Make the same config with OTX1.x : Edit the default factor for lr sch…

…eduler and update warmup_step logic (#3060) * Edit the default factor for lr scheduler and update warmup_step logic * Edit the warmup logic, update the lr shceduler patience * Fix unit tests * Fix the logic * Fix the logic * Fix precommit
openvinotoolkit · Mar 11, 2024 · ec68bac · ec68bac
1 parent 25fc3e1
commit ec68bac
Show file tree

Hide file tree

Showing 55 changed files with 93 additions and 88 deletions.
diff --git a/src/otx/algo/callbacks/adaptive_train_scheduling.py b/src/otx/algo/callbacks/adaptive_train_scheduling.py
@@ -124,7 +124,11 @@ def _revert_patience(lr_scheduler: LRSchedulerTypeUnion, saved_patience: int) ->
                 if hasattr(config, "scheduler") and hasattr(config.scheduler, "patience"):
                     saved_patience = config.scheduler.patience
                     adjusted_patience = (
-                        max(int(config.scheduler.patience / adaptive_interval), self.min_lrschedule_patience) - 1
+                        max(
+                            int((config.scheduler.patience + 1) / adaptive_interval),
+                            self.min_lrschedule_patience,
+                        )
+                        - 1
                     )
                     config.scheduler.patience = adjusted_patience
 
@@ -133,7 +137,7 @@ def _revert_patience(lr_scheduler: LRSchedulerTypeUnion, saved_patience: int) ->
                         f"{saved_patience} --> {adjusted_patience}."
                     )
                     log.warning(msg)
-                    self._revert_lr_patience += [partial(_revert_patience, config, saved_patience)]
+                    self._revert_lr_patience += [partial(_revert_patience, config.scheduler, saved_patience)]
 
     def _change_early_stopping_patience(self, callbacks: list[Callback], adaptive_interval: int) -> None:
         """Change the EarlyStopping patience to change the patience.

diff --git a/src/otx/algo/schedulers/warmup_schedulers.py b/src/otx/algo/schedulers/warmup_schedulers.py
@@ -21,9 +21,9 @@ def __init__(
             raise ValueError(msg)
         self.num_warmup_steps = num_warmup_steps
         self.interval = interval
-        super().__init__(optimizer, lambda step: min(step / num_warmup_steps, 1.0))
+        super().__init__(optimizer, lambda step: min((step + 1.0) / self.num_warmup_steps, 1.0))
 
     def step(self, epoch: int | None = None) -> None:
         """Overriding the step to disable the warmup scheduler after n_steps."""
-        if self._step_count < self.num_warmup_steps:
+        if self._step_count <= self.num_warmup_steps:
             super().step(epoch)
diff --git a/src/otx/recipe/action/action_classification/x3d.yaml b/src/otx/recipe/action/action_classification/x3d.yaml
@@ -13,8 +13,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 2
+      factor: 0.1
+      patience: 1
       monitor: val/accuracy
 
 metric:

diff --git a/src/otx/recipe/action/action_detection/x3d_fastrcnn.yaml b/src/otx/recipe/action/action_detection/x3d_fastrcnn.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 2
+      factor: 0.1
+      patience: 1
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/classification/h_label_cls/efficientnet_b0_light.yaml b/src/otx/recipe/classification/h_label_cls/efficientnet_b0_light.yaml
@@ -14,7 +14,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/h_label_cls/efficientnet_v2_light.yaml b/src/otx/recipe/classification/h_label_cls/efficientnet_v2_light.yaml
@@ -16,7 +16,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/h_label_cls/mobilenet_v3_large_light.yaml b/src/otx/recipe/classification/h_label_cls/mobilenet_v3_large_light.yaml
@@ -19,7 +19,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/h_label_cls/openvino_model.yaml b/src/otx/recipe/classification/h_label_cls/openvino_model.yaml
@@ -20,7 +20,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: min
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: train/loss
 

diff --git a/src/otx/recipe/classification/h_label_cls/otx_deit_tiny.yaml b/src/otx/recipe/classification/h_label_cls/otx_deit_tiny.yaml
@@ -18,7 +18,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/efficientnet_b0_light.yaml b/src/otx/recipe/classification/multi_class_cls/efficientnet_b0_light.yaml
@@ -15,7 +15,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/efficientnet_v2_light.yaml b/src/otx/recipe/classification/multi_class_cls/efficientnet_v2_light.yaml
@@ -15,7 +15,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/mobilenet_v3_large_light.yaml b/src/otx/recipe/classification/multi_class_cls/mobilenet_v3_large_light.yaml
@@ -18,7 +18,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/openvino_model.yaml b/src/otx/recipe/classification/multi_class_cls/openvino_model.yaml
@@ -17,7 +17,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: min
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: train/loss
 

diff --git a/src/otx/recipe/classification/multi_class_cls/otx_deit_tiny.yaml b/src/otx/recipe/classification/multi_class_cls/otx_deit_tiny.yaml
@@ -16,7 +16,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/otx_dino_v2.yaml b/src/otx/recipe/classification/multi_class_cls/otx_dino_v2.yaml
@@ -23,7 +23,7 @@ scheduler:
   init_args:
     mode: min
     factor: 0.1
-    patience: 10
+    patience: 9
     monitor: train/loss
 
 metric:

diff --git a/src/otx/recipe/classification/multi_class_cls/otx_dino_v2_linear_probe.yaml b/src/otx/recipe/classification/multi_class_cls/otx_dino_v2_linear_probe.yaml
@@ -24,7 +24,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: min
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: train/loss
 

diff --git a/src/otx/recipe/classification/multi_class_cls/otx_efficientnet_b0.yaml b/src/otx/recipe/classification/multi_class_cls/otx_efficientnet_b0.yaml
@@ -15,7 +15,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/otx_efficientnet_v2.yaml b/src/otx/recipe/classification/multi_class_cls/otx_efficientnet_v2.yaml
@@ -15,7 +15,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_class_cls/otx_mobilenet_v3_large.yaml b/src/otx/recipe/classification/multi_class_cls/otx_mobilenet_v3_large.yaml
@@ -18,7 +18,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_label_cls/efficientnet_b0_light.yaml b/src/otx/recipe/classification/multi_label_cls/efficientnet_b0_light.yaml
@@ -12,7 +12,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_label_cls/efficientnet_v2_light.yaml b/src/otx/recipe/classification/multi_label_cls/efficientnet_v2_light.yaml
@@ -14,7 +14,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: max
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_label_cls/mobilenet_v3_large_light.yaml b/src/otx/recipe/classification/multi_label_cls/mobilenet_v3_large_light.yaml
@@ -17,7 +17,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/classification/multi_label_cls/openvino_model.yaml b/src/otx/recipe/classification/multi_label_cls/openvino_model.yaml
@@ -17,7 +17,7 @@ scheduler:
   class_path: lightning.pytorch.cli.ReduceLROnPlateau
   init_args:
     mode: min
-    factor: 0.5
+    factor: 0.1
     patience: 1
     monitor: train/loss
 

diff --git a/src/otx/recipe/classification/multi_label_cls/otx_deit_tiny.yaml b/src/otx/recipe/classification/multi_label_cls/otx_deit_tiny.yaml
@@ -16,7 +16,7 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
+      factor: 0.1
       patience: 1
       monitor: val/accuracy
 

diff --git a/src/otx/recipe/detection/atss_mobilenetv2.yaml b/src/otx/recipe/detection/atss_mobilenetv2.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/atss_mobilenetv2_tile.yaml b/src/otx/recipe/detection/atss_mobilenetv2_tile.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/atss_r50_fpn.yaml b/src/otx/recipe/detection/atss_r50_fpn.yaml
@@ -16,8 +16,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/atss_resnext101.yaml b/src/otx/recipe/detection/atss_resnext101.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/openvino_model.yaml b/src/otx/recipe/detection/openvino_model.yaml
@@ -18,7 +18,7 @@ scheduler:
   init_args:
     mode: min
     factor: 0.1
-    patience: 10
+    patience: 9
     monitor: train/loss
 
 engine:

diff --git a/src/otx/recipe/detection/rtmdet_tiny.yaml b/src/otx/recipe/detection/rtmdet_tiny.yaml
@@ -15,7 +15,7 @@ scheduler:
   init_args:
     mode: min
     factor: 0.1
-    patience: 10
+    patience: 9
     monitor: train/loss
 
 metric:

diff --git a/src/otx/recipe/detection/ssd_mobilenetv2.yaml b/src/otx/recipe/detection/ssd_mobilenetv2.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/ssd_mobilenetv2_tile.yaml b/src/otx/recipe/detection/ssd_mobilenetv2_tile.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_l.yaml b/src/otx/recipe/detection/yolox_l.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_l_tile.yaml b/src/otx/recipe/detection/yolox_l_tile.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_s.yaml b/src/otx/recipe/detection/yolox_s.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_s_tile.yaml b/src/otx/recipe/detection/yolox_s_tile.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_tiny.yaml b/src/otx/recipe/detection/yolox_tiny.yaml
@@ -17,8 +17,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_tiny_tile.yaml b/src/otx/recipe/detection/yolox_tiny_tile.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric:

diff --git a/src/otx/recipe/detection/yolox_x.yaml b/src/otx/recipe/detection/yolox_x.yaml
@@ -18,8 +18,8 @@ scheduler:
   - class_path: lightning.pytorch.cli.ReduceLROnPlateau
     init_args:
       mode: max
-      factor: 0.5
-      patience: 5
+      factor: 0.1
+      patience: 4
       monitor: val/map_50
 
 metric: