Fix the Anomaly Task to work & Add integration test for Anomaly

.github/workflows/pre_merge.yaml

@@ -91,6 +91,7 @@ jobs:
           - task: "instance_segmentation"
           - task: "semantic_segmentation"
           - task: "visual_prompting"
+          - task: "anomaly"
     name: Integration-Test-${{ matrix.task }}-py310
     # This is what will cancel the job concurrency
     concurrency:

pyproject.toml

@@ -89,6 +89,18 @@ mmlab = [
 ]
 anomaly = [
     "anomalib==1.0.0",
+    # This is a dependency to avoid conflicts with installing the anomalib[core] option.
+    "av>=10.0.0",
+    "einops>=0.3.2",
+    "freia>=0.2",
+    "imgaug==0.4.0",
+    "kornia>=0.6.6,<0.6.10",
+    "matplotlib>=3.4.3",
+    "opencv-python>=4.5.3.56",
+    "pandas>=1.1.0",
+    "open-clip-torch>=2.23.0",
+    # issue with latest version for torchmetrics with anomalib: https://github.com/Lightning-AI/torchmetrics/issues/1526
+    "torchmetrics==0.10.3",
 ]
 
 [project.scripts]

src/otx/algo/anomaly/padim.py

@@ -15,26 +15,26 @@ class Padim(OTXAnomaly, OTXModel, AnomalibPadim):
     """OTX Padim model.
 
     Args:
-        input_size (tuple[int, int], optional): Input size. Defaults to (256, 256).
         backbone (str, optional): Feature extractor backbone. Defaults to "resnet18".
         layers (list[str], optional): Feature extractor layers. Defaults to ["layer1", "layer2", "layer3"].
         pre_trained (bool, optional): Pretrained backbone. Defaults to True.
         n_features (int | None, optional): Number of features. Defaults to None.
+        num_classes (int, optional): Anoamly don't use num_classes ,
+            but OTXModel always receives num_classes, so need this.
     """
 
     def __init__(
         self,
-        input_size: tuple[int, int] = (256, 256),
         backbone: str = "resnet18",
         layers: list[str] = ["layer1", "layer2", "layer3"],  # noqa: B006
         pre_trained: bool = True,
         n_features: int | None = None,
+        num_classes: int = 2,
     ) -> None:
         OTXAnomaly.__init__(self)
-        OTXModel.__init__(self, num_classes=2)
+        OTXModel.__init__(self, num_classes=num_classes)
         AnomalibPadim.__init__(
             self,
-            input_size=input_size,
             backbone=backbone,
             layers=layers,
             pre_trained=pre_trained,

src/otx/algo/anomaly/stfpm.py

@@ -21,21 +21,21 @@ class Stfpm(OTXAnomaly, OTXModel, AnomalibStfpm):
 
     Args:
         layers (Sequence[str]): Feature extractor layers.
-        input_size (tuple[int, int]): Input size.
         backbone (str, optional): Feature extractor backbone. Defaults to "resnet18".
+        num_classes (int, optional): Anoamly don't use num_classes ,
+            but OTXModel always receives num_classes, so need this.
     """
 
     def __init__(
         self,
         layers: Sequence[str] = ["layer1", "layer2", "layer3"],
-        input_size: tuple[int, int] = (256, 256),
         backbone: str = "resnet18",
+        num_classes: int = 2,
     ) -> None:
         OTXAnomaly.__init__(self)
-        OTXModel.__init__(self, num_classes=2)
+        OTXModel.__init__(self, num_classes=num_classes)
         AnomalibStfpm.__init__(
             self,
-            input_size=input_size,
             backbone=backbone,
             layers=layers,
         )

src/otx/algo/instance_segmentation/otx_instseg_evaluation.py

@@ -26,7 +26,7 @@ def update(self, preds: list[dict], target: list[dict]) -> None:
             target (list[dict]): list of RLE encoded masks
         """
         for item in preds:
-            bbox_detection, mask_detection = self._get_safe_item_values(item, warn=self.warn_on_many_detections)
+            bbox_detection, mask_detection = self._get_safe_item_values(item)
             if bbox_detection is not None:
                 self.detection_box.append(bbox_detection)
             if mask_detection is not None:

src/otx/cli/cli.py

@@ -413,7 +413,7 @@ def instantiate_model(self, model_config: Namespace) -> tuple:
         # Update num_classes
         if not self.get_config_value(self.config_init, "disable_infer_num_classes", False):
             num_classes = self.datamodule.label_info.num_classes
-            if num_classes != model_config.init_args.num_classes:
+            if hasattr(model_config.init_args, "num_classes") and num_classes != model_config.init_args.num_classes:
                 warning_msg = (
                     f"The `num_classes` in dataset is {num_classes} "
                     f"but, the `num_classes` of model is {model_config.init_args.num_classes}. "

src/otx/core/data/entity/anomaly/classification.py

@@ -53,7 +53,7 @@ def collate_fn(
     ) -> AnomalyClassificationDataBatch:
         """Collection function to collect `OTXDataEntity` into `OTXBatchDataEntity` in data loader."""
         batch = super().collate_fn(entities)
-        images = tv_tensors.Image(data=torch.stack(batch.images, dim=0)) if stack_images else batch.images
+        images = tv_tensors.Image(data=torch.stack(tuple(batch.images), dim=0)) if stack_images else batch.images
         return AnomalyClassificationDataBatch(
             batch_size=batch.batch_size,
             images=images,

src/otx/core/data/entity/anomaly/detection.py

@@ -56,7 +56,7 @@ def collate_fn(
     ) -> AnomalyDetectionDataBatch:
         """Collection function to collect `OTXDataEntity` into `OTXBatchDataEntity` in data loader."""
         batch = super().collate_fn(entities)
-        images = tv_tensors.Image(data=torch.stack(batch.images, dim=0)) if stack_images else batch.images
+        images = tv_tensors.Image(data=torch.stack(tuple(batch.images), dim=0)) if stack_images else batch.images
         return AnomalyDetectionDataBatch(
             batch_size=batch.batch_size,
             images=images,

src/otx/core/metrics/accuracy.py

@@ -272,6 +272,7 @@ def __init__(
             TorchmetricAcc(
                 task="multiclass",
                 num_classes=int(head_range[1] - head_range[0]),
+                top_k=1,
             )
             for head_range in head_logits_info.values()
         ]

src/otx/core/model/module/anomaly/anomaly_lightning.py

@@ -134,7 +134,7 @@ class OTXAnomaly:
     def __init__(self) -> None:
         self.optimizer: list[OptimizerCallable] | OptimizerCallable = None
         self.scheduler: list[LRSchedulerCallable] | LRSchedulerCallable = None
-        self.input_size: list[int] = [256, 256]
+        self.input_size: tuple[int] = (256, 256)
         self.mean_values: tuple[float, float, float] = (0.0, 0.0, 0.0)
         self.scale_values: tuple[float, float, float] = (1.0, 1.0, 1.0)
         self.trainer: Trainer
@@ -147,6 +147,19 @@ def __init__(self) -> None:
         self.image_metrics: AnomalibMetricCollection
         self.pixel_metrics: AnomalibMetricCollection
 
+    @property
+    def input_size(self) -> tuple[int, int]:
+        """Returns the input size of the model.
+
+        Returns:
+            tuple[int, int]: The input size of the model as a tuple of (height, width).
+        """
+        return self._input_size
+
+    @input_size.setter
+    def input_size(self, value: tuple[int, int]) -> None:
+        self._input_size = value
+
     @property
     def task(self) -> AnomalibTaskType:
         """Return the task type of the model."""
@@ -342,13 +355,13 @@ def state_dict(self) -> dict[str, Any]:
         """
         state_dict = super().state_dict()  # type: ignore[misc]
         # This is defined in OTXModel
-        state_dict["meta_info"] = self.meta_info  # type: ignore[attr-defined]
+        state_dict["label_info"] = self.label_info  # type: ignore[attr-defined]
         return state_dict
 
     def load_state_dict(self, ckpt: OrderedDict[str, Any], *args, **kwargs) -> None:
         """Pass the checkpoint to the anomaly model."""
         ckpt = ckpt.get("state_dict", ckpt)
-        ckpt.pop("meta_info", None)  # [TODO](ashwinvaidya17): Revisit this method when OTXModel is the lightning model
+        ckpt.pop("label_info", None)  # [TODO](ashwinvaidya17): Revisit this method when OTXModel is the lightning model
         return super().load_state_dict(ckpt, *args, **kwargs)  # type: ignore[misc]
 
     def forward(

src/otx/engine/utils/auto_configurator.py

@@ -42,9 +42,9 @@
     OTXTaskType.INSTANCE_SEGMENTATION: RECIPE_PATH / "instance_segmentation" / "maskrcnn_r50.yaml",
     OTXTaskType.ACTION_CLASSIFICATION: RECIPE_PATH / "action" / "action_classification" / "x3d.yaml",
     OTXTaskType.ACTION_DETECTION: RECIPE_PATH / "action" / "action_detection" / "x3d_fastrcnn.yaml",
-    OTXTaskType.ANOMALY_CLASSIFICATION: RECIPE_PATH / "anomaly" / "anomaly_classification" / "padim.yaml",
-    OTXTaskType.ANOMALY_SEGMENTATION: RECIPE_PATH / "anomaly" / "anomaly_segmentation" / "padim.yaml",
-    OTXTaskType.ANOMALY_DETECTION: RECIPE_PATH / "anomaly" / "anomaly_detection" / "padim.yaml",
+    OTXTaskType.ANOMALY_CLASSIFICATION: RECIPE_PATH / "anomaly_classification" / "padim.yaml",
+    OTXTaskType.ANOMALY_SEGMENTATION: RECIPE_PATH / "anomaly_segmentation" / "padim.yaml",
+    OTXTaskType.ANOMALY_DETECTION: RECIPE_PATH / "anomaly_detection" / "padim.yaml",
     OTXTaskType.VISUAL_PROMPTING: RECIPE_PATH / "visual_prompting" / "sam_tiny_vit.yaml",
     OTXTaskType.ZERO_SHOT_VISUAL_PROMPTING: RECIPE_PATH / "zero_shot_visual_prompting" / "sam_tiny_vit.yaml",
 }
@@ -67,6 +67,7 @@
     "common_semantic_segmentation_with_subset_dirs": [OTXTaskType.SEMANTIC_SEGMENTATION],
     "kinetics": [OTXTaskType.ACTION_CLASSIFICATION],
     "ava": [OTXTaskType.ACTION_DETECTION],
+    "mvtec": [OTXTaskType.ANOMALY_CLASSIFICATION, OTXTaskType.ANOMALY_DETECTION, OTXTaskType.ANOMALY_SEGMENTATION],
 }
 
 OVMODEL_PER_TASK = {

src/otx/recipe/anomaly_classification/padim.yaml

@@ -1,9 +1,6 @@
 model:
   class_path: otx.algo.anomaly.padim.Padim
   init_args:
-    input_size:
-      - 256
-      - 256
     layers: ["layer1", "layer2", "layer3"]
     backbone: "resnet18"
     pre_trained: True
@@ -15,10 +12,15 @@ engine:
 
 callback_monitor: step # this has no effect as Padim does not need to be trained
 
-data: ../../_base_/data/torchvision_base.yaml
+data: ../_base_/data/torchvision_base.yaml
 overrides:
+  precision: 32
   max_epochs: 1
   limit_val_batches: 0 # this is set to 0 as the default dataloader does not have validation set. But this also means that the model will not give correct performance numbers
+  callbacks:
+    - class_path: otx.algo.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
   data:
     task: ANOMALY_CLASSIFICATION
     config:

src/otx/recipe/anomaly_classification/stfpm.yaml

@@ -1,9 +1,6 @@
 model:
   class_path: otx.algo.anomaly.stfpm.Stfpm
   init_args:
-    input_size:
-      - 256
-      - 256
     layers: ["layer1", "layer2", "layer3"]
     backbone: "resnet18"
 
@@ -21,14 +18,17 @@ engine:
 
 callback_monitor: train_loss_epoch # val loss is not available as there is no validation set from default dataloader
 
-data: ../../_base_/data/torchvision_base.yaml
+data: ../_base_/data/torchvision_base.yaml
 overrides:
   max_epochs: 100
   limit_val_batches: 0 # this is set to 0 as the default dataloader does not have validation set. But this also means that the model will not give correct performance numbers
   callbacks:
     - class_path: lightning.pytorch.callbacks.EarlyStopping
       init_args:
         patience: 5
+    - class_path: otx.algo.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
   data:
     task: ANOMALY_CLASSIFICATION
     config:

src/otx/recipe/anomaly_detection/padim.yaml

@@ -1,9 +1,6 @@
 model:
   class_path: otx.algo.anomaly.padim.Padim
   init_args:
-    input_size:
-      - 256
-      - 256
     layers: ["layer1", "layer2", "layer3"]
     backbone: "resnet18"
     pre_trained: True
@@ -15,10 +12,15 @@ engine:
 
 callback_monitor: step # this has no effect as Padim does not need to be trained
 
-data: ../../_base_/data/torchvision_base.yaml
+data: ../_base_/data/torchvision_base.yaml
 overrides:
+  precision: 32
   max_epochs: 1
   limit_val_batches: 0 # this is set to 0 as the default dataloader does not have validation set. But this also means that the model will not give correct performance numbers
+  callbacks:
+    - class_path: otx.algo.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
   data:
     task: ANOMALY_DETECTION
     config:

src/otx/recipe/anomaly_detection/stfpm.yaml

@@ -1,9 +1,6 @@
 model:
   class_path: otx.algo.anomaly.stfpm.Stfpm
   init_args:
-    input_size:
-      - 256
-      - 256
     layers: ["layer1", "layer2", "layer3"]
     backbone: "resnet18"
 
@@ -21,14 +18,17 @@ engine:
 
 callback_monitor: train_loss_epoch # val loss is not available as there is no validation set from default dataloader
 
-data: ../../_base_/data/torchvision_base.yaml
+data: ../_base_/data/torchvision_base.yaml
 overrides:
   max_epochs: 100
   limit_val_batches: 0 # this is set to 0 as the default dataloader does not have validation set. But this also means that the model will not give correct performance numbers
   callbacks:
     - class_path: lightning.pytorch.callbacks.EarlyStopping
       init_args:
         patience: 5
+    - class_path: otx.algo.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
   data:
     task: ANOMALY_DETECTION
     config:

src/otx/recipe/anomaly_segmentation/padim.yaml

@@ -1,9 +1,6 @@
 model:
   class_path: otx.algo.anomaly.padim.Padim
   init_args:
-    input_size:
-      - 256
-      - 256
     layers: ["layer1", "layer2", "layer3"]
     backbone: "resnet18"
     pre_trained: True
@@ -15,10 +12,15 @@ engine:
 
 callback_monitor: step # this has no effect as Padim does not need to be trained
 
-data: ../../_base_/data/torchvision_base.yaml
+data: ../_base_/data/torchvision_base.yaml
 overrides:
+  precision: 32
   max_epochs: 1
   limit_val_batches: 0 # this is set to 0 as the default dataloader does not have validation set. But this also means that the model will not give correct performance numbers
+  callbacks:
+    - class_path: otx.algo.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
   data:
     task: ANOMALY_SEGMENTATION
     config:

src/otx/recipe/anomaly_segmentation/stfpm.yaml

@@ -21,14 +21,17 @@ engine:
 
 callback_monitor: train_loss_epoch # val loss is not available as there is no validation set from default dataloader
 
-data: ../../_base_/data/torchvision_base.yaml
+data: ../_base_/data/torchvision_base.yaml
 overrides:
   max_epochs: 100
   limit_val_batches: 0 # this is set to 0 as the default dataloader does not have validation set. But this also means that the model will not give correct performance numbers
   callbacks:
     - class_path: lightning.pytorch.callbacks.EarlyStopping
       init_args:
         patience: 5
+    - class_path: otx.algo.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
   data:
     task: ANOMALY_SEGMENTATION
     config:

src/otx/recipe/classification/multi_class_cls/efficientnet_b0_light.yaml

@@ -24,6 +24,7 @@ metric:
   init_args:
     task: multiclass
     num_classes: 1000
+    top_k: 1
 
 engine:
   task: MULTI_CLASS_CLS

src/otx/recipe/classification/multi_class_cls/efficientnet_v2_light.yaml

@@ -24,6 +24,7 @@ metric:
   init_args:
     task: multiclass
     num_classes: 1000
+    top_k: 1
 
 engine:
   task: MULTI_CLASS_CLS

src/otx/recipe/classification/multi_class_cls/mobilenet_v3_large_light.yaml

@@ -27,6 +27,7 @@ metric:
   init_args:
     task: multiclass
     num_classes: 1000
+    top_k: 1
 
 engine:
   task: MULTI_CLASS_CLS

src/otx/recipe/classification/multi_class_cls/otx_deit_tiny.yaml

@@ -25,6 +25,7 @@ metric:
   init_args:
     task: multiclass
     num_classes: 1000
+    top_k: 1
 
 engine:
   task: MULTI_CLASS_CLS

src/otx/recipe/classification/multi_class_cls/otx_dino_v2.yaml

@@ -31,6 +31,7 @@ metric:
   init_args:
     task: multiclass
     num_classes: 1000
+    top_k: 1
 
 engine:
   task: MULTI_CLASS_CLS

src/otx/recipe/classification/multi_class_cls/otx_dino_v2_linear_probe.yaml

@@ -33,6 +33,7 @@ metric:
   init_args:
     task: multiclass
     num_classes: 1000
+    top_k: 1
 
 engine:
   task: MULTI_CLASS_CLS