fix converter and early stopping + fix warmup epochs

openvinotoolkit · Nov 25, 2024 · de025a6 · de025a6
1 parent ec610a9
commit de025a6
Show file tree

Hide file tree

Showing 42 changed files with 204 additions and 107 deletions.
diff --git a/src/otx/core/model/base.py b/src/otx/core/model/base.py
@@ -744,7 +744,7 @@ def lr_scheduler_step(self, scheduler: LRSchedulerTypeUnion, metric: Tensor) ->
             return super().lr_scheduler_step(scheduler=scheduler, metric=metric)
 
         if len(warmup_schedulers) != 1:
-            msg = "No more than two warmup schedulers coexist."
+            msg = "No more than one warmup schedulers coexist."
             raise RuntimeError(msg)
 
         warmup_scheduler = next(iter(warmup_schedulers))

diff --git a/src/otx/core/schedulers/warmup_schedulers.py b/src/otx/core/schedulers/warmup_schedulers.py
@@ -19,16 +19,17 @@ class LinearWarmupScheduler(LambdaLR):
     """Linear Warmup scheduler.
 
     Args:
-        num_warmup_steps: Learning rate will linearly increased during the period same as this number.
-        warmup_interval: If "epoch", count the number of steps for the warmup period.
+        optimizer (Optimizer): Optimizer to apply the scheduler.
+        num_warmup_steps (int): Learning rate will linearly increased during the period same as this number.
+        interval (Literal["step", "epoch"]): If "epoch", count the number of epochs for the warmup period.
             Otherwise, the iteration step will be the warmup period.
     """
 
     def __init__(
         self,
         optimizer: Optimizer,
         num_warmup_steps: int = 1000,
-        interval: Literal["step", "epoch"] = "step",
+        interval: Literal["step", "epoch"] = "epoch",
     ):
         if not num_warmup_steps > 0:
             msg = f"num_warmup_steps should be > 0, got {num_warmup_steps}"
@@ -55,7 +56,7 @@ class LinearWarmupSchedulerCallable:
         main_scheduler_callable: Callable to create a LR scheduler that will be mainly used.
         num_warmup_steps: Learning rate will linearly increased during the period same as this number.
             If it is less than equal to zero, do not create `LinearWarmupScheduler`.
-        warmup_interval: If "epoch", count the number of steps for the warmup period.
+        warmup_interval: If "epoch", count the number of epochs for the warmup period.
             Otherwise, the iteration step will be the warmup period.
         monitor: If given, override the main scheduler's `monitor` attribute.
     """
@@ -64,7 +65,7 @@ def __init__(
         self,
         main_scheduler_callable: LRSchedulerCallable,
         num_warmup_steps: int = 0,
-        warmup_interval: Literal["step", "epoch"] = "step",
+        warmup_interval: Literal["step", "epoch"] = "epoch",
         monitor: str | None = None,
     ):
         self.main_scheduler_callable = SchedulerCallableSupportHPO.from_callable(main_scheduler_callable)

diff --git a/src/otx/recipe/classification/h_label_cls/deit_tiny.yaml b/src/otx/recipe/classification/h_label_cls/deit_tiny.yaml
@@ -10,12 +10,16 @@ model:
         weight_decay: 0.05
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: H_LABEL_CLS

diff --git a/src/otx/recipe/classification/h_label_cls/efficientnet_b0.yaml b/src/otx/recipe/classification/h_label_cls/efficientnet_b0.yaml
@@ -11,12 +11,16 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: H_LABEL_CLS

diff --git a/src/otx/recipe/classification/h_label_cls/efficientnet_v2.yaml b/src/otx/recipe/classification/h_label_cls/efficientnet_v2.yaml
@@ -10,6 +10,18 @@ model:
         momentum: 0.9
         weight_decay: 0.0001
 
+    scheduler:
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
+      init_args:
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
+
 engine:
   task: H_LABEL_CLS
   device: auto

diff --git a/src/otx/recipe/classification/h_label_cls/mobilenet_v3_large.yaml b/src/otx/recipe/classification/h_label_cls/mobilenet_v3_large.yaml
@@ -19,7 +19,7 @@ model:
           init_args:
             mode: max
             factor: 0.5
-            patience: 1
+            patience: 5
             monitor: val/accuracy
 
 engine:

diff --git a/src/otx/recipe/classification/h_label_cls/tv_efficientnet_b3.yaml b/src/otx/recipe/classification/h_label_cls/tv_efficientnet_b3.yaml
@@ -12,10 +12,14 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        T_max: 100000
-        eta_min: 0
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+          init_args:
+            T_max: 100000
+            eta_min: 0
 
 engine:
   task: H_LABEL_CLS

diff --git a/src/otx/recipe/classification/h_label_cls/tv_efficientnet_v2_l.yaml b/src/otx/recipe/classification/h_label_cls/tv_efficientnet_v2_l.yaml
@@ -12,10 +12,14 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        T_max: 100000
-        eta_min: 0
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+          init_args:
+            T_max: 100000
+            eta_min: 0
 
 engine:
   task: H_LABEL_CLS

diff --git a/src/otx/recipe/classification/h_label_cls/tv_mobilenet_v3_small.yaml b/src/otx/recipe/classification/h_label_cls/tv_mobilenet_v3_small.yaml
@@ -12,10 +12,14 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        T_max: 100000
-        eta_min: 0
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+          init_args:
+            T_max: 100000
+            eta_min: 0
 
 engine:
   task: H_LABEL_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/deit_tiny.yaml b/src/otx/recipe/classification/multi_class_cls/deit_tiny.yaml
@@ -12,12 +12,16 @@ model:
         weight_decay: 0.05
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/dino_v2.yaml b/src/otx/recipe/classification/multi_class_cls/dino_v2.yaml
@@ -11,12 +11,16 @@ model:
         weight_decay: 0.05
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/efficientnet_b0.yaml b/src/otx/recipe/classification/multi_class_cls/efficientnet_b0.yaml
@@ -12,12 +12,16 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/efficientnet_v2.yaml b/src/otx/recipe/classification/multi_class_cls/efficientnet_v2.yaml
@@ -12,12 +12,16 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/mobilenet_v3_large.yaml b/src/otx/recipe/classification/multi_class_cls/mobilenet_v3_large.yaml
@@ -20,7 +20,7 @@ model:
           init_args:
             mode: max
             factor: 0.5
-            patience: 1
+            patience: 5
             monitor: val/accuracy
 
 engine:

diff --git a/src/otx/recipe/classification/multi_class_cls/tv_efficientnet_b3.yaml b/src/otx/recipe/classification/multi_class_cls/tv_efficientnet_b3.yaml
@@ -12,10 +12,14 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        T_max: 100000
-        eta_min: 0
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+          init_args:
+            T_max: 100000
+            eta_min: 0
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/tv_efficientnet_v2_l.yaml b/src/otx/recipe/classification/multi_class_cls/tv_efficientnet_v2_l.yaml
@@ -12,10 +12,14 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        T_max: 100000
-        eta_min: 0
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+          init_args:
+            T_max: 100000
+            eta_min: 0
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_class_cls/tv_mobilenet_v3_small.yaml b/src/otx/recipe/classification/multi_class_cls/tv_mobilenet_v3_small.yaml
@@ -12,10 +12,14 @@ model:
         weight_decay: 0.0001
 
     scheduler:
-      class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        T_max: 100000
-        eta_min: 0
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: torch.optim.lr_scheduler.CosineAnnealingLR
+          init_args:
+            T_max: 100000
+            eta_min: 0
 
 engine:
   task: MULTI_CLASS_CLS

diff --git a/src/otx/recipe/classification/multi_label_cls/deit_tiny.yaml b/src/otx/recipe/classification/multi_label_cls/deit_tiny.yaml
@@ -11,12 +11,16 @@ model:
         weight_decay: 0.05
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_LABEL_CLS

diff --git a/src/otx/recipe/classification/multi_label_cls/efficientnet_b0.yaml b/src/otx/recipe/classification/multi_label_cls/efficientnet_b0.yaml
@@ -12,12 +12,16 @@ model:
         weight_decay: 0.0005
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_LABEL_CLS

diff --git a/src/otx/recipe/classification/multi_label_cls/efficientnet_v2.yaml b/src/otx/recipe/classification/multi_label_cls/efficientnet_v2.yaml
@@ -12,12 +12,16 @@ model:
         weight_decay: 0.0005
 
     scheduler:
-      class_path: lightning.pytorch.cli.ReduceLROnPlateau
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
       init_args:
-        mode: max
-        factor: 0.5
-        patience: 1
-        monitor: val/accuracy
+        num_warmup_steps: 0
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 5
+            monitor: val/accuracy
 
 engine:
   task: MULTI_LABEL_CLS

diff --git a/src/otx/recipe/classification/multi_label_cls/mobilenet_v3_large.yaml b/src/otx/recipe/classification/multi_label_cls/mobilenet_v3_large.yaml
@@ -20,7 +20,7 @@ model:
           init_args:
             mode: max
             factor: 0.5
-            patience: 1
+            patience: 5
             monitor: val/accuracy
 
 engine: