNyG1qpo(4BmYuPz#7IuK-&sH_aIX*DmOd@enN%7cyM
zDDLzzsj>`ot5p!MZ0o-r;>RWIT6!RyS6af);b^miXyyDfZ{HHYD+<lxu<OK!=XsJn
z5({#Q?_Mat=A0Zkb-PNz!+k(vFp#9Q6HOm<+xF<TN9$}C|8}nIH|6O51qqxQ+a;lD
zwH9}6r9f&?Er&|$UBqAOo##`nuDX5CzK&w^9Tb~KCA0$hmVfN=6K<;h<i4&joVV;t
zBtHk2C?^WNJ1sVyLNW5Ca_iUFofS!)Anm_iS4Ej404zE|wjIC_v?=mt68wcYY4o5d
zZb{RHJy)jj8`ivkWW5=)I|i-RuG1f%2|A6J6@{2xRjf>Y#{Hve^I8I@fvywI_CVG3
zS<n6t4Mn`H$oaxYyv6aI8^HDfaMW7I4h9uKMyJy`B0C!OCB9*yIiU9XK-<D*@p%x1
z-?;zGmYq<ARay;|-m-j-*j!$Q)}{ktZ9}JSLsp7))4$VXp^t%{ZV6ZeO3|jiS7;2o
zZ~788e=QvLCqM<I-@GHWKVd!oE$Vx#d&j}PolzC`#h0r&hYf~LT^$xGDW2`U#?3b!
z(y$HD(*dHr+hej<&~!TcP>-u1h4w*6y<q0%o0?mdBgEb;tf}9FHQ;ys0;KnJ{|g`@
z;`1N0I~%<YjD~j06kn2uTH-9)!0<Bm^IID#lYM=pX{&pZkMuo|dy%-vdnW1Euk^qR
z%>4nJ_)9=_-iA#z^frU`o-w<lTPnSfWtPJIYZ_#Rx8q!ofTfSc2%7XsTu3+o>0xC|
z_fo*WIrf2m3HJ572)}K6b<A-Q{Jc_ut<q`9v0W1=)z|zpSTnP{$f>0ccLK1sy;UXh
zu@xw&Ql68=bRrPPTOpvWF71OJ(}NY8hfS%3yhWPUztH<ysXMylsm@>ag&t&G*?)5#
zehcuxw|PAnX8jYV1wqBsqs7Yi7W!>zq(t?rMm!^1b*@n+eET{73@ic7Yf7)+ji+<f
zPoMNTyR=`yb5rfluQDAFcwF@<rbP~xezNf5#k#)`LFhJOd^8vECKfe|XUTLkdriLt
zVxK49_rP^X2b3@}@iZ5v)bCs4o@nS14t~2n43Iex8H)o-HERMdSpzuH@;qb8M$$8J
zi&UjDE^}gXn&tW!Aj!<krD0n{2K=fvSv@TokF7k-rE4L3ZAPJH^dc04JL28+vmy*0
zHF!B2xqWXTiS%@Q2?Ko#LiY+i1_1G|9>J6I_E!yvIGZ6*C|mb!KyHB`#s8o)3&IV9
zsVmhFXz}ne2|&TkinNpgk!?|}C`Z#{p_M0A9=R39mao1Jx(0+mg@z|0KIqv%cbnks
z9ndGFXlbXGK&y4ujSGJX;`D9a4|=o(@hU-xAc$%C#=pcAt#o=Ala;abiD-DtpZu8C
zu}-ZG_FF^533uaWgajVaRdQ|manb#&x@gbk-TN}-(}A`vEP6_~y;$MwE@Vgp0y|M~
zVrjTR!HVz-I)}^n4vXx2$7diWppj*L9<IMworu<8VJArNYV%6IdLLmxQ@4-Yfb&=`
zZQzaQa)*pR6#nB89jzuCmg+7;J4FZ}u4B`8q2hu^R3Rm2XB;&0DjhQ}_;Q9={#HpL
z(VlXsDWQ=Q#yYuJTi}$*PxYM{ZeWc7vrCZ7g0@(IEs?LR=wunMIU#HKwY@SbnXng;
z%QleM)(2ZC@p3!Nt`J@|Rs`4(`AUk!+=-llPF~fe5vwOsWBjiBGHxMp3<!Xl>INe2
zNDGBfm!Z7a8AFRdu~V=R>zE1cU4R;Oom*la1H?r^FAX9}CuqY1NgRkZ)Y;2Yd`>aH
z;hm+uiiNoG4;-Kgw_Fc@6LF!d#aH^yRRdoDblZ$@=R#{^WllACYaglkDlHrX-Z_!_
zVw#C%To&4^vREtBri<3*+R%HocJ+hD7M96Mt#GvI_A<HEa*nuC6k#b4SHTlY5Mm-V
zf!exh<Cq-R(csn{MDdBKMxcwAarTU_lz(Gc$5uT!;wIr1ndxXs1`k!p<0G<!^MRhZ
z@iC#@WjmuDmT2gfuY9FHk^@0L8{9kXxdBl^DU8RhrO-mGr_)oBqj*oN%^8rWcYr_W
zSv#L6J^r4@%*;%Q?Xm|Ne+nwxtO7T}2m(yb!7dL(Lo>GuurhG-Ft|J)co8(dg(wB3
zrM|~8;&xDYs0D51D=AKyKsky#b|F#doslnA7nE7@IVoSTo>IXgM@tU}ppOZve}iTZ
zvFJG=ZZuyAeEqMYT*sO2h?xe9FL&Q7TX0m^p@$o7XL;{WAz9sx-iI+b1NAT$AQc7n
zmx-LVo}hG3KwH(XOwI=uuPX7JnREokGK*FS);~z&F$lf3y}k|&;Bg+Ruh;x@TKx77
zNLpw}{sUFx;Z<HCnao6>J5Tzyeo->%`X!b@xCWi@KIz8<EcCppN|w+RW$i)Ytig8S
zbFth@@J_@$)hbBjBA%0biPBz19rn!-#?^kmvL%q0kh}tHgWZKxzxkJf1?Vck4{lN;
zO~xsAMxfG)1g*ESULGVJ<+3j>6S$pL{XUa4?8R1%A2<c<{d*CX>&$YqyIS3unJ(<s
zj}jNG=aMb~<$catwn#M(%5xY*<-iY+fs6EmL}}#mQ8w$Pt}HqPK>n_!B?l4#RpVnX
zZr{pKR{=4QuHrR60%`1w4!#H0xLx-RzJQsBXBJlw{0>Nurm2w$Ld#VGsZb4`$=3nO
zd~ENlJ8_!1+BI2xGjr;x$T~!E11xXK?C<jq2330z#!2o6YWIoNUL>1BO29C67+Vpw
z9I(}x+hm>!U@5$c(TJ)9gOhV#7Y~#ae1lg=ZgX(JLn$>A*dT(ug(?64{=<4?1><w*
z?ogs_ST}XyTZb99<(K`PLqtcukcVe+<i@@OvsTsXQ9E7Uy;x3fE68UwAJT7g-;@2Z
z>a()X3sZd+*PB;+3b{*k+jplsGBkb%sGLjZ#U=W*H>oBV_uQ-6;k#JA4fjqKyow+u
ziwv#Ix^n-KM+wHzFWOVI-(Y%!l~2mh#PP9Ljg@)STKhn0CjEn^Czdiqn@D_#e&SV~
zxHVNE*u+qUvG~LIp18_u{e0Wv7=4eH=bbH6*{$&pi7Hvjs|tE&5qMU5t8QMpOnxSa
zSB_fSut4n%O{*HzoRG5G(R9*1lZ)pXJd&Jj$C0O!-9#o&-nA?EV&?i)j$mJUjh1bw
zkx?x1&4b}29>ZSt$stdbX7|z85mzF8ba8(t8S~^8yYL(F+n37Is(;^{Rd@v*NO*Wn
zJ&nU!_f}dwmZm6KYw9HI_9;tJv2316U$rugWi(s@@dw}4Mana34-rqz&!q+qSS6h~
z+<<MSOrGpFi$Ci&o_jlWPf<LvddSzxUt%?8J3v2zd^?Q<hITE`R%t}Ltr0mR)m<7m
z`MTe)(`*)oBd%)^SP91clw;ly@rc@FSA+1z{R%f*b^EQvEkD<d$7`MTmgoJwJ!9pe
zLo~Nk=Ryc3h(LG<)eg48T4{TA*(J<;{t{#2R@ch7nMI&fvJmTX^Loi;pG(->J2MVH
z84nMlUZ1hAv|)-Gn9oUbs@|iWA5f0nG2iLb>>gWK<~s9Xy(RxKJUoQPN=k-tD44rB
z!FnpK#lbyI_JeaDdjw3;b1+PgN^zEB$6kzYOAT7QPDnQ!o!@WV^X_v4Hz+U!A%rq&
z@giQ6;)WVwZ)bLdn3i`wcB;F|xXthv$D@v*@uIH;)CB}ea7&IuoFqnvw?Dd=nQj#m
zC3FL@8o`rZ4%Xr`t%?u)3|oyU1ShR#J|zVaR^am^bb<X}!Z;`$E9XY;Kl9P<Y0drT
z4IzVCTcA=Tj^#T9EoV4@#tmssVgDX~xdeGRhgyR)Fk8=fm{_0VRkn+b*GY3?S3eO`
zxJaQh%9z+Y5abq171Olvw=dYS@~)y5FCDXx9`QRe(#Sff?Pk+~yh<(B-&t!{b8GRa
zzfuB8y|PF23mn7JI|`Z0wB@%?{=4o}Jn(e<l~Cm3t#KoT>3_jOyD+ivJ|-MM{MBce
zeQ()O$ylau3N>c+%n<o(AfL$7SiK_vp9ucV`FiV$75-8SmAiJ$?=Ov^dD+>Z!45p}
zS5dWiC&GwtYgoFI>FXmFQ}0&@GL$}=$vgH3l!{LW&ACpIC4`D~+uL#v>K9w;dm8du
z?!=_(u~AYy>|rgQf1hfTcjJe*SO(?k#Bo>Z5^mulqdN$gD}qFsZ<pfaD{6OZiAdxP
zAEycz*!6Ih42&4vLrq5>eC@nvCSQDM;jidNJotpx=e>Rt#Yz4D=W>m9Ln3nrJ~E@L
zUGh*VKeQ?kMDf#WD=t4~(>}Qb1qFv{gG_KqKxU}~*6}Pn^nusmk_W>$bfe1h9J>Rw
zg=AVMl&q`H-Zc8&kUom0M8%PJ3>oq-KQ6u8-safPqXg0abK_Fs<RU<btV4BfCIqrZ
zV8}P!%tCf9xn1<#<2*I<X%7z>(x~^;)riWSs2A#vOj%ciT{0npZ?!pPrY0IavH!Dy
zY*KC*hZS~exOkRdju$oy-=?T&_6lPPKCnZOeJ}Bt*GG?EO!j6WjJ1+__WbX9(v`2Z
zctGHL)aqdLda4kskv2gIi8cGp4)Er~ebyCb9_8)+3y)DFx4tOC)dyM3x=EZwv(#De
zQ0W$0a?gs-^YR=!jYOZGb3Hk(%{l+)x)&wYskKeosq<yBCFP5!qpz#(|F$1xxk;LH
z+NEPkMGVs^dqBS!HJNbnAPWCG{w(1GOL#3#DW_`oHltwzC*>^~&-gEl%C=i_F_<D{
zdY&CAiHE|QsDi%4Vs4MBhLoLbe~i6bY{pA)OOMOB8$*l^6!V#6vnsk~^^da4aO2<T
z5hUh?Gxn4f*<DI>>=AclMY^O8jOhQ3#!=kai(c$Fx>qu3rtK@^J|RDzadv_r^uHvL
z)B_WS>{T&R_OlDQ5F)ExvooF8Vi8T3ZqHc?{$Hxyeznem%Wrh%6@@kW&OvwD#Ff~r
zU0^-(tZ$aa&s@S~`d;%as5MeZhr6rZm)IV|%3{Sn9Y6!lG-WD`Be1(?xhzbBq(ci_
z;j;2JhS%A3OQsKm%K14A)Y-?3x^&C|{K)eHguvp40Rm;-S?h^9BkWX?Y%Q+ECybW4
zwBE!ixNEoMzSD4`oVsTSPKO}(jLn)d@bsF78ba%tsY)3q8u40eTfLE_t_5HI%@J?@
z>Txrg3~KoBksYR5(ZAU_Uj}~mPfAx2F0&A9BL@3sE#+24E?%z{@PYSScPCApB@4e1
z(XOr~+()zOc5<jWm*3#%7fEHSG{W%o*0R{Lg|-xWp2O4LD0kRF4pRUHnEHXI7TT2s
zj({7pzu&?`0(gz8KarQU{q&cEu@PpM>&&?rEz|zQ?&7i!tOqa4qXW}3;O+kX7v*-k
z56#Fa@r${+-;`(76}D==5Pru0xdq+x<*kPpE!_`J&U&AfUba*$o)zqYj&$apvSkC}
z!m#!J6lGphBbB)rc(9c}y(7Ps&#D-WkmHBgWnlA%hv>@MvBXGje(9o#v10nCeJhfq
zGX3?C$$J|J4>D$NkXN6L|G>PR_j27`UJ0#4l{{aod$%_MRYdR)4mTzZhng>xmOE3M
zf0sl{a7kqd7ag|>qviYk{$6obYx;z#V(@rh0U88;x1O?ZI&HU-G%VKGYfPedNyKan
zHa5&k^)o93oLA;GH)SlUP8D0^+x+O)9GJhj9nHm^7k@}`V&}<y_KrQP)7*mT=c{eE
zx$kkYO;WFYJ%7^q^K5Y~@p-X_zN@2O>sz*A51k0(5IN~L_`@(nuDs8?qQCR_l7e>m
z>r9)LM8b4I+{(8pr4RvJe5Ji++WRlG-|!6%v|14S>7NzW74bu%fE!`^pDFWQy$Z&=
zg+-izk72rSSL|v}?erQ^f41GL$4MVeqQh|pR$QeJ?LC^uCfc3UiLX2l&FFO{qU`Gh
zr{hFZmg>lC|FK+8fWHeRBlAz$UQ;_#wuKEIR&JgU@9<60?gb71`NEf@y8^<Dw^(eu
zp*T{oxxiVmM>Bmw<&L-F7uw|ZOK54>@D>6da?%43|0rF0^we#;?NVtf&8^Qn7do&!
z5`jHeMwM?q>YY}#x!FoS`>YMOghI~jp-8q}4qJqAVDfxP8{m<V&yp)C@k>|+_5DJ%
z)l%6f_}a`dVe?%56be;BIYskvXRXStyCS)p7lu99elP;s!M9(*ZZY&nh@xP*m?S*J
zx3PGfUf4ezn86s%f0+y@`6g50>{3*A0$6&yPfOFH#IB)Np+859e{Bk&J-ylU-+wHf
z;<q$XSdC9C3+puu>2CM`mOayg^5LjmB~pSpM-A*lj(b7^OOEK~--?g*eO^@1^<=2$
zv2--J3x%4|Zaq!ZB9GnyQz4!5@>+PV4Z(YUQgd&f<9-w6g=kW)fyO3?yroag)03xe
zUkUj=p|RgnsN~RsbI8w-VMY%G`NOJ-ADEr1mu?S7(5>TRoh+6N1g+-k*n#@>;^bP1
zaIfyK4+#4-Yh#>b@x9`aPyt2+(<*j)KClX<_*AV`7xFHl)7t|L_*NJn^8g<=vSt+q
zi~_3&KDGHXwx>@}Ff$$ug=Ss02g886K=qedEUw$XD2O2|Xh<{UEyDV0(Zv}L&B&l1
zhEO&vEdwtCfXl;xtHgGeWv)E2-WA|jJIF(*GL~b!NG)H6Pr|*S-vme^wSO((Bg_gs
zN0kZijm&;kk;7Le;A^XYjqg5*RidEOMeW~1eKKSfUlNV+Q28X^)%W3MDhdS1yaXY2
zH+WWvAVKj^cB@40$=>qjQXMFG#1l<|j}d6g^HOa0X(UT=s=6**UC1Ai;++6c947jI
zmFI|peiOXxc=>fLeyN;pM6$CNA}yYp{x{L(D&LL@arV;3cd}&2P#?x*Zn=!?FF-@Y
zZcpgX(hc-8>_mF~0m=x#sF1Nr8!dlio2L!+CRVou5|@X(cWp#J04)@n$C(b3ve&MI
z(1_r=UIbrXLpKXS&gucHg4NqV2;CDd{aE!;GotO<9{3{s&*jlID^9%pZm41Ww!wCv
zTn^|Esl8=?s|xm_oE6`Hq}#e;Lb$yLj$G4cU$gju6y<Jr6*E}FJ3BVBW&z#(Yzwrf
zSu3Ar8bRRZBDH>om6XEqNd79@Zzb&fKIKi=qF6j+ScT=OVT2lsXB1<H(kEhqLY{h}
z<Ri{&4hYe`+SH)?&wA#@zZvosI}9w247};LU>`Ka&w{JemLk-uA>b+O>f3NcIpT5b
zBeaMBSg>_mmKWRYK}V&-I5ZXfGWO<Yqebd*?AeLq%fGi)pb9ZoV9n=O@|ZH@%Y<D_
z0RvvMsmBdTreL|#h=`KUF;$thgMYX)z>f64nld1AntOB5W0RcdowRBiJZ%REHiJho
zy+E)p6XK^Me#Ar<rd%%|6^?(Qwft|ugcPSGl>T_Ee>=sF=u6)L$=D8=P9y(sgan8d
zDAe&-`>JkI?!x593DxUnim7Fp<-20{5uo#n;BP25gX}1B_XW$LqDOXq(jawDUb4?<
zKW*~%Y1Ge$8t0(o)6-FP^&`i|2r|LW4IYtc?qJ*Q{%%XIG_2Vd9nRfSP|;}X?f%>a
zvDxY*vD*ScEjzov=^+~K7s(Lrsid}atxCG;O`~st)lTeSwUc$*1D_gEPx4#J*mu03
z?OoPp&mOFAfzDnA1YoMsa}_1+es39KsEOYGxFF4GL~t;nR{Iw7S)idzll|XqJJ^@Z
ziF89_|M@<etP-0#TudCZt|-Xxy~xbBE=!yqm-m=>A{4Vc=YTmi=2(KG6y&oek{4U}
zaf=Potd-=&ws@;^EO9M4_Hw^w*Z3`j1lyjYlh2J#+|j2bE4v9FtCxyg(N+4=bFYt8
zgZ@FnE~>S(&o-Rd&yK%aA8PScy|lZZNsUh0ku~-AfHsLR`T0d~b&KB|XjR6p{W)Vc
zY4i&^h4MWhvER(YHz)t+ZeLyBQlG&eiITT$i-q3^kAX4tOP0Qio0+plDsY3}JepNz
zqKR%dO2!;JXQZ#(tg=dd$xaQK{q2IJJZ0w7sHmJ17GaO6DuSWkt7_P)D^7NC(m&xw
z#e<Y>p2N9O(PPjg_~6>;P>><<qvVOcHrxGM#|tJJzZ+H>1$x!~R8nqk1%jqQT1Iug
zs=1CjI_7&mkQ{**mcPBLr^XeeQw6bA)>fpyM5L0SRfpg|oV-V?W<@w_j;L$FIZlZW
zrtO*l4h=DvKk0W|)k=rG(sws-;TNJq#$SbnAU>4#+oqe<eb$cVKDLEhNaXn1T>asB
zcQw&q)$NYacWEm|lN3a;lR5ROU%4zig$7{Cp1$9+k6ura_t4V$Zjx#YS~Q4*r}vbX
zqUwzZHt^i=y0cV%nE+pc4Ihlkvo@@BjFtDa_xl#bmcpAIL1(z{%Pw@KMMPn9@=+-b
z^N~;z{#I!pJ3rgvD>VViKHM4IEaLN))GF&|-%X7i&`;d3MmT^orUE&IGPV#=ERjS?
zlSCN`pZ^Gj(F<S2HLqibcbtSH_H#4-5r~$EW~a6vt<Sri#}@Ktk2K4$ufHA5Z*U(F
zsd=hkKY;4~WY&Y=uGkR%h)`5jtEr!2HQDcPg4a*CqTQDf>ss_5P>Ge3MA<%`;R_TF
zh`k{%;XqP%+SD`4Rkq+yvpbF+VbH3QDDMHl4M67BU)QNB&B|x9<hjW<cU||%HPSX@
z;}8oUjIeWhId|Sy8m)F90hNYqyiM-S+rmw3o6Xs<?iZBsnV1_A<i)#=Nu#ILVG7VD
zNk*uARKxC7@=2rQ=0yi~_X1Y2Z*QGZ_FHr^L-oHXCIdhyiG#;ir{mQdhtMz}N#`Nc
z${1a{QA<<@0NdZ<jkXn{?&kQV+cWRL%J5nif^^<NSKeaM{pq^DPL4@AItJ_vtjo%;
zKNq^_FeGI)qC}bBa=3}KPIB+vp_hmBhwnZOiELwIPvkMctryJQ3$Ub4K1=`<6yjx3
z#7kM##grXV=}XY`a~~(f0m+K8f%QQd0o)b>y3Rgr=0bOP`_lo<5P$5~eeWubhGip%
zTr4!x#yYZ&?REDX&1IvH=@?2D_GV?l%}RAP%ccvzH9-*n6|3lg`b(6GW-hd}e${3Q
zzrb1SR3<ECH>ez4I2zd7%hH<Y-;q>@#@0v6jpVIbk-{VqR`Co3bs^9cL#lmbTf~rO
z-py$WAVW-<)3-pKNECX`mpNg4MHiZ$8)BZ~EWB=Jc9|a^m5=o`D5ORdr$Vk<S<$9w
z-+pqVA(6a<Bc0s-dhE4*3=F>371NFWYe(>JcPQ}6*HVgKTqP!4?|{JX*AN#$w{q*>
zEED+D%m9c$f&#`9N?2~&F#;(6%iub+Ii-4U?hK@)+U}iGGGtH_z1}n1ov>2fUptKW
zOk}L<Lt$kRbY+vv(BeTk*KB>mfXE%eyP+hsBD#<EEAa28#`IdF0SMCon*uXiDEgcw
zZkvcekNlgP<=DOI=QlbCj|~z0^TP|)6*-|(qo<r*GLuUpxWE%-Nti2glZ&<G<;PWB
z=1{CrJGXN|hCE{BlMYDM^TW{&UC=P3GzqYf$=&;cKiUtC+LzlNm25#9r7Zie`v3tI
z(&cI?XnXM{VS|K%ijc_34XBC{J+RqxkMrm3Mw?!Sd87ND{2Md7guq?@ux+FX1O^pT
z5K?|KPU^<cZ64v^L1~-c=+4LJo+l4~v#vmO=#7dSA`G)av^A;E+Az&9Uynvs_5sRT
z^Y2gPk!?Q~GNHnJ3fS_z1ncJ4zw<SFmg*H~?&w&V9rrS6zLvwtn0Vq2SP$y`vWVqb
zws?{`6cWba1YHxD3H>nLF3Ut1MMvkax>LOZv5Ff|xE9r)bKf)bDxu2f`Bft&KW&)n
zR0i6H;5rrs*bB>{y=DBRpG0;o9nVpve||Y(N*M2d>Iq4HOx0yG%DdGP2Zv<HV9bnB
zLY&W_igrNVTp{qkqg=<lpbZJ;Im|YXb{D*JB4evH@^7u*!j=2G{ilDwE{S;&5_toK
zxR-JcrCSu~qX}qB!a=^42q<VMckAa{CA0p4#3ySr2uXADIT$G|EGE|3c$>G*yz~E&
zrU<QGzwVtVLg3TvYOO0QS0zVNG#w{TU5nSkX`zjFHLg6m40%(|G`x@w>yA@Gqose?
zYO~u_)p2bWT$4kjujx>Uc5w<|{rd_AauV5npj(%LdI_uC)Njpd)y}D#%Le^6deYCD
zFlO&R?u0&iL#pG9%(+Vcpe0I-LnDW*0T3om-kV+JbgHt|eu|=Ab>fM#EwrXF`_#vz
z-?D(3fBty^Hsy>v!mczO_s)TY1e}Khd_x<Y$)pEnM3r44vh2REXk`7U4Ecy#{ypLB
zQoACPE}~n_3*cgpbAoll+0EmZ+LDAVUVxE33?&Ns?;fS>ukEHCy~4Jg5sx&RBZAJc
z7lZ;N*SlDRb&owR@&As{0{#ug{Epv%c9y505~WC+k-RIuf~<L>8-5Gz3%xXvEZX7)
zJyTHd?EvrED|MaO;Oy>2rEvB-ZFbOo2OSf#t+rS;&8gF=EwnWlcesTDebXPBmtM22
zkv$BY7kX35i)M}X;~9yYY||~F5B6}L9e3pipQFw^X1^*!rrS?XpStVb{4mET6GBr(
z?9{@tgoln5<7=4HO5fASvc*qM4lJV0;-v2CRIJtNlKvEHFD-<|x|@r3W?_<H88YzT
zRyMrl%cU#Y_XuJZZnsbQCt^b9y^bp^w~e^HE*aB_Bw<(~lQ?sDjp)y9V0=TRTPDp9
zvfHM5_*G0>ZGQ#r!+1%KSa&6^Eaq^#Mx3x!`WcZk+B5mcjCSQIP9__zA!<LO0dfGm
z+(o{mq4}_*+O!~pnwlCY)eFxrX8S_V>o|=}<}Od|g2)i1>iF#Wf2z6mcPP^^Z0pKa
z7}YMbRIE*#mR6R+jMIi%$)vOxITmSbF^q8>Q*C8C2wjvzPAv_EA<8fg)0%QVuH#g~
zg`5p0LrmG{{apPKJHO6+-#g#?obLO6-tU=h5mv&h=?o6fX9bck%(>{m(jayxt76V6
zIbJ5Ej;z@Yl@H8=VpvrZ9iW!_euxB8B=j4AmAgTJv-Wv=b3vNt>mbdPk*AJ@@Khrh
z4f<w8d2asVhds&fQ&QV~YTv5D9D^2o=C0ienCaX=m`_iq9!c05e6ZP&m<jlAf2R<|
zXiMBP%=i&DU^i$;)hwFET*{%elN<5TY)Fy%>@Y3D>62Im5FY>q1{}nk7rIKDRO}M=
z;F&~%BTSovtsxsQm1cV<4+g80a4DT+nQUeU(&f#_H-N!}k7@#y>WnVqQF_SMET*p6
zm#@-7MDhfX9l+yR6z}ZR<z{JAOT5&$vP(3#&?lnUHxF2;sX-I^y68#bZ`U+lhEhQE
zp^*m2@KQE9zbTdBY_!<19>GN*N2DRaV_4xZTA>V)saTqc)Z^=RJOQfE8O}{kC7Cd^
zEv2bAjtw9Pu5#cht$oGtZ%77;PCMF4at74;kcqT#2G#RAa-1S6K-(>Js5JFAjT?cn
z+!@U1hL)d+0gTc@54!X@#&wg+C^Qh<ps-c%y(sv2hT-FR8@sp=^ik-5wTv4RtFT;i
z#^#H;m6EX4wT>lMJ`xpn8FqQMD^GHm^o6!p{j2SobD?)FQ3+|N#%Je^qe(Ec#f9%W
zR0@=X!gKmj0L$NwLLkUikt<7f)(xYPR>+{$Kz05X+!~d_<F}^J4`L&rTLY{E8~HI`
z(1gcjfqtdt;Jb13oAhs!FsEAqHo@PQK~n@q3g%rwnu^X{I#$E{Mi*UiPZR7edzB-(
ztLj38!4_5VY2P&$)2M7vvko8>8uf%U<_>C><6Kbktj#tTtT`C+q*HFB&G!eR`{x_v
z(i^dGoLX6~_5=0quHDs<5-1QL03=~$J7QgF*zjRa1r{yhdo-yd_)qq-j)nd!qB(zC
z_S%{PY3kz5H^}!}6BSx{ADkG9@fa``IHY~$49qT@DO2OZ6Ik*r)-EX9KowGAf_FY;
z0cfzE2_u-SL;ube=L_T`!DD3G3xM%X2Rz9Og=?EH`U_JZ>Z<7SQ!{`^NE+t?x^@Z2
zd;R{2)zq&Doux{K2OY4R_pgPzB*ng|H;g1`_v*a`c){K`>>%^??1QzR0IIttg}a%E
z2B1L)476WKbr!K&_fEQQ3KkRMO1S050nnXC0?$f!Lkflb!Y?kT_t_x{JAHo&mO+bb
zb=-di7d2c+6kBRbT#e>%sv>J^l!7&65I{U|fle}G!3(gDunZ7?9XN*&v{vEx@!@z+
zWqjX+K!%sH_e-pmLi@^4Cy5Q(dwxJ&(U^U!>(a<ER0cKhSF4Ju(EON_EU!DlP?e`Y
z>q$x{^)Y!ae&86kBp$YsD|+#Y!oUYMk2OQF6`yx3nXWgUi@TzW`eYf-Xsae>Hn@O%
z+3*nQZL6R&?<6qPnLoC{kqq$f-|FgW@K>^p(>jtiIfGeqBMW55!v#J1aFiw#iHjLg
zI|z&{Hj@Gj*zkE$4&oO*Tl+VU6G3dkLWP67h`czXA`i54V3>e3?!~2AIf5GWH_^p~
z5HSrl+Ht8ib`Q2<BOs*ZDDnw_9Tt6Znpu93AA{0Mr>iM3QKf@P+Wlg_-b}tZFY0)b
z1r}jJKq$&VS9#)*OZ(}W`lN|}xDEd`FcmkwbCxZBk9=1b_i@IL!f{}jOEUH`!8P~0
z!i>fnMA|lXV(#`uchEA_BDPnR=!Ee_raX!8@jLC9Pu_tQ>R2rNqL8fn(i<MmlBpdC
zz}M%nA6vOdOEL-ihevP->pb$I_yS}}BX^JU?Vusl*LPMF0Qo~|#&afp>>b-2-|AFy
z7NEkAaQ!VTc*v-?&S!ITA)SNk)?0N~<Uy%GQW5=Sa$BWG$9vyC@o*Z{wO6Y~7~GQj
z91b-_Q(7xWxk%lU3fmX(Lx}?7g8&G)-32(xPOiS5xMsRll`y4pA|8QPDdlO?SurSf
z6YXsw6yJp#L|7q!wfk*EbqT1Y<<X9II#iDjCZnJ{vBTXt975a{WQ8R;AW8GT_cIK_
zY6Bc?_)nk>1f)JK@E*VCnrR!w?Ce}5U8Rs(ZTgl2N!0hE<vQe-1M0w&#n9<1f{Xy}
zw><xcunb-k8nR+AP+My$fA;G`bCIT)a^)6@+3$aM)<$UwhH$N}w@zESFOi5GrQl;_
LNwQ!ca{u!`++y(2

literal 0
HcmV?d00001

diff --git a/tests/assets/car_tree_bug_zero_shot/images/val/Slide9.PNG b/tests/assets/car_tree_bug_zero_shot/images/val/Slide9.PNG
new file mode 100644
index 0000000000000000000000000000000000000000..da9448fc35f18eb6c4c799bc2306afa2bc8e3f2a
GIT binary patch
literal 26796
zcmdSBc|4Wt7e4&ZAPJSCBGRM?l{r%-Nr@yOQ!<lz9=9e=Q7UEJ2u0@1v#p#+=6SZs
zuua>RnLWMh(fOY9e%{af|L=YOIG=M)J<t8DweGd9b*<~Zd45Y>g>~nlod|-ks$RQt
z8$q_ge=@f0V1$2N<h#xQ|6*{tt#TR3YT}uK|JiDJ>E<N_$q8kln4pg`J6_XsK@hfT
z^dE*=htFmRVs%ON%B4G=Mno#}JMHDzNuqJ>(rbSQyt)UD>W=p(yQhd*w;jLLapfGh
zn1Cebwtd>$I2U*E+k|X+ynFYhFRxV(-np}5=g#d6`*(}m{FS(UkG0;1ONpPiKDnYv
zTVExoV`G0hCH=UiZQ0YDR~bC6;FdOxtw`@m3nk3rX?%26)T|=`VEVPK47<RMqdV@y
zAML$_Y(bFMI~WfjNWhidRq#j0A0yil<j9__XW@UI{~vkf5ME3cU+UIaKZ3X723uHm
z+&^Ue8NXObw_20x1M7Y;GDg;T;G2zD8RV#Qv8|He5a;eH@yddj)n`bh{qNxa{;07Z
z!@=mel#%)0A6dFro`6sUJUkX)?VC+h<Y6qfp@g-%SFU|`@%h$ZwBQ-3pflJ>{Plf`
z!@gu>MWs53cQ99WV=LGa1GeC4Pe{*@tOI#6eNE<NM!kH|EBa3ytXs|R+3YN97MzsH
z{<2Zn*7phCPQ_J_GaTa|XSpc}WaySo^j47lx&8QIq4uq3Zyw$17-@Bk>63@UrJ|_s
z6?0ZnR@y-c>>uO0b5cg<-Rw4m9SjTRg~r`R+Hn$ABV*gxg-pPEX$G~hwTTm<&-VF}
zV~@4E&w0Am&xsu4l5lk(e@$;6UAZD)=Hf(>)nAR|t{g7zWv_4=zO`frTWlc^V7*s+
zG=A6mY9Fn-)m@}#`6rD*N0j|jUn$|S*6dJH3Nc`RrcZ;&<UNmSVkocNyBlSY(hGq|
zvg)qL4cE_=vDsraEO80f1VlZH75vJYt-)sRV#&Gusj>_=@Z!A~LA2j@jXL1t#e*Qu
z(ePqOT!F;W5bWuT!n~d~iRSP-O%L_X9qX|6&X8<%<t<3fcehS|anSIpi1P!;I&lH;
z??B#tpT8=W^^HOc^E$zWg30rl#jCNgD)ws49m#R($0YYHV|ChGl5I~iAQ$uYREa;I
z;lv#{??9aJYjt0d@tm9wNRKEwlWi+|U)eiavLF3Uara?<oTaFjoVOZejMiRw$t&Gx
z=Sz>7=!hiZnLH1?Ls^KyofbO_GQVwGQvQ0%P29!~J_V;|?{g63Ln`zA=quOow>)|e
zR*BAfSNLU@WC}S<OZysn2V2-&ep)mqDYms_`*tMY<OztgtrUsH+GFq+R}wKJXXZWi
z#Cx&eG)~DyRa|Y{>t6|vm|p4g%3p_dnRmdtf&l_&`u5ma20aO4!`RLXE6WqycneNf
zDCU{;vFo+xmfJ5p4oa#8R1%kj_;^|n^!cZ+HuoKZR}whx?sJXlwf;MBX`vy`8axgb
zX%ElxlJqVHKa)^D3C0%=ys+<K)$4FxjIguUGwSxApI(}1C_J5Nc)O)wj;opYoL3w{
zT09{ye-%)Jl$x|(V$Sz9<*O9U+Cyr`RMxlIipi#Y-UxpiCx(}HJmozJiyuQHUxb(0
zdvES>ef(dAVuF90sg5^%o9=Iu6|+rQ{CqRBtg+m=rf;$pkmltTr2Woq87@1sj=#OQ
z?`bzQe0!FyU~^kw3+~Iz@+31r&dW>ih)6Nzd5bfdy^_n$^2d-L&aSgSL|}-5Rd^=v
z;H-Td;M9OIGWoUyt!}q7TEOts{BxnpY7;H>xKwLJPdD(ti|%W=1Re!woeZ!RthV=*
zd^)fwdN`y|Y~-0?c2LMesri&CyC@;8JAU=gk~o!{p}eKdVcGNEU0t`FnrD55`kJ10
z)Z?bKqm<fBCbtOJ&lqq5P#y|Kl2Z+e*1q3zf-Md%j4@QlQDUOs8Wk;9KCoW$ML!A1
zBAAL)k}R{Y_BWzC2elYG6ec5OP9?FjV&-oJHYY)zP>#TJMOj`l#|;44?A_@yzY2dD
zd3;<oYr08z!>u3QovDix_J$;>It}>t{yRI7nOVxA^G2)tVqY1lsDwr%bnMd>c9lnC
zvJZ^UbziJ(ap)HIKi6k(B-G_#t7jD9qN8}hyKhO~=$59(8ueH!2keA;I5?4kU+|tR
zI>%C|(%(4qD>Z#*`b0{E*(>#wJ>2(PqY5|Pa3u-=++xo#sA+jY!X@B06HY4o?a!!B
z1s<Ow%;wacI?I2%bSrB>Q<2s~4Oo1y#6Khwy>uskFvS1&KpdxGvfF`;i{4i!=Js`-
zaXthUj=g{`Xh$E}?EqjDQW(<;s8E;|`DW;~jQpp98#sxz@W~QrRsuH=%Lqu_wG~;f
z+<{gf6@R{8#Fz*<PCKb2ZhK7r7sAac)~I%^F8a{p7p4Ec@}@PfFt87`7{KwyvVMMR
zNT7y+rmmm<bQ$4mK3{IKa)h-5kMuryc?22$xMTQ?bMMXm%fyF6v=}SdRqc>op14+9
zGCPE2+^5<>DAllS_v;dc)Mp>!2Y=5@PN9}5BbIMa#Idp5M4Y+wLv%CGtBrhBvE(6x
zdu93jM;7nj3-2!3yBGXD%79f)##e}aY|4G<lSCYdNR@IU7$aQE*op_#!v7Z3?=I)L
zEkNP6BUGwYr0S#1S%srY&#bqmg5S+zunaQNFUnhMRf<WZArFDFjD}_%pmjVT@7j*<
zuP^W)`e*G`?!1i87Me=(6JMVq?rdntQ4@FvH!QBj*4zJ#=h`^GoIEFFnIW8Zujf`M
z_@?~<1Do}qDbOVM+gPG~(|xYKBc>E!v;4EOz0$>rjgMJd$^j~qQkY=feXy>70V(#D
zEZ0qM!GG*~vD^>7OKdgE(&%Rpx{Rna0|9~bkPa>{q&t3&P5Z5JwZFt+aRvX>vV-C<
zO@qdJ`JDTqoP-2^7w_vf5xze2!mTcU^0etF>F~Orpl9NrxRC}|`*&9z4peyE7hj^F
ztS$92zD&g19!jKcZ&mM!{c^2W=WbM_lCK!I`<ryn;5_X$B9|5MGCp;qDjLP|Zb6eA
zubiO_0FH0ilCBe*v2m}eyPrj3po#vQK@lx461<H1G17faPJHC#IRNFLR~6Js;V>>a
z1DAD;>)uVfav~Hm1ufpK+`HA{Vid{d=Wu#jM%F#ZVV8mQTC~q?ZJtu8jZ!)lAA$dp
z{FzW57}79k-R|IZN!f5i3-#<3(j=0K&^jwUP|wnetw!17AidV>CE1I;y_Dt(&twZ_
zL>mvxylU$!>ti=MoV@tsWvYf;jq8Dx`(>BXuiHixX5G`hBO~P-yZePslWvT6jyQ$e
zehM1WY@W0Ii!_IW()M09-H@&+m!Xi*s+jm+RcH^4Hq&fY-qzZR&{eHET0Q|zpUJ2c
zeqZ*o;Ewj%1-lIqTSG6cwGE<AW2|?7@~@xVE>y#oSQR@PJ)<Jm#=O=c08I5q;6%h=
zzOY&I1QGYV(507~i48C-v)+8BpIwjoK1+&|!tdR`q(5ZR^JMB}_asN9zbf(VJn`P$
zftJ#Ri5?XfThur2q5VL~K4|V3o^^(nskD+YST5SvbSa?=`dHz_Li|(bbWBc>jfvP5
zQfs@{Sk1R-{pu}0raUfjY2({Qe|>L_1D>M(63l5Q4IRw2gpz^G-7)Ax_R}v7G1wQK
zXYOS<Ih0FV<!w$azbOTdY&d%vS`*uPh$ZJCop!rVY{Zj7+|MD1qs>-F_YxqW3BKQO
zz>;*SKUlV88S&mG3ULkH0vJJ)+ABH8V7ovEDMGY7AxB-ze&Tdm<@h(v;KT}mLhN}4
ztYNs_0QXPJu8Oim;+aAq0HcH~q0|iHIvH07*uJ(=;TtDEuIZEx=?;*adWgM>Mzn26
zz*T{Oy!yUI@?{FzVytAzJpS|_P*2_l<;9ggLWJ<Hq4(q2*@aU@lzOhA7K#9W{IC2e
zKilyWP3Y@2fCcqQ`uJVHT>#P&i4Sw3Qc<o}>}*lLP?{L6Obp6UWiNafcXyBg>2sj$
zT#Uf-0*R%}DAXO$u;l=_zKQjif5^qH2av!gkL%6w=JhWbhWCETRDaONO0o8wPY^UM
zvLv9y>hpB+2#3xUA3o?p#~G!|C^mRm{zsclAl7;YQz_<lDJd~-v5!sn&5(78M5d33
zRi~f*N=B1g$muCJAF`2)A(;z0Yb<OZ+RyiX-X!FKAJvzrHaM^4?2Ws^523AlCyf7c
zJ(M#dWa!X9le%0qrGtxG5smBTI1JR`!*58Qg`qe2l*^}}m-72L;2dti($Gcsua+hO
zp=PxUpK1615iM%$sRT&~F&7)z6F8^<(9rr7xW}RQNYjN>&KHu=no1kck2LR<iK}GC
z`po`Asd5)927rN{Ny8F5jKS3ljM6iAfF+H!{Am^srlKvivNyszMCwW;ved1<aZnbX
zNtyNU?OI;em#0%`i)_3$U88I;SRD<>Qx^=t0B~BBJ^$3SoQbJ+A;~T-XycPo+&aIB
zMs%BNZ1Lc!%mT~}8gemQ-GLy{@4&nyOFLH0rG}=WMB+(~wp9vkx!z!Z=b7u*Hj+_%
zc-C3uS5nUkqeH0Q=XC*BPaxLgNu1~e;CdgJNiu&xUX!_$9gAVU7uf;Um=7kb7)aY_
z3i>TtN2PN_1N{h;zXhPP%ZNKQgI<yOa|U{+6UrNM07#-kzy_3Fd~AgzO(b3|fc#Rj
ze>M+Lg$I%}buK)AO8kKmmCiWnWUATh)ktUb{5<!z@5Bg$n$f(Y>q9W&<Vi4Msmvqr
z;=MvAhy1~7Y`hnIVcg(!%@mqcNFS_<Pp;K{Ah1M;O^=&9@2ojvFIK;S=O$aV-|V5o
zbYgFK&h-0+UF!&HP#hS9QYi7;D23Y9dbD}DTdGv(32Cjd@N^nh9@9HfP2!6xA0F9V
z4e#1b(<NHKQ030W>ebSv4RY)f3xYJNA(pY`+^d1NquoeZR^$%3QuEa!Cj8zP=`b`r
zzc)4QnVjGK7?{o{_YNSW6`s>+Uy@#UH`zi~&c>xEplLB@i0N`#|E%0Rx8^n-18PB|
zDy)jmq(-IaOuAC%JzAB3j$7ZyUGbsOP>g`u+{A1}PiRyX*}C`QzjhYobFZ~)VG872
z9rWlv)jBN3c}m$OH@fAx{RAM#dzptb7AL!;l%x&a+fT<HK6E+#qq7Co$D1khI_X@P
z!ifs|m)jhFCh5Dc7JE8f>pFWvf(6R93(z;GS$^fCheIS(YKAV3IGj-3Dg{Y@{t_68
zcKe++Xt!Ts?t__dw*#<UXfPKBS9`)}oRdjt0)6hfk&VNBBsA>KdG<ap*k^FhiVX^6
z_+y9NoqKjk1+u^I5IGHbFqgQk6v{dGHEa#8Z&m$ATJTBv{dRms=tkfkx?Jim1vF4!
z+Fx&`zGJxe25+c4{+xW-Xt&e=%_=^9&IWI5FK8bcx41FHGv_!SQZQmWZ3KhwWvDll
zKv?B%5{O!XeSJmS@oi}juDTdC33R7?Cd`iR8jge+n+#bkuGSNx01ld+W5Bu%$u5xZ
z#3}dl5y0<m{K~ap#f*NlSOQ1%?y9?=7?FUn-BrD(-dncToTL6Kk%y)yJSo}!Ku6%h
zKOGMx!-jw|W3mj`muWSM2X_(+sI?cKFU^xx61n_sIo)_m&dC!SeyhO*|FohA@V=D|
z9i0KwuKushCw**Mv(r151Vi?ersXTlA5zXmyaj>)!&XPhL-Mw^r}ELwu*0;>o(S#6
z&FdGm9K@xZ|6W1LC^vJ;*8tsOE0{W&8&mk~a(_8hs&$3$Q`T?t`-M_TYIsT!R_6KX
z)HqB?FiIZRRS?T0l%?*pLs{wyUvZ7YsQRU%gW&<d8?WrC>b2S!lvPR)?k|=qSg(1c
z1YK?@u|6LNM;k{IcIo5YDsT7NY%L!cd#*8p<MxJ&&VS|4JFJn6<%&`u6}sdn3i)zR
z%8xXpszEPpixboG0nO&x4tN;te@E@0|J|N3+<zB8aarZ1Ebui$8B@FbGVsF;{ICMM
z_}fYiO0%FnqcdiGBX+|Ta1R(d+Bo_KVaPncWB8#fxo<=H%Y|Kfqjt?V&uvO0=SI=k
z4+6D3{saX9e9@<3@2`BfiEox3ioKM83THBQ(b)Xd-@B{y#$e~UGz_(IeJ*_I@=~3y
zyd^p2UW3qE6Jr6Aj|U=6oG{i?LCZNkDVawd?#!_ednpZ{68-$y2P)SJ2O)h90QwDx
z!kNi?-N0W!i1_<)qxjrD9lY<<@~R5?FgU-L0qccM*)TRcWy@X&?c0|w(O!&po-gU>
z1Yy+;3#lFr5Ss!AZVOOLyEGToRI~ekstCVWfs}oHmQmVtNZN4{=)}Fd4X7-X(XOVY
zhR$VrG>`!H0$kNV%O<th+Fx4LI~*u~U`D+Ciq<_JogRV+9RTh5i2$+t%g03b0i;4j
z0%6kq7<DqbDM_+mPGKgnpYP=>0MFNQjKw9-w&#s$FKR&W4XDs%_NpUN@ldD-LH0P2
z9gUebrV%AwQ1Rgyym9<3Vt<l}{qUUM_qMy?<v;)ST!%{EvDMLK$5f}e8TD~9N~LVY
z+B-(ap+z21Mx;s%vuEkh#Q}WhfSP7Bs)>a}U||3EF%0Az^(=d<Hb;Suk~{2jwWb~i
z+r5Qh;3)`hWgil=KHDU0ix-msGAcCV(@oVVUqE#$O)%u`bwuj@{+Xiv;9#2GM}IMi
z*sBjurQYlL;`Mc010<*AE(uWw80e@>_~2I@;Ol{P&MFR0;*L}fJF_O@svRoW0SgG+
zINzxFiEFU;cj7ZCe9;UP7CHcut%^?!@OrOB#X3|_o8C<RfN!n~;Wwqi1iSNO|8Br7
zj~sl623qe)U(h-3Gi4P+Qd}DXD$M*{7Gwk#5zJ9{`h1v}#n%;dc9*9BJyD5=-5-17
z%QD;|$!~$qPF7{TpQekT*-kl`M`L#S4Bc^?dOP{qH?n($&);<>X7O&tK7Wf!uR7Bl
z4V+b*7I>ixjdOBj68s9_iHK*wrdmOVGg*#dZBNE8I37X6>U#fGb@oa{o`3yJ-rV5(
zjXf_K15!1NfmOCtH{)E-Zt9ucWBN$Yb>@=XO*UZRy3T3~lUu>zFGuCEYKt0>aUf}4
z_<n<r=sVtm(o%NzvXS1NZgviPtp92g*?+~RSbqs{@nTnf5yGaocjNfS8~K9<x0n0v
zThf-)&L50w3!a;8v7H-P<#sVlM|CJ?DT*sAh>5fP&kh3MT^pa*iuk%Z1X1CG0I$Km
z1s&1bBnD?X99Gn9i0Pf7<+~5ZB>x85^R2nZE!bGDGzbUbNC+ZidRtS&yA6E-p>nTR
zZkzuVwDNQZ^FbavDf#6X`F5|4<ZnMqw$uuTzPg#`N%}!jVh`dgXAX8Z5T8ozQL-ss
zn+LY+`r8EBO9pHS+O4M}!84#`srZa&ATB55Ei0Xlgb~IoArD%r6Fh6r^u(2((%BF<
z@b!cdb`svTvVi0Ywf7S<s{!jk(YT%F-_LFoYb1SAi-YaT>^EHRhml{^2|h~H8eW&_
zU)3W?2~kQio2elF627~8Z-UU}hgbS<T2KbujL3~~>(w#wmd0{!+OD3t+LI$HMU@9%
zctC^ZGN4_V)%z%-P`;KRKT^9^aSRu#0E4*Iaj-zvSZ3elqaDimgW?u2uLFi=dA^g^
zYMWVMs~}jcG&T#2fV9``;N>@z7V%AER!Y?C4E3NyfBzZ-azti9I6z7Bspn@<GV%wv
z+5>z3jE8wCned`+IG;OTkmt>^V)Oa~3QVk**_%r}bS6j8{dX0DvEdrLjE5e3>&kvp
zNtOPWcEHaP3V=Z*Gey=717}b8Le2oa;pMB{FsZBjfLyt4xX(^J4My{W(?uWMRh9gr
zLNrvf^_xkGq+x!Y0qC6yS9!NP+GGx1HCCEz6)s~b?LwJuwrE(db-Rb&CFQ|9S^vX_
zLHY^^4G&_#maa|PlDQX8e0{CUpAXZoD*B|V;p*fHx$p`T<mhnZ!r1^7&`hVT2tqFl
zkIVy;BLak)bW#8qKmE^Aw^hArfvmsSK)I=^t?NhpQ6Xt1Rfa-)11kG|*Z59=(<MYK
zDmHO?U7d=~o!5A`&%?&=db9-GEQzQk<~XnrzF+JDn{kAZ@yR@`du1J|0Bu#ze^w)L
zkX<d;Hl;(2ub}CpA~ibO_(e6m(n4`DEG&q@7bIz=lfwR18u@;?sFeHkF3`dDnR+w#
z+d9<4GS=p<uYucuguLAmB<bt?Pm+jX3&45+@LlR4alxVe+Jj56{6*5sfY8#+UI(gB
zMLNfIQ<3H<0^WW@(6V$=v+yQF{f1A|E){_b>aeE3QT8ost(gvix)-Jp1rXA8$c-Qc
zgaz}OckWxQ%=?fCYH{qeY(JR#H~52wCm~{UT8#ey<nn(~^v3z9SfC_DljrTA5&hYK
zzsF4wlJEGfYsn^3a@=<DfAy34_g=2gZPt?@ZqKyoEU*VfBIyYz3y!y8uZ%%~xuey&
zCXhrNU73Al%mrYdX&8-$v>h1jwm}O6Zz{F*5^!y4J#Y{eigu)+sN1KYkgA1Y0YS4n
zqHeasc<HnAin5M~JnZ45G(b5vLvGjbk@p&{HNOvDFj2nrObe4|1)ciN?aevOU<o>*
z)xPzERglg>kG;vUqkg?DskmNAnB?+WltUrsuD*+BFhC_}O&D3TblA|&e%ElO4~Ask
z`>^M2GljQsKq%xdT!W~loKf6@9z*db#$l%J=L0$9y)aPOJ&-u9--hn$b??Esw`liE
zswe!DaVlk$K#8o47;Q)&@eUGjdKDEg>%*HdO9X}S*S9y5Nb%dWijfsGGn9*OX)trV
zZcEy!VulV#dmu1Ya^MB3iP%i{Fi?QwEj##U>AeFJpP;(wEgU*q9Og$IdLREM?`}QA
zBoW!ghNs8To-Jpi<^3c|?Pl2{Mb>3mbXG>^E~|&IY%pIRN3;!KNg^@njNxL&46k`J
zAVnI`DoAXVAi$*$%wTOch&g7Ro5-kWMXJPe-syl*LDk`GSgL9#>AhIj)y+q(;ZeoX
z<;Gigp~L-EsQ9_|doT#+GmXu?XQv^Ftt8=_s4)L`=cJx~5>Y%=O3?$=mII?gLN=fd
zqqY3E&o}9?J*zx)c6rK*9C}+wk!!W|n0&R`)ns&-ElmSb@?jmmgM&HZTE^xaff2V~
zQ3CmE7Uov(=TQ^!9s5Hw)<MM)PoeA@UVTMTxYLCqWgqPcmwN?oV){+o_|yrDXiSGF
z(26}^)1|b6#FvW2jva}Qd0n{AH<+1i{?%zO?#=ue=)=rcbR-vK=bAuvB>O|*r~$5I
z0nnbxf*#LCw<D%MjD%j3;E#$H(+C7uJg=N}7?tcMy`jKHa3JFJ^ywVLLSesODy{gC
zXb(G_ngm_5g|Z+QAmu=B1nuqFk_|wk_uD9on5JV=;T>nrOvx+)b|#s^gRhyvAar=|
z=!RB)vZ?>pt#f&BL^Gvr5p8*y%Belrk+3>|`gj=HX}~V95e2K=_k;PoQ?+$O=@u4%
zXgazwWRNme=E!!gcObLR3pi}qY}(srq?}A=cwH(e{S@@T1XV{v@43$3qFykA=MF{`
zF`!M0zZ!zjWk0OZUfm#Gji$?o{+H09717%Cga^&;4Q-!CxyJJXzI;LX1~oCY=?6F%
zmDAD4WxOq7ptLlvzU!cUtbSdPBPX^g<_T6zs&kPc2XeLGZS-^rPCtdj+g-JA!!FAd
zIJ4@810h$~RU}sxvSeg6w0OnAMLo!bJ(<(jz2q2>Roih(1n-Xz4<C1fc|OD~+WN(M
z^;;b=K~cSjUoGK*7<hnA_^bZ~%)UUl_nFo~?>nM1W`fXRXymWEu*>gAxJ|MXdhTUT
z4fk9B#(l;MY#MtCXoEuIGL~y|6o8|;U(@S^m)^%ItIA<<(O;};;bb7bcm3`k3pIer
zz;2?ap6IMtU~$L;RMNpxa)twYg-gq_{sX?9V)v)pVKX(^VP|_!;u$?beG<lWegw40
zht3~igu#IEKnYyoq&yb(5ry0`mq<s`wOhs2a1^FC$OqORgU-!lP*MKb#w6pbT9O)^
zJKeD1gbF=O{)pdQn##)yrPQOdBW!vEB}WfmDL`_5&>Pc6x{YZ?wuUkO^`@=n>V9xa
zsiiAZW=bkS^^g3r&}KOC1$~DEWlw40<(0Rc$?)-b#=9waYrr;G9rD^=mComt0LC5;
z$_~~Jm-&wN_4I|(N2B5K8}I#<*hT%Vnx87mqDtayneXD;S|A?!%WWIN*)t|!MeDdW
zV(ediXhu<Lt^-Fb>9!LtRg0+@`)wsE6$gb|nGlo?qdK`OmA4Y)Tp(zR1=N&7L-2L@
zHW0h;E|HgR?_{Q|70qsHjt{h@MWN>f`N1)5{Ujb{d7c*@4UdT`tMl6;6kA)W(E~|8
z7!|9W>f;fA7>xB@qq-o`a~#hLNWp>FGpY>qeJM6Gi-NS4>X;(>$$ccy+S1{sz{>nL
z;gOa3R!2U0NI{-Z3iI$M+c-PzbH~C}Cypmqe_;wvpV)t=qtUofrmsZr$Q^qnc4NnN
zj+|N^Z%?pcBnid6cP>jZV7v8#KtOG?q0Sw7#V+NED{x9Z4ss1-X*sGo#5G=rhD1R{
zqTjDz4w~defkufP=P|b0N`Tkv!W9_??3Q9dQwVwXT84(|b*hcuPTnJ>r9R^|!i_JZ
zr{eOI#v8oB@&zE>NSL&~g7Kxb8X&}Moug{oem6MBeEB|9rU_L>K~KORbgqAHG$AER
z1?87jp2sKIM5N4G-&GM3Om*RSfJa3xey6U06sEYPm1sg5bR;JQ;hUTDEIlfnQ3(JY
zP^L$KmS2ICMBX0PVDtpekbfq72JJq5QK7LwL_v{l1$>$)Y0(V-a-9=~QLVVG^Kf*?
zsonIdGghVsbORB$8WL-fqW65uh?OFAzp-pNW>jHv7*s;|NbkCaI&;fK9GcdkD5ZJ9
zl}{C%f1hvWkq1@$^sJZG^da23cMZULRnX5cgjdnS(ByqE!e=|pRhDMh2Fl1R4*lDU
zT0g>&N;uW-(q=KOvpkdxl4aqXWF812u&XcK55gng#xxMF&NDBx<C#$~1T_M(MG{Vi
zV6Z>c2W<$c8x?Nxl10bXIX9JScoJL1IEvKrVB70E>v=owLPqK`oODNN2~75viZ)m%
z-pyjL_I2<ZJ!8{u_{;77`4pNfo;Qz0aN1xPn&c5f_)Fjlz`HU*5Y9QI*5L|0Qnm5%
z7rv|N_e#){J41ksD!wsQ1U+`&?T_jQw%?wr)sD;n4dGq62qNZY9cY1-*TMuC;6BGc
z=VdS6n+%~-;q57M0d4Va2A-1$1%bSJE5N{G;9ngLy50l7$XYGj50JA#c~=ET14iE{
zGGH&T7J~$zbE^S2CFg7^=qGz`48k;AJg$KVjKIMXgaN2sX=N*n=FA#0-V3m-FBa-#
zc!9fX-%5EZU}}zi!xQv)3se=DQr#v`_yHEFUp3k`pQW<`<j@x2BZJq)8?e%_ZKf;G
zJnJL6mvwdN2pc*OfyM;3KR6f_-0oiW;_dW0zn^HS^AvjS6UPtV6F33$CulO2eGW#N
z%fdO44GE~RAdS2bv<zCRfE`D>qr4P6tA43+o8uQ6L00KP_%cq4GrU(3O(HnXfO8gT
zH_)#=f$Ch<fWad`3+59Z^kxy8#5`Y@du95^>E!uHAyK{AD`<d{U&k}>j9uAM1kLa!
zNqNL6{B9$UqQ@Y3r4q^u57dMgG4OmmR1B4#GutqWp1D9@0I|Q|Og0O<14I#kl;8Vr
zuuDA3mu>hyEw_>Xd~h1U1I`N1T#50z5IUd>Y!BGx?LmD78#KsH;b<o3)h!}y=(!_{
zig22-UYffqK{MuoVP6Nwx3Zgbp9^f5FCSL~;7fO@lW~yTofYtnMp60}X><4n9HIh2
zGlY(V?yZ{LKx*iae?3dOgV3x(I^g=Y_lpla4nPBe(jF+C7zb&yfIjTgsss%G!|4+z
z0-i{*@f?CO3o>$1ddIm^ep=iH-63mlXO2pBxXlDdv^IYYb?q11FmUSzMRc>zIw?&C
z4zRXzw@Iy(n4ot>fF?CTTL8|0lItZ$$O@}F%sYp-gM_%)HbVjdPm!zWW@BS!IwwHB
zos?DQZ>+!ojBzccsRP1Le%9IBQeC{O-z{NfphPD``zbJPsNkoLJ^jg=$^Amg)!}<&
zV~)d~nj2`0Uy3|uz&ZoQfG}|u7;(ZqxSwSr_<%b(e<K-Af~_O(?iYLDv7@A)kXCk5
z)N>6cPzo@Fk#uF_)xS|aX?H1bXBQa7f(GXwo(Zh}Xmm6VP98S(6w4)R_y+c`vF-D%
zmumy{=!>xlFGAxSeWuyaabn}?$vn7ZF;i6T4@08j^od_6M1#WQ7zCHy7W-&n4i1nq
zAZW`Bn;IXb!A&u*=fLNNd$u~NZ@};l*8nCmZH+2tlkL#!7t4;3ff0XI4}ObH(g0kz
zc9sFlj>_feUOgNgn(`p=fJ$>FP>aK!osV^ud3D!nr~(L2N6u|^^xx=#S6_}O86Aa-
zB5#!tfy-bIcrgvV{v|XY^s>gx$k{2;JvdZF6N8;uc!sEd2G%<S;x!!HctLkVuMTil
zn+VoCWK-pt+wBLyeeo4X)fI)M6TcX=;$JC1&wM?)^(>#r^4{MY!Z-a}O~Jzn{s36+
z{|6TuVK@K3cGJ?VCrr`jmrv(t^}E$W+!pgK%-tbTLuWnkPOD1kUC9=seN2hXz_vLA
z0v_#RG4F@g_iM-lPBu?Z%7{vQf~#zT)8B}-k#863@L9Cqn68i;?-#!(VX#I`>=|p;
zN-~Pno}^Z(Ey$6tThG2WfptAT_reX%$($76Iv|Y=o(016zDwlhO>db9?WVOJ=eb+I
zM{hN8b#+*}M4pSFrD9WF#Sgu1G%@N~Fk3psfc)@dJn-Yowmb(`%|rErSFyS>MR4aZ
z0bBhsx8%!0@~x)tgAUh7V@U+t!CFj;_hrEtpQTFg+0ZPzQ-l&Nhh1OGorqf;<mjG}
zPkcPCYOup7!4J;GNKhxj$J<RiqgmUX)=EQ#3BP{!PK^&_QKHA+)1#vUAF^rbDvbAc
zDpr0td7;^>XEjCg-Qrq~*qm`r>4*d&o=%Cf!od{QgF4F32z~g$@2Fsov|^Rz{SF4d
z#|^nKr`0BhS{f3t?Rh0|oS5p&rgcvy?*qlu$i-gY3Ct6z$Wnd3$c_=uOu!1zJ?|<>
z2K81_zgq<{BJ*$)_q9pF_&tx9gTZwMSQrxDcSewK3qmGF<r&AvY(L|B{9ehkW=jTH
zqc_^d4d8x)(q>4nM@&t{`YvI+O&^>h+>wk}txV>RsUA4HJh9~LKRzCT3i_`}=<>%@
zZVIMhxo><mz`~(qlE|SqJBRHe3R(B}K}34|q8~u}h%{;f(sxso&?V3HC*r^6`10g*
zBOR8oYlDhLZ_{*ePN^0{^cwEgRz~E9Il5b4l?0y31uGkhz&YOyo1l-3YXrqfo(H*J
zjB6hS+}|z+iDV3>kq73llwaS4j4?6nGSkiWAH95`rtb__N<INAL}t6bZGNK_zhkfm
z(7MOZ*&Z$t?1{VP`WY^lVv+6VXBBO;>SB;e|BIw0<kcWQR|?GeaG&`HH~M`Ro2=7w
zewG<!&Q34N8m)E=%LcjGd6`HW^-pcmX7l}FiAvixvE{XiozTh6c|b=*U9*V$@~Y%k
z$DR;aQGh495M7ZAT`~8CDkSdrQX#ICFEjCso;6j|<;=s`%Ln)l3xXDpmWyqtxebVQ
zri1T;+o5hclPcfu9G+oX5}7&&_k%V=%++DQ@$d62gOO4U<^D@jw;6ni0Scew34hPM
z4(4DiCM%ErYn>_??|IDC6j~yz>w;$L{1x8yuR2-`zN*C6iD}z%$Ya0q=SuEFDR0K#
z39S7X+wU=^g*cSu6XO}BMYD)^biKlm?dW~p?dMPLM{iN>p2@<-%<pU4Qe3(bw7#cG
zS>D1}%);9>|DX6yC1~3!Nx1{F?lP2-A~I#OX&l$anRwqZe#6)8F8KK~M-_5$Xm{1c
zlb;xi+y7$CnfqC?c_$db<dN4XJBGXVOR2@n$zHgJATsL6mMGWS$FfRNHHi>F*s1S^
z*|s}|Uk#otm?}Ht9&dnkbmeTQl=%1QeZK@97^SZ`C?Q3ekyxp2_fpfZvH$s}*X|&4
z3dhfBSR#$y8*KISRYHR7AHq;Fs~{r<oN|Z87&wrW^v@ov0?NTPNUoVI1MKc9`Lyh7
z?8D_Hh5yB5tE0(k6!`ERO&jgy(A&#p#I{0+3(Aq>esl?IsHm_YM{1ex9|_ye9(zmI
zOKUT^Y&Z>X%N~~xP?E2i3v&r8Eo^Th$|nB%S^fb1RRajM0ruCl{-HlrP<jp{Vo&*}
zS~hp<XP=X@MDPR18?8Q<rec4p)@E8<mU4%)W%qb&3=Bbm0(;TK8GuAnEB!DL4ao<o
zrM_)`S`)u>@IaXoT55L<vqIVSBNyj5U|;;o2y1{Ef1Hvug@nC>;AJvw{kJ*-vHdA{
zQaNZRa%2_tsPC?tV5Z?A)XqeEsi$*xJBGg;M3<BnhvK;U?{n}Hx?f!o_HE<s;Lip_
z1xwj@_4k)I;O-uLZ}JucGNu6Kd>d}!{)D?5=#ukS?uKsMRf21ek^9j_H5(J5KA-<<
z=(q~v0au>o-Yh5K0DzpR8xbp5jxr6@BO_|fzh;`Q;C^47tXX-YQ(G!TNCxIP@|;3p
ziv3dxl;UEiOU+(Sg4F+d#rY=$HM0Yi$DmgGKiO$9$W15iS$DEaN&H!Y7PSDBNo1ye
z9NS<IdPf4yUeE$OkBS=IeY<lwZJH0P;&PAoYK)xPOgpp*MF!o?WF72*8r1a|AIRq`
zzaaPK4mEH}(@)voV4(OzBIG%;eg)ZbFl^^A$Z$2Nndq$+rg^E-(B{x!ZmxIuv$h+J
z|E;JjHF&7>9Gb@fvUd9CO85h0-Rr}Gxr37O?yqG2_uy_FFODjHZ*Ioo;u$wD7pGBK
z&Vhr6h5uwkNR)z?`^*dO2d$7o_sW)h-kWkWO1ql!Zgm7!{f)!Rs<Us>>YuG?kT6Xy
zN{ig{|5gjn0UsuQ)~9eTM;_kHS|g7;I`p&JxPB447(Rvp7-&bHqftc*wL}9v74w{Y
z`t4`TX7b+2n|#~HKgGZ^y_vNg1+nncB(F7~)0~4s&=C%C1Bg}aqy!@VV@$@ZgON;J
zQuc7j?RgYO9!&VxK(*E28aVKIfj({BEYr<43E_m?*i)K=0u6a=;2jKJa&I=1Z5t$8
z!#0WTv73mo;s+1G0G(G9qsf|{!_H2Zk)rAI5*I+L^n#yvIk;HBJ9@)50yc#!<#Ru)
z6l<2+w)xoV(g_tLqH9R|?zz+dKFYCyqK_+zK6B9mV5I)JbN~LOR+rmtRQy6=e&O2`
zD}ygJVCdQ4^aGC{D0&dUm@8<-i0-(5#1!sR--Y=L^5xf4fn31)LyX0rYco*uUf6I_
zQ&3P2hM=ExHHFKC5JeaC=yD0E%q4wji*tFfTKD~QMr9_n*wzGSt}Kv+Av#9>GXIZ`
z5LfxpPcT3<_3&-DlRB$VB0ZwFi}wOQx=X0<Q8_N;S1m6Sa^YZ3QVqSuxK~OIo%giE
zpFwP)<=@;k|2g>TSZzV<3n4-XkNJKp<VZF+AO<XKY_;@NR$6I(a;bl17qyoESm?RT
zpD9HonPX}*aZ*7K+1^AMHNi2?hL#Mw9*1qo$=LbB)zv<~bKl-UbnY6*LEe=1H`vyw
zg`$ON1U<u|{5NH^UWo2SV%PszdQMWqxXqtB=^)+^@~E?$eKI!MmT5^<!X*;Q&@bKD
z*IR&3UU|@dN3TJbi^Y$Lvmjup2@%R@?;4-zY!14Np_aE92IhbG+%eu$2`vENX}=p<
z<zi^O?jWQxN<HgT&#64X^r>;*_Tis_&TQ?sn)tO83$gu6pIOEY`drl0u<f~C*08nQ
zxiSv}Wu@c-ng&k>>OVLKvO%uz#^c_it7s2E_f^cO&r@f|aG-s|TDhm5s<(_NPn&QB
zmfUH86(qcvP8XT<y6G(tB$M4A%3TRC=6Iu8zo2^GU3}NACX)vQnc7@2pObPd@FQ7c
z2O)v4?ik*$9dAEbGZ1^LDU>pt&KfgC>f-Ir#t6owr(y0G$rqXysj@aw657*zz5C;&
zl5=I$Kn!WCmywT=1@XBn{iVXl?*YHs!A->2Oq0Vx7K6s*XMy%=!%rXDdw%s={_G%O
z<`RkWv7aSCQWLPjWkwN0bt)RFMWvGRmJ;t=(0dUN?a`A28MuTf!dc39A@x%x-@O9K
ze^co<Ww4!uo6O5ODb6RGZPc2Ea!Uz{Vtc)pW430hbJpid!FMj+$1#)nnOMz^#pFAZ
zMRO$|_+x^7@COw(ev}4nc4e-uXT`&cjQTj=W}B?Y6)wCa@Hs@%@#LT+@`)NzF2g*N
zDlStu%V>Kenx>|=45mZ1i9d!roT<OWPX#3@3y+aNi?{a{r$!QlyI05$)<R7i{0fW*
zT-_p6YF4i`YPpE0ixMhuy8W*k$;JbF{g!$5#dQZ(#s@1$*S+m^`yKHu^cTqp&w*lS
z`H_Hxs9EpH6>Tm*h46zWD+U{U{E~+zrV@-jZLW#OFkwvm;+voLo-I~BMIe*0VGy)t
z>Xyn}>xW)2^2pO)EIHdlVmegh3w!HE!4IBkDLOoAYV&e(rP7mChq*c6XFg8CH=b@~
z-Ns!Xq7^k5K{*|xlqC-PA+X(FL6E?2%sgJ6J0XTL_YB_KYd?x#1UmcM*Np`}9Ljlc
z*IR~fpiInLA?1v3ZS8i?ZPGU^`+vApvtLCCF-+Eh7JDB!+u>bIqu#u+21{fs>5rWK
z^XY`!<2<=C4KLC(Qalz~7h5J@g&MTOv;isJ5WL!~AxiLIFXtuxrPv37+dZ9AYmsYq
z8Nut2zi*-|WN70@+PiK8qs6J0LX|M(@xN}X_tL^Ocj1NyR$cgGx>+}zGr~gSSR?%H
z%t@jdpErugX>=mZh+dhsBF4HH%pXr5|I6j}BISq6iXm;h=h(T%vm=pNytfZfXBm+(
z^b5zkx2%06bk^Vd7_r#3hww4NWqKQ7*-wFd3*Vew0;#61;5|p8Mv-l3@(?hI3i_R<
zkzhHAC5>^)x6dw0IkmjcWoXPn7L6@$CoQ><BgQ-KTb<M}P^>zgEg<fLS$pKHZp=Gd
zrx?Ciy5US4pjIeM94GjA)B33uiay56DO%%{i}z$4CKscGm2fS$n&``ZIu>d1urvPc
zora0=hDd-fKWv}e`N|-~KF(|XE&G{_L6}hwg)`}Q6I|Gr*A1@y(iN<=Q0`38>5Opk
zlMIqo^wfOUXu>;3?uAj3qf&F$WNhhae5Xo{_7%l4xKwl*Q4;f7TOcXOjIrc2oYj+<
zVzI3aXJWYo6PHjLO*s5#j-DoUO1nZKOpp+N7qgL|XyFp6vf=juYg{SLUBBK+i|>~+
z^<2thL0*fZ<s?ykT}kEp`RcK+e?y@tOmC@wB5@`oR*;|z0J_>qA#2b*-!5v{Dxjuc
zewq@om>yB^m;BqS*!h6?Xcwpd+e%6|;&=rE8sjbS<9w$GJq!2QS2K-uWwo1ZLruND
zfK42aWwe{p6?B_onCPk4Y7d#XX9ZHb)_y_ha{RloX3CA0MK2p(%pfra@vulU8(B*-
zNY&gWGHAnh4swKdW~^r*?4B+;wEjsZZS+Ly9C3Z>X7^IhV$aW|U>Oy~pD;b#fIu<l
z@CV$qux$Rsr^|cB&xF|cMDbRWV&;0}_}95~>aS#>O6m`ZLEeM-V+~OG{<OI%*2O+C
zb}6$(8duR3^a5WL)SpZ_y$l~{v7>Ge%U)CG2W=5dK2A$a;Y_x=x;{$yW_h{%@t@J5
zCC59Vwr?9ZVW5z32Ckq~NI}}6U`3^Y(7G2=cElr_#EjWsxgVB~mRruV0u-BCNGu6h
zojZOFwoU9Ir}|yTJ-C0amz)UW&h@H&p*J<o<3E^aEQbWJ@2zU#VVfH~NPCP~e@p*s
z{L#6I$>+iu6?S7xoLxg5qW5?oD$!lKptevTl@l)=O-`@uzSYFq%VW{4R0KO$#$*)g
zZ^{rLcLLmcip#9#`{je_dD@+MuBP0(7kmm9&WQavvEj@+|DB16qJpmTuz1NfriNlE
z7!zOYf6-ee2tPUL!HxVl&3GVMNiZO0>04$^WLR9!x4x<RC-TtLx8U*%@m$rUg;)FY
z67hD(krXH_HXmO7gEXsG)JD6Iy6<{=i@gz&L;I=rTM!v#M9F0bMug?Q>4NDuQZBMR
zfCn|N+>=fL)Z`$G^&JRvS*+Q=TyI~o(oo$>l0;r>LFNV#NO2+oFYBK4lB+rS1B4%M
zLu9U_YcWtR;zVm`ZzXHS20%o}j3`vb^&=P6_dxvOqFsImGBEZ;V&hS%W$QeO=Kq<&
zK$)*{F~z8In>yRyW}wVH>Q4IC1I4n!RnsjBT;v<5L3yL(<!**o{y7W*2FJkM%_^5X
z$!-A%yX`D8CW7ww2%jH0mEZU;>Iv|=r1!!vESbso#K!mSi1SM%aN>)FZYl!roTMBL
zGA2xNK~yCms@2m8-<#Pus%pk4PZ1-Jy7g93aM>1h8APU1JF{1rj1`3|c{=VJwN3Ff
z+;uu$!+(m8Cf7ix#aq%t>Aw|ev@py4^hE(CPAc^Tk!E7vXc>3l`0UNc6VD&TW6TVc
zmx5W5u|r@Z|JL~xO_|+Q8tk%CW_Dx@H9fCMS~!0$Dy5rF)@-6x`#q*F;uOw}P%wQ8
zKcBt2No$56to~Tz{nx4zv>Pybc7Hd1K|K+xz4-SBQCD&(j#<KIc9ME3^?(9hjNaX7
zNYj`V;ttbTyhd+Ulpy@tOd{0X0oa1DMh1!_t{Fqkql=Br929t8Ii;?cNfXLn!_Xes
z5iL#e-y<oJtXnEi1uz`(V7_nV%(ObUi=mv*Mu*_W{$5`;7v9cE^}E)krGa-%&34yJ
z#|^qUb;+BXsyp#cdo;P+IH`grK^7#z*pMgf_wFjl{G%%ot`>ynmyFjac&c7fIEe}C
zwP8)4X>MM6Wff-Wf+2Fn;e5Q8%KilFrpW#(vIyOIR&;7t(@Q)x0z#GPWV}W_g>FTU
zhtx?A0!@K1t9wppWAM)&pImLLDZ7l+(n6Ys4cg}T5+&_Kwgl4@Rt|*Jwd!duPt9{5
zM$@n4Dxws|f)UC7ys_F7-x?JyBuh2D^Qmz%q+OjnB!i*O)aTCJ9InCseI}ct0~D|X
z04~huH~c*Rs*J#jrR&v+8WQcsyN``?4ec7_4NJP^?C%3%`gx=U(ICWsbiH;nk-MQ3
zotyFU^4yG`*5fHj2O6==1iIF&W)vl^pdh{?w_VW)ti0_4@v~bLjo(ENq_Zi$mhmXb
zs8D<e68e!8H2BTJQRP{5E+`bK#fN;$-}9QRUY*E(Ma*c|aCtsl9{*k<p4m@-4M#t}
ze9U9zU<!4-Hzma%Ipq7RL6Is>Yc$5ne3>N0A6R<U(_QyXb=&;c3fcec(`YR5tPa@4
zG`}9dj%#`tE)gF}T!JuQXww<DMstEJ7w-G}=>?z0QwW%~^t^A66k4~{Xr)q0aOAxC
z{kiK4tz{y#+dGEw6_o=2?gJO&+M7pFzAQ_LFEBUEuS)Jn9W@)4i+P=aZXJD@@Y2yR
zQ%plqnDR8_;UdP8@YXsN@n5Z)*@PRk_DbRn`^h;B@xjy+gAX}=jbW(kQJO|(FN4T_
z3M<k#Uu`A?PL8}(<A=F)Y6`a=Eaot`fO&uy8qM?TfDt{(_#twH2@PLp2<gpNWUhOI
zbZF#R$yIaZNlK(47hR=@z7{k2LGtA<49)CrEO)uxM*zbCXhC(EWn+|-K(mR^HDlwh
zU!qMfvK8N*DBl4ZG2j4wq1#ewnXWIhxU{h?|8AG>$(2Dd322+&rum<3R_PPeerWV!
z`>Uip))g(W;fA90WUOC}^=2+w-OOtEUpjjY={as#FI#T2tm{vx7F~!=IPu>SWTyOq
z3rEF*?-$U!aXYEI-(Lx5)(&f&p+<*Be|jIUAjIL27QfJlc_sgLB7S$(jEIclCfL13
z0#6o|FBliSl#_yz6Khz}G;RpI!FRYCb|YBcB5XVlSgfV_+~c;*N*cAE@OoH9kbEf|
ziUKUpaR|XPWau|_m%+OGPbs=g%$A<A4=(n+bF1rU$QvqCj^q)x1R=EEf{K6Re2>KI
zDZh0-D;T$re53SWvPv+nu5VQ_1-rP4;q_@aX-i=5ZohCWPf>HE3^Og-G@u~3g)Vzq
zf_;@b3w+>}f6<>$TT|zj<)d~Ei_-8IUN<F9mqU(OYK=My?^)sXnUHoAe`|R4^BMyg
zjQ3R`ui>{fh-adWpV4!1SX55pqyJ@QsfPLpaXYd(aog)Q8CGcl3jnpGE#cdmq>1@A
z`lw2?dnB9XI!IVTxY~c;1($(f-_-H-0ZOFA%xBuQLjD!_$oL+fAeQR2!2T5*of72#
z0lP~3qC87lNU|h<7QI^XdeELShx?niq^=N!McR0Fb5mMTdpzv+L9b&=)Pi^nFrCf3
zhP*JVBIC$J?q<cIv!Zt8LKN-6v%rCe-~5*U_VCbRJhuq-9?H~#4b%qG2;%Qt(xNPV
z3GqpHe>{knUUX%^Qe9NbFh2sN6jNB(Ex)bP_xWHO&jKhBzLN<mxEkal+uo|zH-_Zr
zy}e~fFfA=PtDC#{$r!H>2IFYYKJU-f-7y?u79vC2G<Ku8=`S{2$@LhVw3u9BPm<lV
zp3!<XYT1U}R>T<+iIVU6``4*`xjpf*Uf#*-6XLZp!GbhmNQS`;xnrV?LC6r$z)6o?
z$d5~m2Nv1`S-BH$p|+0Z(=zFj%W*_56->^^al&AP+5OgsiwTPrF0>1_qRGpt@W~ac
z7+0f;4CNieQgIl$jl*01eDWPPgKyVCt^)EbOp9u<4HQimOuUQaF8fx;w=UIh`%E6x
zKOvTMD8E1C;%-Rcvo!DR#hqhcCHu~vlNy|*CO{cu1p@dL<*7^Unn%rY#G0f{N4`C4
zJw{6k6vI}}dObAXP;H8_+~uIlJL}2S1@tR$b2e;dlO1_hz#*+wx!!DP;e-U$c40}6
zhg_yGPx;c&lPSg2KLL(+Ck-9ezU=Np`Ca2BPlPopr+sSkzO0m9GPc*%C683b!)Z+B
zu_jeff?otiCR}x094|}AP8xmpUR{xxpN(0;z-9CHo6-X#mZQK^pZib13B~aMeh2Ef
zG0lA6PQBTxM_XE}8()ta>XKhLD2kVwZqiJkzc3)ffnE|-F~X>+UnMrisL8?L&6aTr
zDSjhwQrbKnEl~@7mJ9o<Nc6d))=)O&HDumcri0&%zUB$DbPfN7uvXK?OE@1t4M-B*
zbLsps@%raw1PQjK&z$}5U(xeE&L=NqtHAM7<~>YD-1Co|0SrQf+1<Mi>iID`UhnbA
z2^zeTXE#<da$h-!5BYHu_P#6My1hyI7<cw?--iz$j(BINE_B`KZ>ui>@qyypD$&Z)
zWLmEzgID}OE0wK{M8|kPuP^9iU9map)<<CH7w)6P=z@AL5pA!tZ=_mx+(&+Ffz|JZ
zXFqY<m`c&WN}u!2;0dCN+5@TYj`NbPz@l@}ohBU*=5b06i*2_kyX5>0)^Up2JxQrE
z85LhB^eaJv!Ld+~jht2g!|AVZh7yJZ`i76EFyJ<((_6QEI*lqXrlJ|Kymd;!bdi{d
z0QL31qf#{1ClqV17d)Q;YN;!qFWr}dT(pNGYpz_3C1KsV^fZOc!o-^87AcF(p}Z&b
ztcM2lEG@^`HA;6^MdXu%bW<f;<zRkfr%(2`4l(@m*+$Z1qjFEDrBij+S-3<(*t9uW
z7~~R{NLr-cSIag@DC;Ylnil!r3vv;$qKuent$6R%L64-MYj^Kec)Pjg=l1kkV!Mbp
zFtM~IVn;^CdP5+;v9)xaOrcy)aFT|F32}E75NVHm4-latyPC~Mf~zdCPBpMsiM1+)
zeP2wJT&r?>^ekf~d{$;n&ty=SC_;&0*iZ)(A}XR(ep#9LiupM~_VL?DmEltHVfwS#
zcn|O%oE8%dRAc$>JXb=v8X9sYoP&W!j<z=1dQ8CyCS?-!d*6~Xt^Hdc1}4ISFNIK;
z22hyv3J7@oMn_66A-YOKk?Oki<HrLgHXeW?ca~5w^r{S>3Zit>tS;TlGTv|H$D3nG
zwO6{x0|W)G!zB-w>ook&duJd>?IlF%cE~}olX(C{8&o`<jPa=u2@2tnIS)6v<`Z`A
z+?iOr6LI9(diMA=Jw3f-_3gC31r+v7d#uh4wkq*+4Y)MRxBjlbj(#Naoaw(0|9G^8
zsY!ly;5Dn}MRA~=7iO|))7c#kO?9;P04kiH2wXvwa-V**p6Gt0wQM_*&eNHuoj&W!
z<?4B7h`tNKo?_Vbd~vRf<WG>rTgRm_FO!Qf<f3UQ4|8JKhidlu+1|V)6=uZI7cA?5
zLRV78`fp67Vy%uQ<Y#)npFux<#&^%!Y+myU98{s})fIRzPitT{+Hba2S;o5EoeZ;X
zPgcf1Hwb?Q-p7k{w=*K@Ah=!efeaB)#~yG_%xT)W=`YQ$J8N(>RFT(@-5Fm(i0kig
z;FR`9(*RE+;sf~xC%-ZvJW>q1C=i#jG7l2(`1iUCGZ}f`a9jqCUr|iYzEQcOR^07f
z-rryuhx~cCeE<jaXq+sX;|y=z+Rb=OJCAQBM6BuB;38NJ>#VL2`;)SjhYpHas7Yd(
zdHmbZg1HjPj0|su11)>_7Kea(`S4?wdj&T8&1-|0PcJNIW8??g%VIeaqeMTr<>%2H
zRWo&r{Up3E!UeEv$QHh1OY?zA#krb{&lV^hDGKjXyB}4r%9Cf|M*<np-=wg`v0U<D
z&u+o|b#7^GT)Iy7D_$m~N_9^af(=NC_umMMr<g6`xca{7AXTYwJc^AZ_7~CH6{+pL
zmhp!RthIK2I$YWM?HNlFi4}3&31=e$55Z6Z)<s<l$Hr4h!_mn&SrQjwb%yakK$z2T
z)gwOId{C;|O(@!#!7`6Ek=E5M=r5~*zg=Bw*GS1aJos0W^k@ifBOSL;XbBf%neQX&
z+JAMv&PXuS81EeJ_|=A4^}tBk^@)p87!mk&Wm{ULIYo^IDtuP&H1gtFD$wm-3&qd~
z7&d+?WYtw;HsnJNEOV&+&jwH-ZdBAYL8m9UwP4&1I;}UJ+G|X|RL$~zL;>AIDEJK5
zZHV|5hFur6ydXuS9`<~39Qk%mWlsfiaS=iRcQ*ge|9&yB;{R*caghH*zmY96^f!5n
zPZAw2T)~BnNK-G5MMH8n<!4#wAg>*WWn*U;8dNr@{SLgvE9JYZNb)#^Qn=4}#FF{G
zLL2F>`i^DtN-RFK5bj|)M<El{hu4=`68te@g+F9qvoHGZuJSF$H%;?0uEj39B%;59
zj7K0~eRpa>Hk%r?i%$RH(|QxO11<(|SQ62@^rFo7e`WZgSLRPJG%)Fs;fGy+=rBgI
ze_7cNSHQBjp2a-ad@hvv{?V>VFnb{eIq7V*e!$ikZ17S;v`9a){jE2j!7<;T*4tdc
zhcS}wznasZxw*1D1J)IXAvfaHtp0qI#60{lZ}S7Y8Kqk{KYu(3Q5WA_vyML?tiDeN
zyh+?K{4{s-F>yxegg=kHMC_gQJrvjfUgYM?ju%_bs?XBxp{uZLgjA;HntZb;KIeOC
zfPcWQJ6-fjeAl}_FTF<WeaURIhDGM)gYu>}DZa$V>GzV8YyWPW3kyz9e!mi@#~HJE
zO<$=xHfT^+0kizBrJ5ww*juvc^WJNSRJ37s<yk@{-)hvt@@LO%A!aAp`9gK2IXQ>1
z-0Z`IUH3Xt@#QNLVtuahx^FiwC`PusG)|Nq%6YmOgX3X{f;Q8TzstC($F68v?C$ja
z950m00w-UZ6YCnk#QPSI8j0!25k{VFmsXk06S9oGOg)wiHNp$BVGsY_TP5HC9;q4o
zn@<k0)D~o=+UC#6dk2cj<J{HDZgr3Nyu_KVrD6phZf-=**7-WmU+b(9joGX+_q^;>
zab85ziauAd`zEzNPuqQ4rT&gRrg8MiAJd``g>%{EIpj2brj8U+$FF3cy49X`%%xi1
zuxXu!v@*+v1P}JBp|jGP*>UYWgKzne;=)iOljXZmyFtI(bsv<YRPjs92M6qYHR1Qa
z*{x^CE3^OpznZ)9cqsR_KM2Q>6v`3NN@c6WkfoE<$zCQTThXzFDN7i`s8duD=a6M=
zC)+gHW#7h<MwaZ1oel<L2HE%bx;@YPe4c;b|K7jv`~EH0^1Z&-HNW3A(f^Wt97;p(
z*oB9Zp5>1$7=0>Cn`6Xkk;=tQqG-cgdg{GhmH#vnI=}trimTr}+{>nqQ>n+4ZFz`%
zu-Yx%YCn>{EnMO4eTeJ~#fFzDv+}#&x;PsVD((F3U6z53W{g+j?RHzu-nZ(MH{!U%
zyCd#lUp=Azqfv3(me<$9Tx~|mI)u~y(6^Um6cbNB$DZ65hH8{!ICjDVWYmCPQH0_;
z2HoP3aU)U0#U?&x9D^|>{F&luZgSnQ7O_PX;bg!+$gU$&qXOT|D4>J8%PwtCseke}
z%lOT|L{d$7ciM~lt0eYMdo{WI(Lg<D2Z!53E>vJ;q$SLY$0`1HyK`}|uNoen_L^hH
z+iJ(SV-zbqs%YIC*{&|TPledTzLM7U^~XEEvzZo^x|^)L3mwd5bO|TX0S3Zuvlm-!
zcR@=D-4=ODn&+A#r?A!FNB=9?+DAx3o@Ubn_h2`B1ge(O<S7?!+h}xQow#JDJAc<}
z1j*MB78A@&z#2}ETJ=1+Tj*H*v}pKEf*N=lT=?IyTB98~#=9Xu`Aa%=Gx{dw{Zr9C
z|H<zQxZyk9pjhvQV;d_f>Ts_0#+wxGIS{stg!wPTocF&hV9!b(f}=!YWx3!1>XPpQ
z_hyyZzuAi%P(9klh96;~LH#o-1s4wm`6&H?ofXqW3t?xL&y9Y0%~Lm;?9SSlM=T7v
zOmtM@)TIFGfYRx);ki^gHSgR`niI}(t@zmiFDlkhe>k?^OVqlP(Pv=3&9HCF0tl*C
z{mM52ib6bRTNc0cUICQNyijP|{6;CaU({~3ec_?&Y2AtweY%hy{A#7+fRgauXl4mE
z;98pK7=ms1T|P##)gFW9fEP!~Lgc3v#u?xZs(r{={I%on+0-TqufV3n%eWStYr>I{
z>qltJT*QVO9_r%sPWP8?e08vgrHs8e$u&o8_BusxZkfU;BsZV`w#PO4QPA|%?_h@#
zYvV+RDCD)u&zx}PeJDMqlk9XiA$xQ4J%%7T7ai$0;olToMdm5MR~yUrxR{ABl@S1s
zDn=o?5w^12Z*;EPmnY`t_Mp_kI@iq)yY?r;2ZwdXf*x?rfjTo1I!AvAK1SbKv~E3r
zk=a^DuXEk`&{hRIAG%?JYOz*NwZY#KklWn26@K?;#`8#ufnv_WQ#g2P5%o7pV{SRt
zl6uI0=0d&CW#bNAY03PzgWfabn6@+qEy<62`Fa`L(LKt2>U4mq#*^RHifgni*C(&i
zwT3y5NY3OZgzFN{k?63s;AKgeF-!X<sz+}+V(b%PR}WWFD-|VvGt0JyKTZ62{9?h>
z@YTQH(?MOH-wYJ=>s1IP+1T9gS`@6@e|tLhAK9IyPY#I?b;cr6_RR55EM}Byyz1HB
zYDJcW-+UF%6-NZ=`laCqMc*Tvh<J$_^13(6X<9Ht`P){8E}%;8aWnjGQ$LE4D=)n_
zzpk>|x2pcgtDi;`V`qQvwOdOM<T(PR+}0$6-2PB|IpSa;gQh913Bg!J>HynjGN<2J
zk`Pr>ZnOj3NW<N5;`PlFtnIsX&%a9yt6b<S0W(oKzQeKFTE^W-J(Mt+MD@9RP>|Pm
z_|BoSd1a}kebW#*_A_ljl@?(;Y~`+v4dJT8iWHL5UB~gxaDs6;HKqMkFh$}&uc$$o
z0ZbEMpgT6c<gvmHH3Bg5?1<Gm$iuQ8RBZB(yjLjgY;q8USA6YyzgqH;@8b+yRUvNu
z`xQ0*k*l-f<yD+v8qVuO1KUCzt4X~szIvzK^#LQ`3Z2ic$oE&(MkPoDqXpTOXJTj}
zQWviZCkj_;(rVi(!F}speZ;F8m7p_AHBXhq1r993<29RD8hxz&jIMmhvLfjCN#g<v
zyuFE>;TcEmIL$U-?^>2+%gbNkdk$$Of-;k7for&NI@0(2CQ~9GgE=K34!R`g@b6Wa
z8cyF-C0WY!>qTQOb1N%$KZA+gu4!;1F2WmP&ojv%kFn;wqzIw)O?0np-#@#fCUSTW
zKdm?xz&$*q((Np%$BcUtB2YyxGE(IOh^qHrlc++r@=!@I&F*?EEBv($I^pwPYn(sr
zs{sehkY^1ubPpQ-JQ7N5AlDgp=-8qY20TQq%bh!(Igk)Br-PUQXhS(;c`)a5b5@%2
zKtO24o}zg7;m~BHg-KZEkIMnJ1$IUqzdNT_qKiTYz1=XFo(Z5MTB=82gXTOVC%=Ic
zzED?ixJ9BdrO*ZLeO-FfLQD@Z2j_9Ik&^~aQxgcdT`m#H$=Oc7_OI{8-BIS8>&eL4
zU%z^hD{6@28tOhXL+8L!M3`RiZAQcno7fMTL(Y<&Scef<nEF*#2H~nFt-sA5ZUNWp
ztcg`!`RNlIuj-d)vZLXpZrB~pkV11Kdm-%L)`R>Gz@a>&1nx&bHvA?EgE`G>NC4jN
zO+{DYN8tbuNDx#%oH&?#V3-jrfl;r>-$+TKg^-m0^AIyznnCYuzms{&=lXypfpa9-
zt|(;C|7<wUumDb1q4F#exaacpd2xC|RRm(^so=vaTP@$%5bbKhlM@sR!M==KIqz6*
zfJ*&j%8tsxs$JlnLKJ+ag@g7ktlj?eptZDwDOkv#rkw5~CqbrCV7lDk@fDzuJD*Ou
zgLPLnMn~5A|LCGINy_9RV+vuc!XXq<VJfJw)c4D6gWUZdTetOOGRO}%E!)6EiP}e!
zkw!MV8U(2I8SUaU;1JLM&;igPhWDkzP#O@rH>%?j+E}pWdK0iHAaDxLtsw|e^JE<{
zDh0)DQ4}5huBB9vh0LzG_nw7|I;(BZG<Yb7A5D)BZ}VRI;);!!QYm2qU((t*j2OkK
z;{ll2XReENiEa-637xl)NOkotf`bm$yXL30XAZ(*WT@<nLeZb{JK@s6$+pfm>pxz`
zMt2IdB8%zk0a^8)J-{ZWXNGeO)z*e;#a01<T9g&ncT#>Q5CTNTY$sOB7jAwX4TkGK
zUKQ(_9r1f&o88-H3N63ni@DK#AWA;`m=@7Bs36(bxa(YJ*u>sYY`wddzTytRX}k}^
ztyRcN*9W!=c(dj10FBNMMhxiAGj?wVOG<S7HVqs_!rPn#G{b2cD6oZC<wRSDpV2Ld
z!hTrSVGsEV@s{U7ZCX(pI{LsyYr{R+p-X0%pny<S^-EFdv)Qt*j0YYgzOq!An}N5u
za`tiA8<&IA=mCMV=*36TV93DIBZ<IE)I?ckDyhnQ+~Isgk8``@H2QK7Il&5QDt<}v
zyfEnCUk9tm3Qw=*ebHq7qVS!S{bGaUREWnjZkE6I4BdIHGy3hAmNe;5(B6G2hii&`
zvc6md@09dgaF5QhFMTTw)aXjSJF-FZJa*;i0yli!YSMmeZk#Gklib}|p#9=u*J68E
zM6POFD*`EiJ!IjqTGNN^!gjSMvaQ#AbTFBlal}r}u;n8(QKlLA>r7<vZg}kqzl7j+
zV+n0Bre7`vTVD^jf&E-0N3t^9XI-fA2kD(sU2IZq(IgzE!w0ANQ$1%zTp+C4=bIxd
zvii=heCN|z7U*I1gQ*7se@lV%A#~)C7T!83^p>!P47khr_e*5?hJ&$I3Vj!$D1Se2
zYb^b~ga#1eHVt4|^*HUlBlk9s)JjVVQLRq!m53&xQ?NgcT%T=Emb&-=&ORP9*=TYR
z-e}b}Ts0b<N7s(OWe8U%bWUxZn7yVd1jfWTbuj=yKF7|8y4c7bB$gR{hXx=R_Jj|u
zRGacwnG<e`Oz9g~nOF}NSc3Gbtl~^s|7y}T{%<dLW4_!C2NXBCk)7n`1st=1TC`7a
z`-J)e(Vk0Y=9e1Pu@!P$L4Vv*KbZUv)7=hBW07NAk8%K=*B#=F37*54nHEgk?WbOe
z7I?R84z&GN;BPc0^^~A6;d0+AFk_YDZERu#-m}F!VjUm*Q<C;}W1RvAe5|5!Ge4$n
zO<DF*RSMjvTQ*!0*7*c0Eg>gB`az=H43anyqrTc`Y;eZ0v@5~T0ic)M7`@&_vMnNX
zJ`b`~uejKW{S;rMhM!s+{2~mb>Qvt!Fk{JylK*162TT&3K2M>ux0{=&pX59o+Sae7
zPz#*`6%$1fSkJNzF5d)e9ZQ01XQ77{Rn~QG^q#A*@6P7+n_-;eE6>CL?fo$yLJ%3X
zZ34fr^!BW7pt4>rABC`MpzVN?x4JBqKk)apc$<)@Q%Hn@3kkFkvdBC&_@JqQ3k#P`
zG*Ys$L-*`%>$U)g@)rAzUKNCZi;uR=y6h_=$(ewLmUH<Pmlk8pDa&11rq(Z?9N(pG
z7AOha%Dv{4x9VBauQ;s=k(8vhl#edO4;HU)8M`j*PZHwa_-ZQ3>Q)QKEi8P8f<rvC
z%Lw~JK)4DMY1OwxteL>p3&yv`S$@%y#{?0;p^Y`N{vIx$@E=9%y?liA2Vnj2YaX0?
zNSDV?ng@S!ce2YbdE}h_nkTc?mgQ$_me$Rf_Ok^>5o`qsalRn9fEy-{eRX0ZJ2U#G
z5N~-#{Qb*o(}cH<z0AI(_rWc}ejj?%W}BCye#b!q+$Ik99_!8EW+DVmGmXM?u|1WS
zaKPt+UJcTmk(i9&+2Y6>2Y!TP=I!B}+w8)e{UWR=BmGXW17rgj*;F5!7$m2Wy%DNQ
z9a2*6^FxeF?#!;L%O?67Dfczktsr%rpiJKSFzk9+PUQTjek7ssh1NSqsern|>dGhR
zR@?ge;90v&Luu;yG)v>w7EHP(d4^5R<9Za5c&0`Pc0h2Fv%RZiUK+H{JWtjiCJ)oR
z?qTz1VbUL($Jtk_`mqzOU7<D#9`in<)%;*4Gs|hwa}_u}8DqE=1qqH8D!V`L!o>$^
zS;@ZBg5r3JkG}umIj|GRN#92IGGT=4s=!{3hMqhgz+j00OlxPXxH*6mRysWCr-g*5
z0%Q|LyD|xQdi_dhxiAx$UclyGFHtN=jw<XU>;94tN&{;nmZn{L>H{H~p`Ofyz!^uu
z>Ey#AP(N;MbI(<B_{K8D`Xq^v(Z9pIwnZy7?5(VD4}F!YE*XI&=WW}t|9T7r)?-Xf
z6#YAA*x@t?>i;4<e!#?+Ytnh;tXq4uk&OwA5s;1L^{-0$tmPw0W;RlbfKZ-A^e7&#
zQ}$NY0zvLrWC5120I<kX!=}-LIW?W%9DDAtA!UHf1gj^IU4bH3-=UxUwgCovlnq!M
zJv+fjC6#Ov=9J(VQS@ujBA+#hQJe&cvOi)+@DmT@z5mN5nwdqv_H-6ZX{?PU)_`jo
z>&sWnO+Q1j>T>_Lyhq^WufWgnU6~x3+jYcqMUraUA04GOg%BJDx|L2e+6kmrqw(0m
z$an3+FmxUatvj*QFcpjnBF7owH~Az=i;$8Ksmlus&;$g-xsk9X(+4l3cp_o!-pxK@
z>3{}zG3a3kEuKDlb#XbiM?<Z-LEA-hC9nj4Wf-zzaq!`M<QoC`19a3%9I+q!l;!4`
zK3%E=Y(9br@fz16!;5g4o`_}gcRmFhyP~yplG^+x*C3$NqkLExltvb#qifm3UKbsE
zEIN_I_(t)tntdD`7-(N)w_pk7xHCA|@>|&*4sX~HlBlLPh!KaNiKyNhZt($3uWSZd
zh8KN6_CX+o+?=L&<hb@UR>JwVaQJrQ!8idmtD1w}gBYsK{TM+cZX=LaDl45V1$?OL
zYr_s|U#a2gGO*D*<`bV&3{d;{B-G(zEXu&-IJ@`y4{2#&5e2^+=pm$o<bMJOYzr0l
z40sff&G29nTM#!Mc%HOG*?&EPECcE@Gz11~ntmX8rJWIJZX}rmtExJ4Lkq=i>nhsm
z?j&c)r(Cf44r!!^I;?BNuI#Bea`Rp8x28UFd`A|f#|Sl7hleNweRvx20JH`Td*fr`
z1vM)+p8o|x)!MyUj){yw?b9?gc_ZKODm4eJyc=o4hYFFt!iJw5{6jJPQa2EU^k(PS
zE5Vm6l?#^;V5I_A{nhryO%yk*{vU9QKsgQh?zmnBNZ*FHsHsu%lcMT2Fw9|XKu3tu
zd~wV19~qy#R3@h#f3L`5&NR0mlgOFDk9j~bd9$2Upcr_^$F3H9Ls+~YQlaaiQ1T7<
z5?+kjmnKuO2$OIj$yH)Z{5pn^E%#_z9@ZeJZG%GiwS?>>=6vNjUjwdvibI|jZM1%_
z-}nz8-$#B%?b`>;0C=az&f|HcjB_`C&3?os-e^g*K@A7}`cW4PpF=F4tr`-CGK_`k
zh3_oyKb7cv@6hI9)@hbVGBln+5pTP3DARK3QBaaFv9u<_&q5oJ$wKKia6LpEuoa~*
z2_gPD7t7Oy<SM8Gl`f(54$1ffUW815k((5g3cwc!ac%yQ5lGGn`G2>(Fcn6&+t?{d
z1VR}&#ybDr87i_`2>K&ka!&;Jw3Li|jO^-YBrY;rk(^>qwg!E;$YBzxl~tmVNexju
zUqM{U(XEC3tKT8(ae2L$rBGzq^Y|JFx*RHfhx~NmMPe{&I{>VrWU^WW5buRiW=xwM
z&9L%tFfBY9LKOqzThaU}VT-N{_{Zt$2<Lx}+TI^Dq+Sdx2f(BhFa@v^NxKE2bbd~3
z2O~8OXAI8`cuk{ggclQ_MNYt8fhq}vmSy(hO&xk}=G+G=olbrzNPJQsPp`X+Rzx-t
z)&Qt&tS_BO9FW87fMt5Fyy*qWwwXeFJnFt@j@l1Z6YvrRs33r9!(Wfu>Ls4Brr6xy
z0q=iqsIthao<uek=E%PKoJm_~%q&aPYlierlM!eJ)v5a(YZdU=Oc!Z``v2_@ZeU^m
gryro$p|9i)q1ZRtGfjg9o(IiCuNi1%Y2JJGKkH0dHUIzs

literal 0
HcmV?d00001

diff --git a/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_prompt_getter.yml b/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_prompt_getter.yml
index 43f5b4c35ce..e1590800e6a 100644
--- a/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_prompt_getter.yml
+++ b/tests/e2e/cli/visual_prompting/reference/Zero_Shot_SAM_Tiny_ViT/compressed_prompt_getter.yml
@@ -1,3 +1,3 @@
 TestToolsZeroShotVisualPrompting:
   ptq:
-    number_of_fakequantizers: 1
+    number_of_fakequantizers: 0
diff --git a/tests/e2e/cli/visual_prompting/test_zero_shot.py b/tests/e2e/cli/visual_prompting/test_zero_shot.py
index e5471cc7bb7..c0fb4fdb5d3 100644
--- a/tests/e2e/cli/visual_prompting/test_zero_shot.py
+++ b/tests/e2e/cli/visual_prompting/test_zero_shot.py
@@ -23,10 +23,10 @@
 )
 
 args = {
-    "--train-data-roots": "tests/assets/car_tree_bug",
-    "--val-data-roots": "tests/assets/car_tree_bug",
-    "--test-data-roots": "tests/assets/car_tree_bug",
-    "--input": "tests/assets/car_tree_bug/images/train",
+    "--train-data-roots": "tests/assets/car_tree_bug_zero_shot",
+    "--val-data-roots": "tests/assets/car_tree_bug_zero_shot",
+    "--test-data-roots": "tests/assets/car_tree_bug_zero_shot",
+    "--input": "tests/assets/car_tree_bug_zero_shot/images/train",
     "train_params": [
         "params",
         "--learning_parameters.trainer.max_epochs",
@@ -100,7 +100,7 @@ def test_otx_eval_openvino(self, template, tmp_dir_path, half_precision):
             tmp_dir_path,
             otx_dir,
             args,
-            threshold=0.2,
+            threshold=0.3,
             half_precision=half_precision,
             is_visual_prompting=True,
         )

From dc059ab572bedfbdedf72bcd81db58eec2d86f4a Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Mon, 19 Feb 2024 17:25:28 +0900
Subject: [PATCH 14/28] precommit

---
 .../model_wrappers/openvino_models.py         |  1 -
 .../visual_prompting/tasks/openvino.py        | 41 ++++++++++---------
 2 files changed, 22 insertions(+), 20 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 7c0e8da2b8e..4400e6bf1bd 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -23,7 +23,6 @@
 from openvino.model_api.models.types import NumericalValue, StringValue
 
 from otx.algorithms.visual_prompting.adapters.pytorch_lightning.datasets.pipelines import ResizeLongestSide
-from otx.api.utils.segmentation_utils import create_hard_prediction_from_soft_prediction
 
 
 class ImageEncoder(ImageModel):
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index b7aaa85bc51..29342157aa3 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -310,13 +310,17 @@ def infer(
         self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray
     ) -> Tuple[List[Any], DefaultDict[Any, Any], DefaultDict[Any, Any]]:
         """Perform a prediction for a given input image."""
+        points_score: np.ndarray
+
         # forward image encoder
         images, meta = self.pre_process_image_encoder(images)
         original_size = np.asarray([meta["original_shape"][:2]], dtype=np.int64)
         image_embeddings = self.forward_image_encoder(images)
 
         # get point candidates
-        total_points_scores, total_bg_coords = self.forward_prompt_getter(image_embeddings, reference_feats, used_indices, original_size)
+        total_points_scores, total_bg_coords = self.forward_prompt_getter(
+            image_embeddings, reference_feats, used_indices, original_size
+        )
 
         annotations: DefaultDict = defaultdict(list)
         predicted_masks: DefaultDict = defaultdict(list)
@@ -325,8 +329,9 @@ def infer(
             points_scores = total_points_scores[label]
             bg_coords = total_bg_coords[label]
             for points_score in points_scores:
-                if points_score[-1] in [-1., 0.]:
+                if points_score[-1] in [-1.0, 0.0]:
                     continue
+
                 x, y = points_score[:2]
                 is_done = False
                 for pm in predicted_masks.get(label, []):
@@ -354,7 +359,7 @@ def infer(
                 used_points[label].append(points_score)
 
         self._inspect_overlapping_areas(predicted_masks, used_points)
-        
+
         for label, predictions in predicted_masks.items():
             if len(predictions) == 0:
                 continue
@@ -364,13 +369,10 @@ def infer(
             }
             for prediction, used_point in zip(predictions, used_points[label]):
                 annotation, _, _ = self.post_process(
-                {
-                    self.model["decoder"].output_blob_name: prediction,
-                    "scores": used_point[-1]
-                },
-                metadata)
+                    {self.model["decoder"].output_blob_name: prediction, "scores": used_point[-1]}, metadata
+                )
                 annotations[label].extend(annotation)
-            
+
         return sum(annotations.values(), []), predicted_masks, used_points
 
     def forward_prompt_getter(
@@ -379,13 +381,13 @@ def forward_prompt_getter(
         reference_feats: np.ndarray,
         used_indices: np.ndarray,
         original_size: np.ndarray,
-    ) -> Dict[str, np.ndarray]:
+    ) -> Tuple[Dict[int, np.ndarray], Dict[int, np.ndarray]]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
         inputs = {
             "original_size": original_size,
             "threshold": np.array([[self.model["prompt_getter"].sim_threshold]], dtype=np.float32),
             "num_bg_points": np.array([[self.model["prompt_getter"].num_bg_points]], dtype=np.int64),
-            **image_embeddings
+            **image_embeddings,
         }
         total_points_scores: Dict[int, np.ndarray] = {}
         total_bg_coords: Dict[int, np.ndarray] = {}
@@ -393,10 +395,10 @@ def forward_prompt_getter(
             reference_feat = reference_feats[label]
             inputs["reference_feat"] = reference_feat
             outputs = self.model["prompt_getter"].infer_sync(inputs)
-            
+
             total_points_scores[label] = outputs["points_scores"]
             total_bg_coords[label] = outputs["bg_coords"]
-        
+
         return total_points_scores, total_bg_coords
 
     def forward_decoder(  # type: ignore
@@ -642,19 +644,20 @@ def __getitem__(self, index: int) -> Dict[str, Any]:
             image_embeddings = self.image_encoder(images["images"])
             if self.module_name == "prompt_getter":
                 return {
-                    "reference_feat": self.reference_feats[self.used_indices[0][0]], # only use the first feature
+                    "reference_feat": self.reference_feats[self.used_indices[0][0]],  # only use the first feature
                     "original_size": original_size,
                     "threshold": np.array([[self.inferencer.model["prompt_getter"].sim_threshold]], dtype=np.float32),
                     "num_bg_points": np.array([[self.inferencer.model["prompt_getter"].num_bg_points]], dtype=np.int64),
-                    **image_embeddings
+                    **image_embeddings,
                 }
 
             total_points_scores, total_bg_coords = self.inferencer.forward_prompt_getter(
-                image_embeddings, self.reference_feats, self.used_indices, original_size)
-            
+                image_embeddings, self.reference_feats, self.used_indices, original_size
+            )
+
             # only use the first prompt
-            point_score = total_points_scores[0][0]
-            bg_coords = total_bg_coords[0]
+            point_score: np.ndarray = total_points_scores[0][0]
+            bg_coords: np.ndarray = total_bg_coords[0]
 
             x, y = point_score[:2]
             point_coords = np.concatenate((np.array([[x, y]]), bg_coords), axis=0, dtype=np.float32)

From a6e460d4f39f1bf23e0d1cc61d6f7f2d0c52c52a Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Mon, 19 Feb 2024 20:05:33 +0900
Subject: [PATCH 15/28] Fix unit tests

---
 .../zero_shot_segment_anything.py             |  2 +-
 .../visual_prompting/tasks/openvino.py        |  3 -
 .../model_wrappers/test_openvino_models.py    |  6 +-
 .../test_zero_shot_segment_anything.py        | 76 +++++++++----------
 .../visual_prompting/tasks/test_inference.py  |  5 +-
 .../visual_prompting/tasks/test_openvino.py   | 25 +++---
 .../visual_prompting/test_helpers.py          |  2 +-
 7 files changed, 56 insertions(+), 63 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index aa3206eec89..99bce5ff412 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -128,7 +128,7 @@ def _point_selection(
 
         # to handle empty tensor
         len_fg_coords_scores = len(fg_coords_scores)
-        fg_coords_scores = F.pad(fg_coords_scores, (0, 0, 0, max(0, 10 - len_fg_coords_scores)), value=-1)
+        fg_coords_scores = F.pad(fg_coords_scores, (0, 0, 0, max(0, 1 - len_fg_coords_scores)), value=-1)
 
         ratio = self.image_size / original_size.max()
         width = (original_size[1] * ratio).to(torch.int64)
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 29342157aa3..2611d488a43 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -481,7 +481,6 @@ def _inspect_overlapping_areas(
         self,
         predicted_masks: Dict[int, List[np.ndarray]],
         used_points: Dict[int, List[np.ndarray]],
-        # annotations: Dict[int, List[np.ndarray]],
         threshold_iou: float = 0.8,
     ):
         def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
@@ -511,12 +510,10 @@ def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
             for im in sorted(list(set(overlapped_label)), reverse=True):
                 masks.pop(im)
                 used_points[label].pop(im)
-                # annotations[label].pop(im)
 
             for jm in sorted(list(set(overlapped_other_label)), reverse=True):
                 other_masks.pop(jm)
                 used_points[other_label].pop(jm)
-                # annotations[other_label].pop(jm)
 
     def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
         """Perform a prediction for a given input image."""
diff --git a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
index d000c818c9f..1db453d6e11 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
@@ -176,10 +176,10 @@ def test_get_inputs(self):
     @e2e_pytest_unit
     def test_postprocess(self, mocker):
         """Test postprocess."""
-        self.decoder.output_blob_name = "masks"
-        self.decoder.soft_threshold = 0.5
+        self.decoder.output_blob_name = "upscaled_masks"
+        self.decoder.mask_threshold = 0.
         self.decoder.blur_strength = 2
-        fake_output = {"masks": np.ones((4, 4)), "scores": 0.1}
+        fake_output = {"upscaled_masks": np.ones((4, 4)), "scores": 0.1}
         fake_metadata = {"original_size": np.array([[6, 6]]), "label": mocker.Mock(spec=LabelEntity)}
         returned_value = self.decoder.postprocess(outputs=fake_output, meta=fake_metadata)
 
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 047dcb4f8bf..3202f3b6caf 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -29,71 +29,71 @@
 
 
 class TestPromptGetter:
-    @pytest.fixture(autouse=True)
-    def setup(self) -> None:
-        self.prompt_getter = PromptGetter(image_size=3, downsizing=1)
+    @pytest.fixture
+    def prompt_getter(self) -> PromptGetter:
+        return PromptGetter(image_size=4, downsizing=1)
 
     @e2e_pytest_unit
-    def test_set_default_thresholds(self) -> None:
+    def test_set_default_thresholds(self, prompt_getter) -> None:
         """Test set_default_thresholds."""
-        assert self.prompt_getter.default_threshold_reference == 0.3
-        assert self.prompt_getter.default_threshold_target == 0.65
+        assert prompt_getter.default_threshold_reference == 0.3
+        assert prompt_getter.default_threshold_target == 0.65
 
-        self.prompt_getter.set_default_thresholds(default_threshold_reference=0.5, default_threshold_target=0.7)
+        prompt_getter.set_default_thresholds(default_threshold_reference=0.5, default_threshold_target=0.7)
 
-        assert self.prompt_getter.default_threshold_reference == 0.5
-        assert self.prompt_getter.default_threshold_target == 0.7
+        assert prompt_getter.default_threshold_reference == 0.5
+        assert prompt_getter.default_threshold_target == 0.7
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "result_point_selection",
         [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])],
     )
-    def test_forward(self, mocker, result_point_selection: torch.Tensor) -> None:
+    def test_forward(self, mocker, prompt_getter, result_point_selection: torch.Tensor) -> None:
         """Test forward."""
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
+        )
         mocker.patch.object(
-            self.prompt_getter, "get_prompt_candidates", return_value=(result_point_selection, torch.zeros(1, 2))
+            prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2))
         )
         image_embeddings = torch.ones(1, 4, 4, 4)
-        reference_feats = torch.rand(1, 1, 4)
-        used_indices = torch.as_tensor([[0]])
-        original_size = torch.tensor((self.prompt_getter.image_size, self.prompt_getter.image_size), dtype=torch.int64)
-
-        total_points_scores, total_bg_coords = self.prompt_getter(
-            image_embeddings=image_embeddings,
-            reference_feats=reference_feats,
-            used_indices=used_indices,
-            original_size=original_size,
+        reference_feat = torch.rand(1, 4)
+        original_size = torch.tensor(
+            [[prompt_getter.image_size, prompt_getter.image_size]], dtype=torch.int64
         )
 
-        assert total_points_scores.shape[0] == 1
-        assert total_bg_coords.shape[0] == 1
+        points_scores, bg_coords = prompt_getter(
+            image_embeddings=image_embeddings, reference_feat=reference_feat, original_size=original_size
+        )
 
+        assert torch.all(points_scores == result_point_selection)
+        assert torch.all(bg_coords == torch.zeros(1, 2))
+        
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "result_point_selection",
         [torch.tensor([[2, 2, 0.9], [1, 2, 0.8], [0, 2, 0.7], [2, 1, 0.6]]), torch.tensor([[-1, -1, -1]])],
     )
-    def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tensor) -> None:
+    def test_get_prompt_candidates(self, mocker, prompt_getter, result_point_selection: torch.Tensor) -> None:
         """Test get_prompt_candidates."""
-        mocker.patch(
-            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
-        )
         mocker.patch.object(
-            self.prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2))
+            prompt_getter, "get_prompt_candidates", return_value=(result_point_selection, torch.zeros(1, 2))
         )
         image_embeddings = torch.ones(1, 4, 4, 4)
-        reference_feat = torch.rand(1, 4)
-        original_size = torch.tensor(
-            [[self.prompt_getter.image_size, self.prompt_getter.image_size]], dtype=torch.int64
-        )
+        reference_feats = torch.rand(1, 1, 4)
+        used_indices = torch.as_tensor([[0]])
+        original_size = torch.tensor((prompt_getter.image_size, prompt_getter.image_size), dtype=torch.int64)
 
-        points_scores, bg_coords = self.prompt_getter.get_prompt_candidates(
-            image_embeddings=image_embeddings, reference_feat=reference_feat, original_size=original_size
+        total_points_scores, total_bg_coords = prompt_getter.get_prompt_candidates(
+            image_embeddings=image_embeddings,
+            reference_feats=reference_feats,
+            used_indices=used_indices,
+            original_size=original_size,
         )
 
-        assert torch.all(points_scores == result_point_selection)
-        assert torch.all(bg_coords == torch.zeros(1, 2))
+        assert total_points_scores.shape[0] == len(result_point_selection)
+        assert total_bg_coords.shape[0] == 1
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
@@ -106,11 +106,11 @@ def test_get_prompt_candidates(self, mocker, result_point_selection: torch.Tenso
             (torch.zeros(3, 3), torch.tensor([[-1, -1, -1]])),
         ],
     )
-    def test_point_selection(self, mask_sim: torch.Tensor, expected: torch.Tensor) -> None:
+    def test_point_selection(self, prompt_getter, mask_sim: torch.Tensor, expected: torch.Tensor) -> None:
         """Test _point_selection."""
-        points_scores, bg_coords = self.prompt_getter._point_selection(
+        points_scores, bg_coords = prompt_getter._point_selection(
             mask_sim=mask_sim,
-            original_size=torch.tensor([self.prompt_getter.image_size, self.prompt_getter.image_size]),
+            original_size=torch.tensor([prompt_getter.image_size, prompt_getter.image_size]),
             threshold=torch.tensor([[0.5]]),
         )
 
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
index d595227fa7c..8b330131b0f 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_inference.py
@@ -343,8 +343,7 @@ def test_export_to_onnx(self):
             },
             "visual_prompting_prompt_getter": {
                 "image_embeddings": np.random.randn(1, embed_dim, *embed_size).astype(dtype=np.float32),
-                "reference_feats": np.random.randn(2, 1, 256).astype(dtype=np.float32),
-                "used_indices": np.array([[0, 1]], dtype=np.int64),
+                "reference_feat": np.random.randn(1, 256).astype(dtype=np.float32),
                 "original_size": np.random.randint(low=0, high=image_size * 2, size=(1, 2), dtype=np.int64),
                 "threshold": np.array([[0.1]], dtype=np.float32),
                 "num_bg_points": np.random.randint(low=1, high=image_size, size=(1, 1), dtype=np.int64),
@@ -360,7 +359,7 @@ def test_export_to_onnx(self):
         }
         onnx_outputs = {
             "visual_prompting_image_encoder": ["image_embeddings"],
-            "visual_prompting_prompt_getter": ["total_points_scores", "total_bg_coords"],
+            "visual_prompting_prompt_getter": ["points_scores", "bg_coords"],
             "visual_prompting_decoder": ["upscaled_masks", "iou_predictions", "low_res_masks"],
         }
 
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index cdfdab44739..717e4a7d703 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -204,6 +204,7 @@ def setup(self, mocker):
         )
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.3
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"]._apply_coords.return_value = np.array([[1, 1]])
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].output_blob_name = "upscaled_masks"
 
     @e2e_pytest_unit
     def test_predict(self, mocker):
@@ -221,10 +222,10 @@ def test_predict(self, mocker):
         mocker_forward_decoder = mocker.patch.object(
             OpenVINOZeroShotVisualPromptingInferencer,
             "forward_prompt_getter",
-            return_value={"total_points_scores": np.array([[[1, 1, 1]]]), "total_bg_coords": np.array([[[2, 2]]])},
+            return_value=({0: np.array([[1, 1, 1]])}, {0: np.array([[2, 2]])}),
         )
         mocker_forward_decoder = mocker.patch.object(
-            OpenVINOZeroShotVisualPromptingInferencer, "forward_decoder", return_value={}
+            OpenVINOZeroShotVisualPromptingInferencer, "forward_decoder", return_value={"upscaled_masks": None}
         )
         mocker_post_process = mocker.patch.object(
             OpenVINOZeroShotVisualPromptingInferencer, "post_process", return_value=(self.fake_annotation, None, None)
@@ -421,11 +422,11 @@ def test_inspect_overlapping_areas(self) -> None:
         }
 
         self.zero_shot_visual_prompting_ov_inferencer._inspect_overlapping_areas(
-            predicted_masks, used_points, predicted_masks.copy(), threshold_iou=0.5
+            predicted_masks, used_points, threshold_iou=0.5
         )
 
-        assert len(predicted_masks[0]) == 1
-        assert len(predicted_masks[1]) == 2
+        assert len(predicted_masks[0]) == 2
+        assert len(predicted_masks[1]) == 3
         assert all(np.array([2, 2, 0.5]) == used_points[0][0])
         assert all(np.array([0, 0, 0.7]) == used_points[1][2])
 
@@ -484,9 +485,8 @@ def _load_dataloader(module_name: str, output_model: Optional[ModelEntity] = Non
             dataset = generate_visual_prompting_dataset()
             dataset = dataset.get_subset(Subset.TRAINING)
             return OTXZeroShotOpenVinoDataLoader(
-                dataset, self.mocker_inferencer, module_name, output_model=output_model
+                dataset, self.mocker_inferencer, module_name, output_model=output_model, reference_feats=np.zeros((1, 1, 1)), used_indices=np.array([[0]])
             )
-
         return _load_dataloader
 
     @pytest.fixture(autouse=True)
@@ -510,17 +510,14 @@ def test_getitem(self, mocker, load_dataloader, module_name: str):
         setattr(dataloader, "target_length", 8)
         mocker.patch.object(
             dataloader.inferencer,
-            "pre_process",
+            "pre_process_image_encoder",
             return_value=({"images": np.zeros((1, 3, 4, 4), dtype=np.uint8)}, {"original_shape": (4, 4)}),
         )
         if module_name == "decoder":
             mocker.patch.object(
-                dataloader,
-                "prompt_getter",
-                return_value={
-                    "total_points_scores": [np.array([[0, 0, 0.5]])],
-                    "total_bg_coords": [np.array([[1, 1]])],
-                },
+                dataloader.inferencer,
+                "forward_prompt_getter",
+                return_value=({0: np.array([[0, 0, 0.5]])}, {0: np.array([[1, 1]])}),
             )
 
         results = dataloader.__getitem__(0)
diff --git a/tests/unit/algorithms/visual_prompting/test_helpers.py b/tests/unit/algorithms/visual_prompting/test_helpers.py
index 5775c7712de..a1b077cf557 100644
--- a/tests/unit/algorithms/visual_prompting/test_helpers.py
+++ b/tests/unit/algorithms/visual_prompting/test_helpers.py
@@ -202,7 +202,7 @@ def set_default_thresholds(self, *args, **kwargs):
         pass
 
     def get_prompt_candidates(self, *args, **kwargs):
-        return {1: (torch.Tensor([[0, 0, 0.5]]), torch.Tensor([[1, 1]]))}
+        return {1: torch.Tensor([[0, 0, 0.5]])}, {1: torch.Tensor([[1, 1]])}
 
     def forward(self, *args, **kwargs):
         return torch.tensor([[[0, 0, 0.5], [1, 1, 0.7]]]), torch.tensor([[[2, 2]]])

From 387edff2f0a2d7abf228e566b81ad601c03c4a71 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Mon, 19 Feb 2024 20:34:28 +0900
Subject: [PATCH 16/28] Revert postprocessing

---
 .../models/visual_prompters/segment_anything.py                 | 2 --
 1 file changed, 2 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index 14535b04cb2..f50afe06c8f 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -351,8 +351,6 @@ def postprocess_masks(cls, masks: Tensor, input_size: int, orig_size: Tensor) ->
 
     def get_prepadded_size(self, input_image_size: Tensor, longest_side: int) -> Tensor:
         """Get pre-padded size."""
-        input_image_size = input_image_size.to(torch.float32)
-        longest_side = torch.tensor(longest_side).to(torch.float32)
         scale = longest_side / torch.max(input_image_size)
         transformed_size = scale * input_image_size
         return torch.floor(transformed_size + 0.5).to(torch.int64)

From 182255d14163d9c60cadfc153c0a1bb3688428da Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Mon, 19 Feb 2024 20:37:25 +0900
Subject: [PATCH 17/28] precommit

---
 .../openvino/model_wrappers/test_openvino_models.py    |  2 +-
 .../test_zero_shot_segment_anything.py                 | 10 +++-------
 .../algorithms/visual_prompting/tasks/test_openvino.py |  8 +++++++-
 3 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
index 1db453d6e11..1b70f12aa33 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/openvino/model_wrappers/test_openvino_models.py
@@ -177,7 +177,7 @@ def test_get_inputs(self):
     def test_postprocess(self, mocker):
         """Test postprocess."""
         self.decoder.output_blob_name = "upscaled_masks"
-        self.decoder.mask_threshold = 0.
+        self.decoder.mask_threshold = 0.0
         self.decoder.blur_strength = 2
         fake_output = {"upscaled_masks": np.ones((4, 4)), "scores": 0.1}
         fake_metadata = {"original_size": np.array([[6, 6]]), "label": mocker.Mock(spec=LabelEntity)}
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index 3202f3b6caf..a7fb4ae5d49 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -54,14 +54,10 @@ def test_forward(self, mocker, prompt_getter, result_point_selection: torch.Tens
         mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.ZeroShotSegmentAnything"
         )
-        mocker.patch.object(
-            prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2))
-        )
+        mocker.patch.object(prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2)))
         image_embeddings = torch.ones(1, 4, 4, 4)
         reference_feat = torch.rand(1, 4)
-        original_size = torch.tensor(
-            [[prompt_getter.image_size, prompt_getter.image_size]], dtype=torch.int64
-        )
+        original_size = torch.tensor([[prompt_getter.image_size, prompt_getter.image_size]], dtype=torch.int64)
 
         points_scores, bg_coords = prompt_getter(
             image_embeddings=image_embeddings, reference_feat=reference_feat, original_size=original_size
@@ -69,7 +65,7 @@ def test_forward(self, mocker, prompt_getter, result_point_selection: torch.Tens
 
         assert torch.all(points_scores == result_point_selection)
         assert torch.all(bg_coords == torch.zeros(1, 2))
-        
+
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "result_point_selection",
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 717e4a7d703..78850916564 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -485,8 +485,14 @@ def _load_dataloader(module_name: str, output_model: Optional[ModelEntity] = Non
             dataset = generate_visual_prompting_dataset()
             dataset = dataset.get_subset(Subset.TRAINING)
             return OTXZeroShotOpenVinoDataLoader(
-                dataset, self.mocker_inferencer, module_name, output_model=output_model, reference_feats=np.zeros((1, 1, 1)), used_indices=np.array([[0]])
+                dataset,
+                self.mocker_inferencer,
+                module_name,
+                output_model=output_model,
+                reference_feats=np.zeros((1, 1, 1)),
+                used_indices=np.array([[0]]),
             )
+
         return _load_dataloader
 
     @pytest.fixture(autouse=True)

From e17179265eba392c12a887c067ff780e2f7f3141 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 20 Feb 2024 15:39:41 +0900
Subject: [PATCH 18/28] Update unit tests

---
 .../model_wrappers/openvino_models.py         |   1 +
 .../zero_shot_segment_anything.py             |   9 +-
 .../configs/base/configuration.py             |   9 +
 .../visual_prompting/tasks/openvino.py        | 221 ++++++++++++++++--
 .../visual_prompting/tasks/test_openvino.py   | 135 ++++++++++-
 .../visual_prompting/test_helpers.py          |   5 +-
 6 files changed, 354 insertions(+), 26 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 4400e6bf1bd..22dee61a7c9 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -71,6 +71,7 @@ def parameters(cls) -> Dict[str, Any]:  # noqa: D102
         parameters.update({"image_size": NumericalValue(value_type=int, default_value=1024, min=0, max=2048)})
         parameters.update({"sim_threshold": NumericalValue(value_type=float, default_value=0.5, min=0, max=1)})
         parameters.update({"num_bg_points": NumericalValue(value_type=int, default_value=1, min=0, max=1024)})
+        parameters.update({"default_threshold_reference": NumericalValue(value_type=float, default_value=0.3, min=-1., max=1.)})
         return parameters
 
     def _get_inputs(self):
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 99bce5ff412..37223101ef2 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -8,7 +8,7 @@
 import pickle
 from collections import OrderedDict, defaultdict
 from copy import deepcopy
-from datetime import datetime
+import time
 from itertools import product
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Union
 
@@ -505,7 +505,8 @@ def _predict_masks(
         masks: Tensor
         logits: Tensor
         scores: Tensor
-        for i in range(3):
+        num_iter = 3 if is_cascade else 1
+        for i in range(num_iter):
             if i == 0:
                 # First-step prediction
                 mask_input = torch.zeros(1, 1, *map(lambda x: x * 4, image_embeddings.shape[2:]), device=self.device)
@@ -747,7 +748,7 @@ def training_epoch_end(self, outputs) -> None:
             self.reference_info["used_indices"].unique().unsqueeze(0), requires_grad=False
         )
         if self.config.model.save_outputs:
-            path_reference_info = self.path_reference_info.format(datetime.now().strftime("%Y%m%d-%H%M%S"))
+            path_reference_info = self.path_reference_info.format(time.strftime("%Y%m%d-%H%M%S"))
             os.makedirs(os.path.dirname(path_reference_info), exist_ok=True)
             torch.save(self.reference_info, path_reference_info)
             pickle.dump(
@@ -758,4 +759,4 @@ def training_epoch_end(self, outputs) -> None:
                 repr(self.trainer.datamodule.train_dataset.dataset),
                 open(path_reference_info.replace("reference_info.pt", "reference_meta.json"), "w"),
             )
-            logger.info(f"Saved reference info at {path_reference_info}")
+            logger.info(f"Saved reference info at {path_reference_info}.")
diff --git a/src/otx/algorithms/visual_prompting/configs/base/configuration.py b/src/otx/algorithms/visual_prompting/configs/base/configuration.py
index d7383c28c69..61388696a09 100644
--- a/src/otx/algorithms/visual_prompting/configs/base/configuration.py
+++ b/src/otx/algorithms/visual_prompting/configs/base/configuration.py
@@ -131,6 +131,15 @@ class __Postprocessing(ParameterGroup):
             max_value=1024,
             affects_outcome_of=ModelLifecycle.INFERENCE,
         )
+        
+        default_threshold_reference = configurable_float(
+            default_value=0.3,
+            header="Default reference threshold",
+            description="The threshold to get target area in the mask for reference features.",
+            min_value=-1.0,
+            max_value=1.0,
+            affects_outcome_of=ModelLifecycle.INFERENCE,
+        )
 
     @attrs
     class __POTParameter(BaseConfig.BasePOTParameter):
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 2611d488a43..988043f1600 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -14,8 +14,10 @@
 # See the License for the specific language governing permissions
 # and limitations under the License.
 
+import cv2
 import io
 import json
+from copy import deepcopy
 import os
 import pickle
 import random
@@ -26,6 +28,7 @@
 from pathlib import Path
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Type, Union
 from zipfile import ZipFile
+from otx.api.entities.scored_label import ScoredLabel
 
 import attr
 import nncf
@@ -255,7 +258,7 @@ def __init__(
                 **attr.asdict(
                     hparams.postprocessing,
                     filter=lambda attr, value: attr.name
-                    in ["image_size", "sim_threshold", "num_bg_points", "embedded_processing"],
+                    in ["image_size", "sim_threshold", "num_bg_points", "embedded_processing", "default_threshold_reference"],
                 )
             },
             "decoder": {
@@ -294,8 +297,8 @@ def __init__(
         self.point_labels_box = np.array([[2, 3]], dtype=np.float32)
         self.has_mask_inputs = [np.array([[0.0]]), np.array([[1.0]])]
 
-        self.reference_feats = None
-        self.used_indices = None
+        self.reference_feats: np.ndarray = None
+        self.used_indices: np.ndarray = None
 
     def pre_process_image_encoder(
         self, inputs: np.ndarray, extra_processing: bool = False
@@ -303,11 +306,73 @@ def pre_process_image_encoder(
         """Pre-process function of OpenVINO Zero-shot Visual Prompting Inferencer for image encoder."""
         return self.model["image_encoder"].preprocess(inputs, extra_processing)
 
-    def learn(self, images: np.ndarray):
-        """Learn."""
+    def learn(
+        self,
+        dataset_item: DatasetItemEntity,
+        reset_feat: bool = False,
+        use_bbox: bool = False,
+        use_point: bool = False,
+        path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle",
+    ) -> Tuple[Dict[int, np.ndarray], np.ndarray]:
+        """Learn for reference features."""
+        ref_masks: np.ndarray
+        self.reference_feats: np.ndarray
+        self.used_indices: np.ndarray
+
+        if reset_feat or self.reference_feats is None:
+            self.initialize_reference_info()
+            
+        images, meta, prompts = self.pre_process(dataset_item, use_bbox, use_point)
+        largest_label: int = max([int(p["label"].id) for p in prompts])
+        self.expand_reference_info(largest_label)
+        
+        image_embeddings = self.forward_image_encoder(images)
+        processed_embedding = image_embeddings["image_embeddings"].squeeze().transpose(1, 2, 0)
+        original_size = meta["original_shape"][:2]
+        
+        ref_masks = np.zeros((largest_label + 1, *map(int, original_size)), dtype=np.uint8)
+        for prompt in prompts:
+            if "point_coords" in prompt:
+                # bboxes and points
+                label = prompt.pop("label")
+                original_size = prompt.get("orig_size")
+                prompt.update(image_embeddings)
+                
+                prediction = self.forward_decoder(prompt, original_size, is_cascade=False)
+                ref_mask = prediction["upscaled_masks"]
+            else:
+                logger.warning("annotation and polygon will be supported.")
+                continue
+            ref_masks[int(label.id)] += ref_mask
+            
+        ref_masks = np.clip(ref_masks, 0, 1)
+        for label in range(largest_label+1):
+            ref_mask = ref_masks[label]
+            if ref_mask.sum() == 0:
+                # empty prediction
+                continue
+
+            ref_feat = None
+            default_threshold_reference = deepcopy(self.model["prompt_getter"].default_threshold_reference)
+            while ref_feat is None:
+                logger.info(f"[*] default_threshold_reference : {default_threshold_reference:.4f}")
+                ref_feat = self._generate_masked_features(
+                    processed_embedding, ref_masks[label], default_threshold_reference
+                )
+                default_threshold_reference -= 0.05
+
+            self.reference_feats[label] = ref_feat
+            self.used_indices = np.concatenate((self.used_indices, np.array([[label]])), axis=1)
+        
+        reference_info = {"reference_feats": self.reference_feats, "used_indices": self.used_indices}
+        path_reference_info = path_reference_info.format(time.strftime("%Y%m%d-%H%M%S"))
+        logger.info(f"Saved reference info at {path_reference_info}.")
+        pickle.dump(reference_info, open(path_reference_info, "wb"))
+        return reference_info, ref_masks
+        
 
     def infer(
-        self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray
+        self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray, is_cascade: bool = False,
     ) -> Tuple[List[Any], DefaultDict[Any, Any], DefaultDict[Any, Any]]:
         """Perform a prediction for a given input image."""
         points_score: np.ndarray
@@ -352,7 +417,7 @@ def infer(
                 }
                 inputs_decoder.update(image_embeddings)
 
-                prediction = self.forward_decoder(inputs_decoder, original_size)
+                prediction = self.forward_decoder(inputs_decoder, original_size, is_cascade)
                 prediction.update({"scores": points_score[-1]})
 
                 predicted_masks[label].append(prediction[self.model["decoder"].output_blob_name])
@@ -402,13 +467,14 @@ def forward_prompt_getter(
         return total_points_scores, total_bg_coords
 
     def forward_decoder(  # type: ignore
-        self, inputs: Dict[str, np.ndarray], original_size: np.ndarray
+        self, inputs: Dict[str, np.ndarray], original_size: np.ndarray, is_cascade: bool = True,
     ) -> Dict[str, np.ndarray]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
         masks: np.ndarray
         logits: np.ndarray
         scores: np.ndarray
-        for i in range(3):
+        num_iter = 3 if is_cascade else 1
+        for i in range(num_iter):
             if i == 0:
                 # First-step prediction
                 mask_input = np.zeros(
@@ -416,7 +482,7 @@ def forward_decoder(  # type: ignore
                 )
                 has_mask_input = self.has_mask_inputs[0]
 
-            elif i == 1:
+            elif is_cascade and i == 1:
                 # Cascaded Post-refinement-1
                 mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)  # noqa: F821
                 if masks.sum() == 0:
@@ -424,7 +490,7 @@ def forward_decoder(  # type: ignore
 
                 has_mask_input = self.has_mask_inputs[1]
 
-            elif i == 2:
+            elif is_cascade and i == 2:
                 # Cascaded Post-refinement-2
                 mask_input, masks = self._postprocess_masks(masks, logits, scores)  # noqa: F821
                 if masks.sum() == 0:
@@ -517,27 +583,86 @@ def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray):
 
     def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type: ignore
         """Perform a prediction for a given input image."""
-        if self.reference_feats is None and self.used_indices is None:
-            self.reference_feats, self.used_indices = self._get_reference_info()
         results = self.infer(dataset_item.numpy, self.reference_feats, self.used_indices)
         return results[0]
 
-    def _find_latest_reference_info(self) -> Union[str, None]:
+    def _find_latest_reference_info(self, root: str = "vpm_zsl_reference_infos") -> Union[str, None]:
         """Find latest reference info to be used."""
-        if len(stamps := sorted(os.listdir("vpm_zsl_reference_infos"), reverse=True)) > 0:
+        if len(stamps := sorted(os.listdir(root), reverse=True)) > 0:
             return stamps[0]
         return None
 
     def _get_reference_info(
-        self, path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle"
+        self, root: str = "vpm_zsl_reference_infos", path_reference_info: str = "{}/reference_info.pickle"
     ) -> Union[Tuple[np.ndarray, np.ndarray], None]:
         """Get reference info through loading previously saved one or running `learn`."""
-        if (latest_stamp := self._find_latest_reference_info()) is not None:
+        if (latest_stamp := self._find_latest_reference_info(root)) is not None:
             # load previously saved reference info
-            latest_reference_info = path_reference_info.format(latest_stamp)
+            latest_reference_info = os.path.join(root, path_reference_info.format(latest_stamp))
             reference_info = pickle.load(open(latest_reference_info, "rb"))
             return reference_info["reference_feats"], reference_info["used_indices"]
         return None, None
+    
+    def initialize_reference_info(self) -> None:
+        """Initialize reference information."""
+        self.reference_feats: np.ndarray = np.zeros((0, 1, 256), dtype=np.float32)
+        self.used_indices: np.ndarray = np.array([[]], dtype=np.int64)
+
+    def expand_reference_info(self, new_largest_label: int) -> None:
+        """Expand reference info dimensions if newly given processed prompts have more lables."""
+        if new_largest_label > (cur_largest_label := len(self.reference_feats) - 1):
+            diff = new_largest_label - cur_largest_label
+            self.reference_feats = np.pad(self.reference_feats, ((0, diff), (0, 0), (0, 0)), constant_values=0.0)
+            
+    def _generate_masked_features(
+        self,
+        feats: np.ndarray,
+        masks: np.ndarray,
+        threshold_mask: float,
+    ) -> Tuple[np.ndarray, ...]:
+        """Generate masked features.
+
+        Args:
+            feats (np.ndarray): Raw reference features. It will be filtered with masks.
+            masks (np.ndarray): Reference masks used to filter features.
+            threshold_mask (float): Threshold to control masked region.
+
+        Returns:
+            (np.ndarray): Masked features.
+        """
+        target_shape = self.model["image_encoder"].image_size / max(masks.shape) * np.array(masks.shape)
+        target_shape = target_shape[::-1].astype(np.int32)
+
+        # Post-process masks
+        masks = cv2.resize(masks, target_shape, interpolation=cv2.INTER_LINEAR)
+        masks = self._pad_to_square(masks)
+        masks = cv2.resize(masks, feats.shape[:2][::-1], interpolation=cv2.INTER_LINEAR)
+
+        # Target feature extraction
+        if (masks > threshold_mask).sum() == 0:
+            # (for stability) there is no area to be extracted
+            return None
+
+        masked_feat = feats[masks > threshold_mask]
+        masked_feat = masked_feat.mean(0)[None]
+        masked_feat = masked_feat / np.linalg.norm(masked_feat, axis=-1, keepdims=True)
+
+        return masked_feat
+    
+    def _pad_to_square(self, x: np.ndarray) -> np.ndarray:
+        """Pad to a square input.
+
+        Args:
+            x (np.ndarray): Mask to be padded.
+
+        Returns:
+            (np.ndarray): Padded mask.
+        """
+        h, w = x.shape[-2:]
+        padh = self.model["image_encoder"].image_size - h
+        padw = self.model["image_encoder"].image_size - w
+        x = np.pad(x, ((0, padh), (0, padw)), constant_values=0.)
+        return x
 
 
 class OTXOpenVinoDataLoader:
@@ -943,6 +1068,66 @@ def load_inferencer(self) -> OpenVINOZeroShotVisualPromptingInferencer:
             },
             num_requests=get_default_async_reqs_num(),
         )
+        
+    def infer(
+        self,
+        dataset: DatasetEntity,
+        inference_parameters: Optional[InferenceParameters] = None,
+        path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle"
+    ) -> DatasetEntity:
+        """Infer function of OpenVINOVisualPromptingTask.
+
+        Currently, asynchronous execution is not supported, synchronous execution will be executed instead.
+        """
+        if inference_parameters is not None:
+            update_progress_callback = inference_parameters.update_progress
+            enable_async_inference = inference_parameters.enable_async_inference
+        else:
+            update_progress_callback = default_progress_callback
+            enable_async_inference = True
+
+        # FIXME (sungchul): Support async inference.
+        if enable_async_inference:
+            logger.warning("Asynchronous inference doesn't work, synchronous inference will be executed.")
+            enable_async_inference = False
+        predicted_validation_dataset = dataset.with_empty_annotations()
+
+        def add_prediction(id: int, annotations: List[Annotation]):
+            dataset_item = predicted_validation_dataset[id]
+            dataset_item.append_annotations(annotations)
+
+        total_time = 0.0
+        dataset_size = len(dataset)
+        
+        if self.inferencer.reference_feats is None and self.inferencer.used_indices is None:
+            # set reference_feats and used_indices from previously saved reference_info
+            self.inferencer.reference_feats, self.inferencer.used_indices = self.inferencer._get_reference_info(path_reference_info)
+            if self.inferencer.reference_feats is None and self.inferencer.used_indices is None:
+                # if they are empty, stop inference and return empty dataset
+                logger.warning((
+                    "reference_feats and used_indices are empty, stop inference and return empty dataset. "
+                    "Please run learn function first."
+                ))
+                return predicted_validation_dataset
+
+        for i, dataset_item in enumerate(dataset, 1):
+            start_time = time.perf_counter()
+
+            annotations = self.inferencer.predict(dataset_item)
+            add_prediction(i - 1, annotations)
+
+            end_time = time.perf_counter() - start_time
+            total_time += end_time
+            update_progress_callback(int(i / dataset_size * 100), None)
+
+        self.inferencer.await_all()
+
+        self._avg_time_per_image = total_time / len(dataset)
+        logger.info(f"Avg time per image: {self._avg_time_per_image} secs")
+        logger.info(f"Total time: {total_time} secs")
+        logger.info("Visual Prompting OpenVINO inference completed")
+
+        return predicted_validation_dataset
 
     def optimize(
         self,
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 78850916564..d548679d2c8 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -49,6 +49,7 @@
     generate_visual_prompting_dataset,
     init_environment,
 )
+from tests.unit.algorithms.visual_prompting.test_helpers import MockScoredLabel
 
 
 class TestOpenVINOVisualPromptingInferencer:
@@ -200,11 +201,54 @@ def setup(self, mocker):
             {"image_encoder": "", "prompt_getter": "", "decoder": ""},
         )
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"] = mocker.patch(
-            "otx.algorithms.visual_prompting.tasks.openvino.model_wrappers.Decoder", autospec=True
+            "otx.algorithms.visual_prompting.tasks.openvino.model_wrappers.Decoder", autospec=True,
         )
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.3
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"]._apply_coords.return_value = np.array([[1, 1]])
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].output_blob_name = "upscaled_masks"
+        
+    @e2e_pytest_unit
+    def test_learn(self, mocker):
+        """Test learn."""
+        mocker_pre_process = mocker.patch.object(
+            OpenVINOVisualPromptingInferencer,
+            "pre_process",
+            return_value=(
+                torch.zeros((1, 3, 2, 2)),
+                {"original_shape": np.array((4, 4))},
+                [
+                    {
+                        "point_coords": [np.array([[[1, 1], [2, 2]]])],
+                        "point_labels": [1, 2],
+                        "label": MockScoredLabel(label=0, name="fake"),
+                        "orig_size": (4, 4),
+                    }
+                ],
+            ),
+        )
+        mocker_forward_image_encoder = mocker.patch.object(
+            OpenVINOZeroShotVisualPromptingInferencer,
+            "forward_image_encoder",
+            return_value={"image_embeddings": np.empty((4, 2, 2))},
+        )
+        mocker_generate_masked_features = mocker.patch.object(OpenVINOZeroShotVisualPromptingInferencer, "_generate_masked_features", return_value=torch.ones(1, 256))
+
+        self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].infer_sync.return_value = {
+            "upscaled_masks": np.ones((1, 4, 4, 4), dtype=np.bool), "iou_predictions": np.array([[0.9, 0.7, 0.9, 0.8]]), "low_res_masks": np.ones((1, 4, 2, 2)),}        
+        mocker_pickle_dump = mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.pickle.dump")
+        mocker.patch("builtins.open", return_value="Mocked data")
+        self.zero_shot_visual_prompting_ov_inferencer.model["prompt_getter"].default_threshold_reference = 0.3
+        
+        fake_input = mocker.Mock(spec=DatasetItemEntity)
+        results = self.zero_shot_visual_prompting_ov_inferencer.learn(fake_input, reset_feat=True)
+
+        assert results[0]["reference_feats"].shape == (1, 1, 256)
+        assert results[0]["used_indices"] == np.array([[0]])
+        assert np.all(results[1] == np.ones((1, 4, 4)))
+        mocker_pre_process.assert_called_once()
+        mocker_forward_image_encoder.assert_called_once()
+        mocker_generate_masked_features.assert_called_once()
+        mocker_pickle_dump.assert_called_once()
 
     @e2e_pytest_unit
     def test_predict(self, mocker):
@@ -230,16 +274,33 @@ def test_predict(self, mocker):
         mocker_post_process = mocker.patch.object(
             OpenVINOZeroShotVisualPromptingInferencer, "post_process", return_value=(self.fake_annotation, None, None)
         )
+        self.zero_shot_visual_prompting_ov_inferencer.reference_feats = np.random.rand(1, 1, 1)
+        self.zero_shot_visual_prompting_ov_inferencer.used_indices = np.array([[0]])
         fake_input = mocker.Mock(spec=DatasetItemEntity)
 
-        returned_value = self.zero_shot_visual_prompting_ov_inferencer.predict(fake_input)
+        results = self.zero_shot_visual_prompting_ov_inferencer.predict(fake_input)
 
         mocker_pre_process.assert_called_once()
         mocker_forward.assert_called_once()
         mocker_forward_decoder.assert_called_once()
         mocker_post_process.assert_called_once()
-        assert returned_value == self.fake_annotation
+        assert results == self.fake_annotation
 
+    @e2e_pytest_unit
+    def test_forward_prompt_getter(self):
+        """Test forward_prompt_getter."""
+        self.zero_shot_visual_prompting_ov_inferencer.model["prompt_getter"].infer_sync.return_value = {
+            "points_scores": np.array([[1, 1, 0.5]]), "bg_coords": np.array([[0, 0]])}
+        
+        total_points_scores, total_bg_coords = self.zero_shot_visual_prompting_ov_inferencer.forward_prompt_getter(
+            image_embeddings={"image_embeddings": np.empty((4, 2, 2))},
+            reference_feats=np.random.rand(1, 1, 1),
+            used_indices=np.array([[0]]),
+            original_size=np.array([4, 4]))
+        
+        assert np.all(total_points_scores[0] == np.array([[1, 1, 0.5]]))
+        assert np.all(total_bg_coords[0] == np.array([[0, 0]]))
+    
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "postprocess_output,infer_sync_output,expected",
@@ -429,6 +490,74 @@ def test_inspect_overlapping_areas(self) -> None:
         assert len(predicted_masks[1]) == 3
         assert all(np.array([2, 2, 0.5]) == used_points[0][0])
         assert all(np.array([0, 0, 0.7]) == used_points[1][2])
+        
+    @e2e_pytest_unit
+    def test_find_latest_reference_info(self, mocker):
+        """Test _find_latest_reference_info."""
+        # there are some saved reference info
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=["1", "2"])
+        results = self.zero_shot_visual_prompting_ov_inferencer._find_latest_reference_info()
+        assert results == "2"
+        
+        # there are no saved reference info
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=[])
+        results = self.zero_shot_visual_prompting_ov_inferencer._find_latest_reference_info()
+        assert results is None
+        
+    @e2e_pytest_unit
+    def test_get_reference_info(self, mocker):
+        """Test _get_reference_info."""
+        # get previously saved reference info
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=["1", "2"])
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.pickle.load", return_value={"reference_feats": 1, "used_indices": 2})
+        mocker.patch("builtins.open", return_value="Mocked data")
+        
+        results = self.zero_shot_visual_prompting_ov_inferencer._get_reference_info()
+        assert results == (1, 2)
+        
+        # no saved reference info
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=[])
+        
+        results = self.zero_shot_visual_prompting_ov_inferencer._get_reference_info()
+        assert results == (None, None)
+        
+    @e2e_pytest_unit
+    def test_expand_reference_info(self):
+        """Test expand_reference_info."""
+        self.zero_shot_visual_prompting_ov_inferencer.reference_feats = np.ones((3, 2, 2))
+        new_largest_label = 5
+
+        self.zero_shot_visual_prompting_ov_inferencer.expand_reference_info(new_largest_label)
+        
+        assert self.zero_shot_visual_prompting_ov_inferencer.reference_feats.shape == (6, 2, 2)
+        assert np.all(self.zero_shot_visual_prompting_ov_inferencer.reference_feats[:3] == 1.0)
+        assert np.all(self.zero_shot_visual_prompting_ov_inferencer.reference_feats[3:] == 0.0)
+        
+    @e2e_pytest_unit
+    def test_generate_masked_features(self) -> None:
+        """Test _generate_masked_features."""
+        self.zero_shot_visual_prompting_ov_inferencer.model["image_encoder"].image_size = 16
+        feats = np.random.rand(8, 8, 1)
+        masks = np.zeros((16, 16), dtype=np.float32)
+        masks[4:12, 4:12] = 1.0
+
+        masked_feat = self.zero_shot_visual_prompting_ov_inferencer._generate_masked_features(
+            feats=feats, masks=masks, threshold_mask=0.3)
+
+        assert masked_feat.shape == (1, 1)
+        
+    @e2e_pytest_unit
+    def test_pad_to_square(self) -> None:
+        """Test _pad_to_square."""
+        self.zero_shot_visual_prompting_ov_inferencer.model["image_encoder"].image_size = 16
+
+        result = self.zero_shot_visual_prompting_ov_inferencer._pad_to_square(x=np.ones((8, 8)))
+
+        assert result[:8, :8].sum() == 8**2
+        assert result[:8, 8:].sum() == 0
+        assert result[8:, :8].sum() == 0
+        assert result[8:, 8:].sum() == 0
+        
 
 
 class TestOTXOpenVinoDataLoader:
diff --git a/tests/unit/algorithms/visual_prompting/test_helpers.py b/tests/unit/algorithms/visual_prompting/test_helpers.py
index a1b077cf557..5d79549c670 100644
--- a/tests/unit/algorithms/visual_prompting/test_helpers.py
+++ b/tests/unit/algorithms/visual_prompting/test_helpers.py
@@ -184,10 +184,13 @@ def predict_mask(self, *args, **kwargs):
 
 
 class MockScoredLabel:
-    def __init__(self, label: int, name: str = "background"):
+    def __init__(self, label: int, name: str = "background", probability: float = 0., label_source = None,):
         self.name = name
         self.label = Mock()
         self.label.id_ = label
+        self.label.id = label
+        self.probability = probability
+        self.label_source = label_source
         self.__class__ = ScoredLabel
 
 

From 43bc6c2569c7f60a528e6eca1750221bd271ece5 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 20 Feb 2024 15:42:17 +0900
Subject: [PATCH 19/28] precommit

---
 .../model_wrappers/openvino_models.py         |  4 +-
 .../zero_shot_segment_anything.py             |  2 +-
 .../configs/base/configuration.py             |  2 +-
 .../visual_prompting/tasks/openvino.py        | 79 +++++++++++--------
 .../visual_prompting/tasks/test_openvino.py   | 58 ++++++++------
 .../visual_prompting/test_helpers.py          |  8 +-
 6 files changed, 93 insertions(+), 60 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
index 22dee61a7c9..3283026252b 100644
--- a/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
+++ b/src/otx/algorithms/visual_prompting/adapters/openvino/model_wrappers/openvino_models.py
@@ -71,7 +71,9 @@ def parameters(cls) -> Dict[str, Any]:  # noqa: D102
         parameters.update({"image_size": NumericalValue(value_type=int, default_value=1024, min=0, max=2048)})
         parameters.update({"sim_threshold": NumericalValue(value_type=float, default_value=0.5, min=0, max=1)})
         parameters.update({"num_bg_points": NumericalValue(value_type=int, default_value=1, min=0, max=1024)})
-        parameters.update({"default_threshold_reference": NumericalValue(value_type=float, default_value=0.3, min=-1., max=1.)})
+        parameters.update(
+            {"default_threshold_reference": NumericalValue(value_type=float, default_value=0.3, min=-1.0, max=1.0)}
+        )
         return parameters
 
     def _get_inputs(self):
diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index 37223101ef2..b54f1180b49 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -6,9 +6,9 @@
 import json
 import os
 import pickle
+import time
 from collections import OrderedDict, defaultdict
 from copy import deepcopy
-import time
 from itertools import product
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Union
 
diff --git a/src/otx/algorithms/visual_prompting/configs/base/configuration.py b/src/otx/algorithms/visual_prompting/configs/base/configuration.py
index 61388696a09..2961fa8a837 100644
--- a/src/otx/algorithms/visual_prompting/configs/base/configuration.py
+++ b/src/otx/algorithms/visual_prompting/configs/base/configuration.py
@@ -131,7 +131,7 @@ class __Postprocessing(ParameterGroup):
             max_value=1024,
             affects_outcome_of=ModelLifecycle.INFERENCE,
         )
-        
+
         default_threshold_reference = configurable_float(
             default_value=0.3,
             header="Default reference threshold",
diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index 988043f1600..bbffbe9ce22 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -14,23 +14,22 @@
 # See the License for the specific language governing permissions
 # and limitations under the License.
 
-import cv2
 import io
 import json
-from copy import deepcopy
 import os
 import pickle
 import random
 import tempfile
 import time
 from collections import defaultdict
+from copy import deepcopy
 from itertools import product
 from pathlib import Path
 from typing import Any, DefaultDict, Dict, List, Optional, Tuple, Type, Union
 from zipfile import ZipFile
-from otx.api.entities.scored_label import ScoredLabel
 
 import attr
+import cv2
 import nncf
 import numpy as np
 import openvino.runtime as ov
@@ -258,7 +257,13 @@ def __init__(
                 **attr.asdict(
                     hparams.postprocessing,
                     filter=lambda attr, value: attr.name
-                    in ["image_size", "sim_threshold", "num_bg_points", "embedded_processing", "default_threshold_reference"],
+                    in [
+                        "image_size",
+                        "sim_threshold",
+                        "num_bg_points",
+                        "embedded_processing",
+                        "default_threshold_reference",
+                    ],
                 )
             },
             "decoder": {
@@ -297,8 +302,8 @@ def __init__(
         self.point_labels_box = np.array([[2, 3]], dtype=np.float32)
         self.has_mask_inputs = [np.array([[0.0]]), np.array([[1.0]])]
 
-        self.reference_feats: np.ndarray = None
-        self.used_indices: np.ndarray = None
+        self.reference_feats: Optional[np.ndarray] = None
+        self.used_indices: Optional[np.ndarray] = None
 
     def pre_process_image_encoder(
         self, inputs: np.ndarray, extra_processing: bool = False
@@ -313,23 +318,21 @@ def learn(
         use_bbox: bool = False,
         use_point: bool = False,
         path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle",
-    ) -> Tuple[Dict[int, np.ndarray], np.ndarray]:
+    ) -> Tuple[Dict[str, np.ndarray], np.ndarray]:
         """Learn for reference features."""
         ref_masks: np.ndarray
-        self.reference_feats: np.ndarray
-        self.used_indices: np.ndarray
 
         if reset_feat or self.reference_feats is None:
             self.initialize_reference_info()
-            
+
         images, meta, prompts = self.pre_process(dataset_item, use_bbox, use_point)
         largest_label: int = max([int(p["label"].id) for p in prompts])
         self.expand_reference_info(largest_label)
-        
+
         image_embeddings = self.forward_image_encoder(images)
         processed_embedding = image_embeddings["image_embeddings"].squeeze().transpose(1, 2, 0)
         original_size = meta["original_shape"][:2]
-        
+
         ref_masks = np.zeros((largest_label + 1, *map(int, original_size)), dtype=np.uint8)
         for prompt in prompts:
             if "point_coords" in prompt:
@@ -337,16 +340,16 @@ def learn(
                 label = prompt.pop("label")
                 original_size = prompt.get("orig_size")
                 prompt.update(image_embeddings)
-                
+
                 prediction = self.forward_decoder(prompt, original_size, is_cascade=False)
                 ref_mask = prediction["upscaled_masks"]
             else:
                 logger.warning("annotation and polygon will be supported.")
                 continue
             ref_masks[int(label.id)] += ref_mask
-            
+
         ref_masks = np.clip(ref_masks, 0, 1)
-        for label in range(largest_label+1):
+        for label in range(largest_label + 1):
             ref_mask = ref_masks[label]
             if ref_mask.sum() == 0:
                 # empty prediction
@@ -363,16 +366,19 @@ def learn(
 
             self.reference_feats[label] = ref_feat
             self.used_indices = np.concatenate((self.used_indices, np.array([[label]])), axis=1)
-        
+
         reference_info = {"reference_feats": self.reference_feats, "used_indices": self.used_indices}
         path_reference_info = path_reference_info.format(time.strftime("%Y%m%d-%H%M%S"))
         logger.info(f"Saved reference info at {path_reference_info}.")
         pickle.dump(reference_info, open(path_reference_info, "wb"))
         return reference_info, ref_masks
-        
 
     def infer(
-        self, images: np.ndarray, reference_feats: np.ndarray, used_indices: np.ndarray, is_cascade: bool = False,
+        self,
+        images: np.ndarray,
+        reference_feats: np.ndarray,
+        used_indices: np.ndarray,
+        is_cascade: bool = False,
     ) -> Tuple[List[Any], DefaultDict[Any, Any], DefaultDict[Any, Any]]:
         """Perform a prediction for a given input image."""
         points_score: np.ndarray
@@ -467,7 +473,10 @@ def forward_prompt_getter(
         return total_points_scores, total_bg_coords
 
     def forward_decoder(  # type: ignore
-        self, inputs: Dict[str, np.ndarray], original_size: np.ndarray, is_cascade: bool = True,
+        self,
+        inputs: Dict[str, np.ndarray],
+        original_size: np.ndarray,
+        is_cascade: bool = True,
     ) -> Dict[str, np.ndarray]:
         """Forward function of OpenVINO Visual Prompting Inferencer."""
         masks: np.ndarray
@@ -602,18 +611,18 @@ def _get_reference_info(
             reference_info = pickle.load(open(latest_reference_info, "rb"))
             return reference_info["reference_feats"], reference_info["used_indices"]
         return None, None
-    
+
     def initialize_reference_info(self) -> None:
         """Initialize reference information."""
-        self.reference_feats: np.ndarray = np.zeros((0, 1, 256), dtype=np.float32)
-        self.used_indices: np.ndarray = np.array([[]], dtype=np.int64)
+        self.reference_feats = np.zeros((0, 1, 256), dtype=np.float32)
+        self.used_indices = np.array([[]], dtype=np.int64)
 
     def expand_reference_info(self, new_largest_label: int) -> None:
         """Expand reference info dimensions if newly given processed prompts have more lables."""
         if new_largest_label > (cur_largest_label := len(self.reference_feats) - 1):
             diff = new_largest_label - cur_largest_label
             self.reference_feats = np.pad(self.reference_feats, ((0, diff), (0, 0), (0, 0)), constant_values=0.0)
-            
+
     def _generate_masked_features(
         self,
         feats: np.ndarray,
@@ -648,7 +657,7 @@ def _generate_masked_features(
         masked_feat = masked_feat / np.linalg.norm(masked_feat, axis=-1, keepdims=True)
 
         return masked_feat
-    
+
     def _pad_to_square(self, x: np.ndarray) -> np.ndarray:
         """Pad to a square input.
 
@@ -661,7 +670,7 @@ def _pad_to_square(self, x: np.ndarray) -> np.ndarray:
         h, w = x.shape[-2:]
         padh = self.model["image_encoder"].image_size - h
         padw = self.model["image_encoder"].image_size - w
-        x = np.pad(x, ((0, padh), (0, padw)), constant_values=0.)
+        x = np.pad(x, ((0, padh), (0, padw)), constant_values=0.0)
         return x
 
 
@@ -1068,12 +1077,12 @@ def load_inferencer(self) -> OpenVINOZeroShotVisualPromptingInferencer:
             },
             num_requests=get_default_async_reqs_num(),
         )
-        
+
     def infer(
         self,
         dataset: DatasetEntity,
         inference_parameters: Optional[InferenceParameters] = None,
-        path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle"
+        path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle",
     ) -> DatasetEntity:
         """Infer function of OpenVINOVisualPromptingTask.
 
@@ -1098,16 +1107,20 @@ def add_prediction(id: int, annotations: List[Annotation]):
 
         total_time = 0.0
         dataset_size = len(dataset)
-        
+
         if self.inferencer.reference_feats is None and self.inferencer.used_indices is None:
             # set reference_feats and used_indices from previously saved reference_info
-            self.inferencer.reference_feats, self.inferencer.used_indices = self.inferencer._get_reference_info(path_reference_info)
+            self.inferencer.reference_feats, self.inferencer.used_indices = self.inferencer._get_reference_info(
+                path_reference_info
+            )
             if self.inferencer.reference_feats is None and self.inferencer.used_indices is None:
                 # if they are empty, stop inference and return empty dataset
-                logger.warning((
-                    "reference_feats and used_indices are empty, stop inference and return empty dataset. "
-                    "Please run learn function first."
-                ))
+                logger.warning(
+                    (
+                        "reference_feats and used_indices are empty, stop inference and return empty dataset. "
+                        "Please run learn function first."
+                    )
+                )
                 return predicted_validation_dataset
 
         for i, dataset_item in enumerate(dataset, 1):
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index d548679d2c8..6aee906b6be 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -201,12 +201,13 @@ def setup(self, mocker):
             {"image_encoder": "", "prompt_getter": "", "decoder": ""},
         )
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"] = mocker.patch(
-            "otx.algorithms.visual_prompting.tasks.openvino.model_wrappers.Decoder", autospec=True,
+            "otx.algorithms.visual_prompting.tasks.openvino.model_wrappers.Decoder",
+            autospec=True,
         )
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].mask_threshold = 0.3
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"]._apply_coords.return_value = np.array([[1, 1]])
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].output_blob_name = "upscaled_masks"
-        
+
     @e2e_pytest_unit
     def test_learn(self, mocker):
         """Test learn."""
@@ -231,14 +232,19 @@ def test_learn(self, mocker):
             "forward_image_encoder",
             return_value={"image_embeddings": np.empty((4, 2, 2))},
         )
-        mocker_generate_masked_features = mocker.patch.object(OpenVINOZeroShotVisualPromptingInferencer, "_generate_masked_features", return_value=torch.ones(1, 256))
+        mocker_generate_masked_features = mocker.patch.object(
+            OpenVINOZeroShotVisualPromptingInferencer, "_generate_masked_features", return_value=torch.ones(1, 256)
+        )
 
         self.zero_shot_visual_prompting_ov_inferencer.model["decoder"].infer_sync.return_value = {
-            "upscaled_masks": np.ones((1, 4, 4, 4), dtype=np.bool), "iou_predictions": np.array([[0.9, 0.7, 0.9, 0.8]]), "low_res_masks": np.ones((1, 4, 2, 2)),}        
+            "upscaled_masks": np.ones((1, 4, 4, 4), dtype=np.bool),
+            "iou_predictions": np.array([[0.9, 0.7, 0.9, 0.8]]),
+            "low_res_masks": np.ones((1, 4, 2, 2)),
+        }
         mocker_pickle_dump = mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.pickle.dump")
         mocker.patch("builtins.open", return_value="Mocked data")
         self.zero_shot_visual_prompting_ov_inferencer.model["prompt_getter"].default_threshold_reference = 0.3
-        
+
         fake_input = mocker.Mock(spec=DatasetItemEntity)
         results = self.zero_shot_visual_prompting_ov_inferencer.learn(fake_input, reset_feat=True)
 
@@ -290,17 +296,20 @@ def test_predict(self, mocker):
     def test_forward_prompt_getter(self):
         """Test forward_prompt_getter."""
         self.zero_shot_visual_prompting_ov_inferencer.model["prompt_getter"].infer_sync.return_value = {
-            "points_scores": np.array([[1, 1, 0.5]]), "bg_coords": np.array([[0, 0]])}
-        
+            "points_scores": np.array([[1, 1, 0.5]]),
+            "bg_coords": np.array([[0, 0]]),
+        }
+
         total_points_scores, total_bg_coords = self.zero_shot_visual_prompting_ov_inferencer.forward_prompt_getter(
             image_embeddings={"image_embeddings": np.empty((4, 2, 2))},
             reference_feats=np.random.rand(1, 1, 1),
             used_indices=np.array([[0]]),
-            original_size=np.array([4, 4]))
-        
+            original_size=np.array([4, 4]),
+        )
+
         assert np.all(total_points_scores[0] == np.array([[1, 1, 0.5]]))
         assert np.all(total_bg_coords[0] == np.array([[0, 0]]))
-    
+
     @e2e_pytest_unit
     @pytest.mark.parametrize(
         "postprocess_output,infer_sync_output,expected",
@@ -490,7 +499,7 @@ def test_inspect_overlapping_areas(self) -> None:
         assert len(predicted_masks[1]) == 3
         assert all(np.array([2, 2, 0.5]) == used_points[0][0])
         assert all(np.array([0, 0, 0.7]) == used_points[1][2])
-        
+
     @e2e_pytest_unit
     def test_find_latest_reference_info(self, mocker):
         """Test _find_latest_reference_info."""
@@ -498,29 +507,32 @@ def test_find_latest_reference_info(self, mocker):
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=["1", "2"])
         results = self.zero_shot_visual_prompting_ov_inferencer._find_latest_reference_info()
         assert results == "2"
-        
+
         # there are no saved reference info
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=[])
         results = self.zero_shot_visual_prompting_ov_inferencer._find_latest_reference_info()
         assert results is None
-        
+
     @e2e_pytest_unit
     def test_get_reference_info(self, mocker):
         """Test _get_reference_info."""
         # get previously saved reference info
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=["1", "2"])
-        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.pickle.load", return_value={"reference_feats": 1, "used_indices": 2})
+        mocker.patch(
+            "otx.algorithms.visual_prompting.tasks.openvino.pickle.load",
+            return_value={"reference_feats": 1, "used_indices": 2},
+        )
         mocker.patch("builtins.open", return_value="Mocked data")
-        
+
         results = self.zero_shot_visual_prompting_ov_inferencer._get_reference_info()
         assert results == (1, 2)
-        
+
         # no saved reference info
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=[])
-        
+
         results = self.zero_shot_visual_prompting_ov_inferencer._get_reference_info()
         assert results == (None, None)
-        
+
     @e2e_pytest_unit
     def test_expand_reference_info(self):
         """Test expand_reference_info."""
@@ -528,11 +540,11 @@ def test_expand_reference_info(self):
         new_largest_label = 5
 
         self.zero_shot_visual_prompting_ov_inferencer.expand_reference_info(new_largest_label)
-        
+
         assert self.zero_shot_visual_prompting_ov_inferencer.reference_feats.shape == (6, 2, 2)
         assert np.all(self.zero_shot_visual_prompting_ov_inferencer.reference_feats[:3] == 1.0)
         assert np.all(self.zero_shot_visual_prompting_ov_inferencer.reference_feats[3:] == 0.0)
-        
+
     @e2e_pytest_unit
     def test_generate_masked_features(self) -> None:
         """Test _generate_masked_features."""
@@ -542,10 +554,11 @@ def test_generate_masked_features(self) -> None:
         masks[4:12, 4:12] = 1.0
 
         masked_feat = self.zero_shot_visual_prompting_ov_inferencer._generate_masked_features(
-            feats=feats, masks=masks, threshold_mask=0.3)
+            feats=feats, masks=masks, threshold_mask=0.3
+        )
 
         assert masked_feat.shape == (1, 1)
-        
+
     @e2e_pytest_unit
     def test_pad_to_square(self) -> None:
         """Test _pad_to_square."""
@@ -557,7 +570,6 @@ def test_pad_to_square(self) -> None:
         assert result[:8, 8:].sum() == 0
         assert result[8:, :8].sum() == 0
         assert result[8:, 8:].sum() == 0
-        
 
 
 class TestOTXOpenVinoDataLoader:
diff --git a/tests/unit/algorithms/visual_prompting/test_helpers.py b/tests/unit/algorithms/visual_prompting/test_helpers.py
index 5d79549c670..445bc0f2ba1 100644
--- a/tests/unit/algorithms/visual_prompting/test_helpers.py
+++ b/tests/unit/algorithms/visual_prompting/test_helpers.py
@@ -184,7 +184,13 @@ def predict_mask(self, *args, **kwargs):
 
 
 class MockScoredLabel:
-    def __init__(self, label: int, name: str = "background", probability: float = 0., label_source = None,):
+    def __init__(
+        self,
+        label: int,
+        name: str = "background",
+        probability: float = 0.0,
+        label_source=None,
+    ):
         self.name = name
         self.label = Mock()
         self.label.id_ = label

From 3966f36f8e66f9f2aae0e200a9ee098f26af8f96 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 20 Feb 2024 15:47:37 +0900
Subject: [PATCH 20/28] Fix

---
 src/otx/algorithms/visual_prompting/tasks/openvino.py     | 7 +++++--
 tests/integration/cli/visual_prompting/test_zero_shot.py  | 8 ++++----
 .../algorithms/visual_prompting/tasks/test_openvino.py    | 4 ++++
 3 files changed, 13 insertions(+), 6 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index bbffbe9ce22..ada64d3c937 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -597,6 +597,8 @@ def predict(self, dataset_item: DatasetItemEntity) -> List[Annotation]:  # type:
 
     def _find_latest_reference_info(self, root: str = "vpm_zsl_reference_infos") -> Union[str, None]:
         """Find latest reference info to be used."""
+        if not os.path.isdir(root):
+            return None
         if len(stamps := sorted(os.listdir(root), reverse=True)) > 0:
             return stamps[0]
         return None
@@ -1082,7 +1084,8 @@ def infer(
         self,
         dataset: DatasetEntity,
         inference_parameters: Optional[InferenceParameters] = None,
-        path_reference_info: str = "vpm_zsl_reference_infos/{}/reference_info.pickle",
+        root: str = "vpm_zsl_reference_infos",
+        path_reference_info: str = "{}/reference_info.pickle",
     ) -> DatasetEntity:
         """Infer function of OpenVINOVisualPromptingTask.
 
@@ -1111,7 +1114,7 @@ def add_prediction(id: int, annotations: List[Annotation]):
         if self.inferencer.reference_feats is None and self.inferencer.used_indices is None:
             # set reference_feats and used_indices from previously saved reference_info
             self.inferencer.reference_feats, self.inferencer.used_indices = self.inferencer._get_reference_info(
-                path_reference_info
+                root, path_reference_info
             )
             if self.inferencer.reference_feats is None and self.inferencer.used_indices is None:
                 # if they are empty, stop inference and return empty dataset
diff --git a/tests/integration/cli/visual_prompting/test_zero_shot.py b/tests/integration/cli/visual_prompting/test_zero_shot.py
index ccedf5c2fa2..33b5b433628 100644
--- a/tests/integration/cli/visual_prompting/test_zero_shot.py
+++ b/tests/integration/cli/visual_prompting/test_zero_shot.py
@@ -17,10 +17,10 @@
 )
 
 args = {
-    "--train-data-roots": "tests/assets/car_tree_bug",
-    "--val-data-roots": "tests/assets/car_tree_bug",
-    "--test-data-roots": "tests/assets/car_tree_bug",
-    "--input": "tests/assets/car_tree_bug/images/train",
+    "--train-data-roots": "tests/assets/car_tree_bug_zero_shot",
+    "--val-data-roots": "tests/assets/car_tree_bug_zero_shot",
+    "--test-data-roots": "tests/assets/car_tree_bug_zero_shot",
+    "--input": "tests/assets/car_tree_bug_zero_shot/images/train",
     "train_params": [
         "params",
         "--learning_parameters.trainer.max_epochs",
diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 6aee906b6be..39c5fcfafd8 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -503,6 +503,8 @@ def test_inspect_overlapping_areas(self) -> None:
     @e2e_pytest_unit
     def test_find_latest_reference_info(self, mocker):
         """Test _find_latest_reference_info."""
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.path.isdir", return_value=True)
+
         # there are some saved reference info
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=["1", "2"])
         results = self.zero_shot_visual_prompting_ov_inferencer._find_latest_reference_info()
@@ -516,6 +518,8 @@ def test_find_latest_reference_info(self, mocker):
     @e2e_pytest_unit
     def test_get_reference_info(self, mocker):
         """Test _get_reference_info."""
+        mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.path.isdir", return_value=True)
+
         # get previously saved reference info
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.os.listdir", return_value=["1", "2"])
         mocker.patch(

From b3448f67471fcff6e21b52de6d5b22756eef89f7 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Tue, 20 Feb 2024 17:36:13 +0900
Subject: [PATCH 21/28] Fix for unit test coverage

---
 .../zero_shot_segment_anything.py             |  7 +-
 .../test_zero_shot_segment_anything.py        | 85 +++++++++++++++++--
 2 files changed, 84 insertions(+), 8 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
index b54f1180b49..5948f2e42bb 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/zero_shot_segment_anything.py
@@ -723,11 +723,12 @@ def configure_optimizers(self) -> None:
         """Skip configure_optimizers unused in zero-shot learning."""
         pass
 
-    def _find_latest_reference_info(self) -> Union[str, None]:
+    def _find_latest_reference_info(self, root: str = "vpm_zsl_reference_infos") -> Union[str, None]:
         """Find latest reference info to be used."""
-        if len(stamps := sorted(os.listdir("vpm_zsl_reference_infos"), reverse=True)) > 0:
+        if not os.path.isdir(root):
+            return None
+        if len(stamps := sorted(os.listdir(root), reverse=True)) > 0:
             return stamps[0]
-        self.initialize_reference_info()
         return None
 
     def on_train_start(self) -> None:
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
index a7fb4ae5d49..5f1812adf86 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_zero_shot_segment_anything.py
@@ -73,9 +73,7 @@ def test_forward(self, mocker, prompt_getter, result_point_selection: torch.Tens
     )
     def test_get_prompt_candidates(self, mocker, prompt_getter, result_point_selection: torch.Tensor) -> None:
         """Test get_prompt_candidates."""
-        mocker.patch.object(
-            prompt_getter, "get_prompt_candidates", return_value=(result_point_selection, torch.zeros(1, 2))
-        )
+        mocker.patch.object(prompt_getter, "_point_selection", return_value=(result_point_selection, torch.zeros(1, 2)))
         image_embeddings = torch.ones(1, 4, 4, 4)
         reference_feats = torch.rand(1, 1, 4)
         used_indices = torch.as_tensor([[0]])
@@ -88,8 +86,8 @@ def test_get_prompt_candidates(self, mocker, prompt_getter, result_point_selecti
             original_size=original_size,
         )
 
-        assert total_points_scores.shape[0] == len(result_point_selection)
-        assert total_bg_coords.shape[0] == 1
+        assert total_points_scores[0].shape[0] == len(result_point_selection)
+        assert total_bg_coords[0].shape[0] == 1
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
@@ -426,3 +424,80 @@ def test_postprocess_masks(
         _, result = zero_shot_segment_anything._postprocess_masks(masks, logits, scores)
 
         assert torch.equal(result, expected)
+
+    @e2e_pytest_unit
+    def test_find_latest_reference_info(self, mocker, set_zero_shot_segment_anything):
+        """Test _find_latest_reference_info."""
+        zero_shot_segment_anything = set_zero_shot_segment_anything()
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.os.path.isdir",
+            return_value=True,
+        )
+
+        # there are some saved reference info
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.os.listdir",
+            return_value=["1", "2"],
+        )
+        results = zero_shot_segment_anything._find_latest_reference_info()
+        assert results == "2"
+
+        # there are no saved reference info
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.os.listdir",
+            return_value=[],
+        )
+        results = zero_shot_segment_anything._find_latest_reference_info()
+        assert results is None
+
+    @e2e_pytest_unit
+    def test_on_predict_start(self, mocker, set_zero_shot_segment_anything):
+        """Test on_predict_start."""
+        zero_shot_segment_anything = set_zero_shot_segment_anything()
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.os.path.isdir",
+            return_value=True,
+        )
+
+        # get previously saved reference info
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.os.listdir",
+            return_value=["1", "2"],
+        )
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.torch.load",
+            return_value=torch.nn.ParameterDict(
+                {"reference_feats": torch.zeros((1, 1, 256)), "used_indices": torch.tensor([[0.0]])}
+            ),
+        )
+        mocker.patch("builtins.open", return_value="Mocked data")
+
+        zero_shot_segment_anything.on_predict_start()
+        assert isinstance(zero_shot_segment_anything.reference_info, torch.nn.ParameterDict)
+        assert zero_shot_segment_anything.reference_info["reference_feats"].shape == (1, 1, 256)
+        assert zero_shot_segment_anything.reference_info["used_indices"].shape == (1, 1)
+
+        # no saved reference info
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.zero_shot_segment_anything.os.listdir",
+            return_value=[],
+        )
+
+        zero_shot_segment_anything.set_empty_reference_info()
+        zero_shot_segment_anything.on_predict_start()
+
+        assert zero_shot_segment_anything.reference_info["reference_feats"].shape == (0,)
+        assert zero_shot_segment_anything.reference_info["used_indices"].shape == (1, 0)
+
+    @e2e_pytest_unit
+    def test_expand_reference_info(self, set_zero_shot_segment_anything):
+        """Test expand_reference_info."""
+        zero_shot_segment_anything = set_zero_shot_segment_anything()
+        zero_shot_segment_anything.reference_info["reference_feats"] = torch.ones((3, 2, 2))
+        new_largest_label = 5
+
+        zero_shot_segment_anything.expand_reference_info(new_largest_label)
+
+        assert zero_shot_segment_anything.reference_info["reference_feats"].shape == (6, 2, 2)
+        assert torch.all(zero_shot_segment_anything.reference_info["reference_feats"][:3] == 1.0)
+        assert torch.all(zero_shot_segment_anything.reference_info["reference_feats"][3:] == 0.0)

From 5c6ffd27cc5088a84a660c52268fcf7471627aa7 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 22 Feb 2024 14:59:37 +0900
Subject: [PATCH 22/28] Refactor model loading and state dict handling

---
 .../visual_prompters/segment_anything.py      | 20 ++---
 .../visual_prompters/test_segment_anything.py | 88 ++++++++++---------
 2 files changed, 50 insertions(+), 58 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
index f50afe06c8f..9327f7f2b84 100644
--- a/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
+++ b/src/otx/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/segment_anything.py
@@ -138,26 +138,13 @@ def load_checkpoint(self, state_dict: Optional[OrderedDict] = None) -> None:
         Args:
             state_dict (Optional[OrderedDict], optional): State dict of SAM. Defaults to None.
         """
-
-        def skip_unused_parameters(state_dict):
-            if self.config.model.backbone == "tiny_vit":
-                for key in [
-                    "image_encoder.norm_head.weight",
-                    "image_encoder.norm_head.bias",
-                    "image_encoder.head.weight",
-                    "image_encoder.head.bias",
-                ]:
-                    if key in state_dict:
-                        state_dict.pop(key)
-
         if state_dict:
             # state_dict from args.load_from
-            skip_unused_parameters(state_dict)
             self.load_state_dict(state_dict)
         elif self.config.model.checkpoint:
             if str(self.config.model.checkpoint).endswith(".ckpt"):
                 # load lightning checkpoint
-                self.load_from_checkpoint(self.config.model.checkpoint)
+                self.load_from_checkpoint(self.config.model.checkpoint, strict=False)
             else:
                 if str(self.config.model.checkpoint).startswith("http"):
                     # get checkpoint from url
@@ -167,8 +154,11 @@ def skip_unused_parameters(state_dict):
                     with open(self.config.model.checkpoint, "rb") as f:
                         state_dict = torch.load(f)
 
-                skip_unused_parameters(state_dict)
                 self.load_state_dict(state_dict, strict=False)
+        else:
+            # use default checkpoint
+            state_dict = torch.hub.load_state_dict_from_url(CKPT_PATHS[self.config.model.backbone])
+            self.load_state_dict(state_dict, strict=False)
 
     ##########################################################
     #     forward for inference (export/deploy/optimize)     #
diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index 8f3f727e533..a1a49b0faea 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -153,37 +153,21 @@ def test_set_metrics(self, mocker, loss_type: str):
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(
-        "is_backbone_arg,state_dict",
+        "state_dict",
         [
-            (
-                False,
-                OrderedDict(
-                    [
-                        ("image_encoder.weight", torch.ones(4, 4)),
-                        ("image_encoder.bias", torch.ones(4)),
-                        ("prompt_encoder.layer.weight", Tensor([[0.0]])),
-                        ("prompt_encoder.layer.bias", Tensor([0.0])),
-                        ("mask_decoder.layer.weight", Tensor([[0.0]])),
-                        ("mask_decoder.layer.bias", Tensor([0.0])),
-                    ]
-                ),
-            ),
-            (
-                True,
-                OrderedDict(
-                    [
-                        ("image_encoder.weight", torch.ones(4, 4)),
-                        ("image_encoder.bias", torch.ones(4)),
-                        ("prompt_encoder.layer.weight", Tensor([[1.0]])),
-                        ("prompt_encoder.layer.bias", Tensor([1.0])),
-                        ("mask_decoder.layer.weight", Tensor([[1.0]])),
-                        ("mask_decoder.layer.bias", Tensor([1.0])),
-                    ]
-                ),
+            OrderedDict(
+                [
+                    ("image_encoder.weight", torch.ones(4, 4)),
+                    ("image_encoder.bias", torch.ones(4)),
+                    ("prompt_encoder.layer.weight", Tensor([[1.0]])),
+                    ("prompt_encoder.layer.bias", Tensor([1.0])),
+                    ("mask_decoder.layer.weight", Tensor([[1.0]])),
+                    ("mask_decoder.layer.bias", Tensor([1.0])),
+                ]
             ),
         ],
     )
-    def test_load_checkpoint_with_state_dict(self, mocker, is_backbone_arg: bool, state_dict: OrderedDict):
+    def test_load_checkpoint_with_state_dict(self, mocker, state_dict: OrderedDict):
         """Test load_checkpoint with state_dict."""
         mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.freeze_networks"
@@ -191,30 +175,18 @@ def test_load_checkpoint_with_state_dict(self, mocker, is_backbone_arg: bool, st
         mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.set_metrics"
         )
+        mocker_load_state_dict = mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_state_dict"
+        )
 
         sam = SegmentAnything(self.base_config, state_dict=state_dict)
         sam_state_dict = sam.state_dict()
 
+        mocker_load_state_dict.assert_called_once()
         for k, v in state_dict.items():
             assert k in sam_state_dict
             assert torch.all(v == sam_state_dict[k])
 
-    @e2e_pytest_unit
-    def test_load_checkpoint_without_checkpoint(self, mocker):
-        """Test load_checkpoint without checkpoint."""
-        mocker.patch(
-            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.freeze_networks"
-        )
-        mocker.patch(
-            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.set_metrics"
-        )
-        config = self.base_config.copy()
-        config.model.update(dict(checkpoint=None))
-
-        sam = SegmentAnything(config, state_dict=None)
-
-        assert True
-
     @e2e_pytest_unit
     def test_load_checkpoint_with_url(self, mocker):
         """Test load_checkpoint with url."""
@@ -228,12 +200,16 @@ def test_load_checkpoint_with_url(self, mocker):
         mocker_load_state_dict = mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_state_dict"
         )
+        mocker_load_from_checkpoint = mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_from_checkpoint"
+        )
 
         config = self.base_config.copy()
         config.model.update(dict(checkpoint="http://checkpoint"))
 
         sam = SegmentAnything(config, state_dict=None)
 
+        mocker_load_from_checkpoint.assert_not_called()
         mocker_load_state_dict_from_url.assert_called_once()
         mocker_load_state_dict.assert_called_once()
 
@@ -263,8 +239,34 @@ def test_load_checkpoint_from_local_checkpoint(self, mocker, monkeypatch, checkp
 
         if checkpoint.endswith(".ckpt"):
             mocker_load_from_checkpoint.assert_called_once()
+            mocker_load_state_dict.assert_not_called()
         else:
+            mocker_load_from_checkpoint.assert_not_called()
             mocker_load_state_dict.assert_called_once()
+            
+    @e2e_pytest_unit
+    def test_load_checkpoint_without_checkpoint(self, mocker):
+        """Test load_checkpoint without checkpoint."""
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.freeze_networks"
+        )
+        mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.set_metrics"
+        )
+        mocker_load_from_checkpoint = mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_from_checkpoint"
+        )
+        mocker_load_state_dict = mocker.patch(
+            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_state_dict"
+        )
+        mocker_load_state_dict_from_url = mocker.patch("torch.hub.load_state_dict_from_url", return_value=OrderedDict())
+
+        config = self.base_config.copy()
+        sam = SegmentAnything(config, state_dict=None)
+
+        mocker_load_from_checkpoint.assert_not_called()
+        mocker_load_state_dict_from_url.assert_called_once()
+        mocker_load_state_dict.assert_called_once()
 
     @e2e_pytest_unit
     @pytest.mark.parametrize(

From 539a37cbd3ec058d70e86704ba159d7aa9f53d26 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 22 Feb 2024 15:03:23 +0900
Subject: [PATCH 23/28] Remove unnecessary flag

---
 src/otx/algorithms/visual_prompting/tasks/openvino.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/otx/algorithms/visual_prompting/tasks/openvino.py b/src/otx/algorithms/visual_prompting/tasks/openvino.py
index ada64d3c937..6f50ef97940 100644
--- a/src/otx/algorithms/visual_prompting/tasks/openvino.py
+++ b/src/otx/algorithms/visual_prompting/tasks/openvino.py
@@ -491,7 +491,7 @@ def forward_decoder(  # type: ignore
                 )
                 has_mask_input = self.has_mask_inputs[0]
 
-            elif is_cascade and i == 1:
+            elif i == 1:
                 # Cascaded Post-refinement-1
                 mask_input, masks = self._postprocess_masks(masks, logits, scores, is_single=True)  # noqa: F821
                 if masks.sum() == 0:
@@ -499,7 +499,7 @@ def forward_decoder(  # type: ignore
 
                 has_mask_input = self.has_mask_inputs[1]
 
-            elif is_cascade and i == 2:
+            elif i == 2:
                 # Cascaded Post-refinement-2
                 mask_input, masks = self._postprocess_masks(masks, logits, scores)  # noqa: F821
                 if masks.sum() == 0:

From a01345a3c3d24f5ea9563dc49ccba568d550fdad Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 22 Feb 2024 15:11:16 +0900
Subject: [PATCH 24/28] precommit

---
 .../models/visual_prompters/test_segment_anything.py            | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index a1a49b0faea..66b56f6dcbf 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -243,7 +243,7 @@ def test_load_checkpoint_from_local_checkpoint(self, mocker, monkeypatch, checkp
         else:
             mocker_load_from_checkpoint.assert_not_called()
             mocker_load_state_dict.assert_called_once()
-            
+
     @e2e_pytest_unit
     def test_load_checkpoint_without_checkpoint(self, mocker):
         """Test load_checkpoint without checkpoint."""

From d0304275439c04bc46887bee21fb758458a8615d Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 22 Feb 2024 15:55:57 +0900
Subject: [PATCH 25/28] Fix

---
 .../models/visual_prompters/test_segment_anything.py          | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
index 66b56f6dcbf..eecddf412a1 100644
--- a/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
+++ b/tests/unit/algorithms/visual_prompting/adapters/pytorch_lightning/models/visual_prompters/test_segment_anything.py
@@ -175,14 +175,10 @@ def test_load_checkpoint_with_state_dict(self, mocker, state_dict: OrderedDict):
         mocker.patch(
             "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.set_metrics"
         )
-        mocker_load_state_dict = mocker.patch(
-            "otx.algorithms.visual_prompting.adapters.pytorch_lightning.models.visual_prompters.segment_anything.SegmentAnything.load_state_dict"
-        )
 
         sam = SegmentAnything(self.base_config, state_dict=state_dict)
         sam_state_dict = sam.state_dict()
 
-        mocker_load_state_dict.assert_called_once()
         for k, v in state_dict.items():
             assert k in sam_state_dict
             assert torch.all(v == sam_state_dict[k])

From 9d144fd914d40cd7d3753ddd093abf34bfe62aa8 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 22 Feb 2024 16:55:16 +0900
Subject: [PATCH 26/28] Fix

---
 .../visual_prompting/tasks/test_openvino.py   | 54 ++++++++++++++-----
 1 file changed, 40 insertions(+), 14 deletions(-)

diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index 39c5fcfafd8..bc11947b9b7 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -842,7 +842,33 @@ def setup(self, mocker, otx_model):
         mocker.patch.object(
             OpenVINOZeroShotVisualPromptingTask, "load_inferencer", return_value=visual_prompting_ov_inferencer
         )
-        self.visual_prompting_ov_task = OpenVINOZeroShotVisualPromptingTask(task_environment=self.task_environment)
+        self.zero_shot_visual_prompting_ov_task = OpenVINOZeroShotVisualPromptingTask(task_environment=self.task_environment)
+        
+    @e2e_pytest_unit
+    def test_infer(self, mocker):
+        """Test infer."""
+        fake_annotation = [
+            Annotation(
+                Polygon(points=[Point(0, 0)]),
+                id=0,
+                labels=[ScoredLabel(LabelEntity(name="fake", domain="VISUALPROMPTING"), probability=1.0)],
+            )
+        ]
+
+        mocker_predict = mocker.patch.object(OpenVINOZeroShotVisualPromptingInferencer, "predict", return_value=fake_annotation)
+        mocker.patch.object(ShapeFactory, "shape_produces_valid_crop", return_value=True)
+
+        dataset = generate_visual_prompting_dataset()
+
+        updated_dataset = self.zero_shot_visual_prompting_ov_task.infer(
+            dataset, InferenceParameters(enable_async_inference=False)
+        )
+
+        for updated in updated_dataset:
+            assert updated.annotation_scene.contains_any([LabelEntity(name="fake", domain="VISUALPROMPTING")])
+
+        mocker_predict.assert_called()
+        assert mocker_predict.call_count == len(updated_dataset)
 
     @e2e_pytest_unit
     def test_optimize(self, mocker):
@@ -857,43 +883,43 @@ def patch_save_model(model, output_xml):
 
         dataset = generate_visual_prompting_dataset()
         output_model = deepcopy(self.task_environment.model)
-        self.visual_prompting_ov_task.model.set_data("visual_prompting_image_encoder.xml", b"image_encoder_xml")
-        self.visual_prompting_ov_task.model.set_data("visual_prompting_image_encoder.bin", b"image_encoder_bin")
-        self.visual_prompting_ov_task.model.set_data("visual_prompting_prompt_getter.xml", b"prompt_getter_xml")
-        self.visual_prompting_ov_task.model.set_data("visual_prompting_prompt_getter.bin", b"prompt_getter_bin")
-        self.visual_prompting_ov_task.model.set_data("visual_prompting_decoder.xml", b"decoder_xml")
-        self.visual_prompting_ov_task.model.set_data("visual_prompting_decoder.bin", b"decoder_bin")
+        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_image_encoder.xml", b"image_encoder_xml")
+        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_image_encoder.bin", b"image_encoder_bin")
+        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_prompt_getter.xml", b"prompt_getter_xml")
+        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_prompt_getter.bin", b"prompt_getter_bin")
+        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_decoder.xml", b"decoder_xml")
+        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_decoder.bin", b"decoder_bin")
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.ov.Core.read_model", autospec=True)
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.ov.serialize", new=patch_save_model)
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.ov.Core.compile_model")
         fake_quantize = mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.nncf.quantize", autospec=True)
 
-        self.visual_prompting_ov_task.optimize(OptimizationType.POT, dataset=dataset, output_model=output_model)
+        self.zero_shot_visual_prompting_ov_task.optimize(OptimizationType.POT, dataset=dataset, output_model=output_model)
 
         fake_quantize.assert_called()
         assert fake_quantize.call_count == 3
 
         assert (
-            self.visual_prompting_ov_task.model.get_data("visual_prompting_image_encoder.xml")
+            self.zero_shot_visual_prompting_ov_task.model.get_data("visual_prompting_image_encoder.xml")
             == b"compressed_visual_prompting_image_encoder.xml"
         )
         assert (
-            self.visual_prompting_ov_task.model.get_data("visual_prompting_image_encoder.bin")
+            self.zero_shot_visual_prompting_ov_task.model.get_data("visual_prompting_image_encoder.bin")
             == b"compressed_visual_prompting_image_encoder.bin"
         )
         assert (
-            self.visual_prompting_ov_task.model.get_data("visual_prompting_prompt_getter.xml")
+            self.zero_shot_visual_prompting_ov_task.model.get_data("visual_prompting_prompt_getter.xml")
             == b"compressed_visual_prompting_prompt_getter.xml"
         )
         assert (
-            self.visual_prompting_ov_task.model.get_data("visual_prompting_prompt_getter.bin")
+            self.zero_shot_visual_prompting_ov_task.model.get_data("visual_prompting_prompt_getter.bin")
             == b"compressed_visual_prompting_prompt_getter.bin"
         )
         assert (
-            self.visual_prompting_ov_task.model.get_data("visual_prompting_decoder.xml")
+            self.zero_shot_visual_prompting_ov_task.model.get_data("visual_prompting_decoder.xml")
             == b"compressed_visual_prompting_decoder.xml"
         )
         assert (
-            self.visual_prompting_ov_task.model.get_data("visual_prompting_decoder.bin")
+            self.zero_shot_visual_prompting_ov_task.model.get_data("visual_prompting_decoder.bin")
             == b"compressed_visual_prompting_decoder.bin"
         )

From a99d3b97ec417cbb5c182d5bae134670f03c52f0 Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Thu, 22 Feb 2024 17:00:45 +0900
Subject: [PATCH 27/28] Fix

---
 .../visual_prompting/tasks/test_openvino.py   | 30 ++++++++++++++-----
 1 file changed, 22 insertions(+), 8 deletions(-)

diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index bc11947b9b7..f3ba6dae44b 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -842,8 +842,10 @@ def setup(self, mocker, otx_model):
         mocker.patch.object(
             OpenVINOZeroShotVisualPromptingTask, "load_inferencer", return_value=visual_prompting_ov_inferencer
         )
-        self.zero_shot_visual_prompting_ov_task = OpenVINOZeroShotVisualPromptingTask(task_environment=self.task_environment)
-        
+        self.zero_shot_visual_prompting_ov_task = OpenVINOZeroShotVisualPromptingTask(
+            task_environment=self.task_environment
+        )
+
     @e2e_pytest_unit
     def test_infer(self, mocker):
         """Test infer."""
@@ -855,7 +857,9 @@ def test_infer(self, mocker):
             )
         ]
 
-        mocker_predict = mocker.patch.object(OpenVINOZeroShotVisualPromptingInferencer, "predict", return_value=fake_annotation)
+        mocker_predict = mocker.patch.object(
+            OpenVINOZeroShotVisualPromptingInferencer, "predict", return_value=fake_annotation
+        )
         mocker.patch.object(ShapeFactory, "shape_produces_valid_crop", return_value=True)
 
         dataset = generate_visual_prompting_dataset()
@@ -883,10 +887,18 @@ def patch_save_model(model, output_xml):
 
         dataset = generate_visual_prompting_dataset()
         output_model = deepcopy(self.task_environment.model)
-        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_image_encoder.xml", b"image_encoder_xml")
-        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_image_encoder.bin", b"image_encoder_bin")
-        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_prompt_getter.xml", b"prompt_getter_xml")
-        self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_prompt_getter.bin", b"prompt_getter_bin")
+        self.zero_shot_visual_prompting_ov_task.model.set_data(
+            "visual_prompting_image_encoder.xml", b"image_encoder_xml"
+        )
+        self.zero_shot_visual_prompting_ov_task.model.set_data(
+            "visual_prompting_image_encoder.bin", b"image_encoder_bin"
+        )
+        self.zero_shot_visual_prompting_ov_task.model.set_data(
+            "visual_prompting_prompt_getter.xml", b"prompt_getter_xml"
+        )
+        self.zero_shot_visual_prompting_ov_task.model.set_data(
+            "visual_prompting_prompt_getter.bin", b"prompt_getter_bin"
+        )
         self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_decoder.xml", b"decoder_xml")
         self.zero_shot_visual_prompting_ov_task.model.set_data("visual_prompting_decoder.bin", b"decoder_bin")
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.ov.Core.read_model", autospec=True)
@@ -894,7 +906,9 @@ def patch_save_model(model, output_xml):
         mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.ov.Core.compile_model")
         fake_quantize = mocker.patch("otx.algorithms.visual_prompting.tasks.openvino.nncf.quantize", autospec=True)
 
-        self.zero_shot_visual_prompting_ov_task.optimize(OptimizationType.POT, dataset=dataset, output_model=output_model)
+        self.zero_shot_visual_prompting_ov_task.optimize(
+            OptimizationType.POT, dataset=dataset, output_model=output_model
+        )
 
         fake_quantize.assert_called()
         assert fake_quantize.call_count == 3

From 826408f1f7e09140a65345c66ffaa6f2bcaae5ef Mon Sep 17 00:00:00 2001
From: "Kim, Sungchul" <sungchul.kim@intel.com>
Date: Fri, 23 Feb 2024 13:43:03 +0900
Subject: [PATCH 28/28] Fix unittest

---
 .../visual_prompting/tasks/test_openvino.py   | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
index e231e1c7b88..8dab6141cab 100644
--- a/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
+++ b/tests/unit/algorithms/visual_prompting/tasks/test_openvino.py
@@ -847,8 +847,20 @@ def setup(self, mocker, otx_model):
         )
 
     @e2e_pytest_unit
-    def test_infer(self, mocker):
-        """Test infer."""
+    def test_infer_without_reference_info(self):
+        """Test infer without reference_info."""
+        dataset = generate_visual_prompting_dataset()
+
+        updated_dataset = self.zero_shot_visual_prompting_ov_task.infer(
+            dataset, InferenceParameters(enable_async_inference=False)
+        )
+
+        for updated in updated_dataset:
+            assert len(updated.annotation_scene.annotations) == 0
+
+    @e2e_pytest_unit
+    def test_infer_with_reference_info(self, mocker):
+        """Test infer with reference_info."""
         fake_annotation = [
             Annotation(
                 Polygon(points=[Point(0, 0)]),
@@ -861,6 +873,9 @@ def test_infer(self, mocker):
             OpenVINOZeroShotVisualPromptingInferencer, "predict", return_value=fake_annotation
         )
         mocker.patch.object(ShapeFactory, "shape_produces_valid_crop", return_value=True)
+        mocker.patch.object(
+            self.zero_shot_visual_prompting_ov_task.inferencer, "_get_reference_info", return_value=({}, {})
+        )
 
         dataset = generate_visual_prompting_dataset()