Refactor MaskRCNN rotated detection (#3858)

* Refactor MaskRCNN rotated detection

* update rotated ov model recipe

* update rotated det recipes

src/otx/core/model/rotated_detection.py

@@ -9,11 +9,12 @@
 from datumaro import Polygon
 from torchvision import tv_tensors
 
+from otx.algo.instance_segmentation.maskrcnn import MaskRCNN, MaskRCNNEfficientNet, MaskRCNNResNet50
 from otx.core.data.entity.instance_segmentation import InstanceSegBatchPredEntity
-from otx.core.model.instance_segmentation import OTXInstanceSegModel, OVInstanceSegmentationModel
+from otx.core.model.instance_segmentation import OVInstanceSegmentationModel
 
 
-class OTXRotatedDetModel(OTXInstanceSegModel):
+class RotatedMaskRCNNModel(MaskRCNN):
     """Base class for the rotated detection models used in OTX."""
 
     def predict_step(self, *args: torch.Any, **kwargs: torch.Any) -> InstanceSegBatchPredEntity:
@@ -93,6 +94,14 @@ def predict_step(self, *args: torch.Any, **kwargs: torch.Any) -> InstanceSegBatc
         )
 
 
+class RotatedMaskRCNNResNet50(RotatedMaskRCNNModel, MaskRCNNResNet50):
+    """Rotated MaskRCNN model with ResNet50 backbone."""
+
+
+class RotatedMaskRCNNEfficientNet(RotatedMaskRCNNModel, MaskRCNNEfficientNet):
+    """Rotated MaskRCNN model with EfficientNet backbone."""
+
+
 class OVRotatedDetectionModel(OVInstanceSegmentationModel):
     """Rotated Detection model compatible for OpenVINO IR Inference.
 

src/otx/recipe/rotated_detection/maskrcnn_efficientnetb2b.yaml

@@ -1,5 +1,5 @@
 model:
-  class_path: otx.algo.instance_segmentation.maskrcnn.MaskRCNNEfficientNet
+  class_path: otx.core.model.rotated_detection.RotatedMaskRCNNEfficientNet
   init_args:
     label_info: 80
 
@@ -28,29 +28,40 @@ engine:
 
 callback_monitor: val/map_50
 
-data: ../_base_/data/rotated_detection.yaml
+data: ../_base_/data/instance_segmentation.yaml
 overrides:
+  task: ROTATED_DETECTION
   max_epochs: 100
   data:
     train_subset:
       batch_size: 4
+      num_workers: 8
       transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            size_divisor: 32
         - class_path: torchvision.transforms.v2.Normalize
           init_args:
             std: [1.0, 1.0, 1.0]
       sampler:
         class_path: otx.algo.samplers.balanced_sampler.BalancedSampler
 
     val_subset:
-      batch_size: 1
+      num_workers: 4
       transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            size_divisor: 32
         - class_path: torchvision.transforms.v2.Normalize
           init_args:
             std: [1.0, 1.0, 1.0]
 
     test_subset:
-      batch_size: 1
+      num_workers: 4
       transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            size_divisor: 32
         - class_path: torchvision.transforms.v2.Normalize
           init_args:
             std: [1.0, 1.0, 1.0]

src/otx/recipe/rotated_detection/maskrcnn_efficientnetb2b_tile.yaml

@@ -0,0 +1,89 @@
+model:
+  class_path: otx.core.model.rotated_detection.RotatedMaskRCNNEfficientNet
+  init_args:
+    label_info: 80
+
+    optimizer:
+      class_path: torch.optim.SGD
+      init_args:
+        lr: 0.007
+        momentum: 0.9
+        weight_decay: 0.001
+
+    scheduler:
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
+      init_args:
+        num_warmup_steps: 100
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.1
+            patience: 9
+            monitor: val/map_50
+
+engine:
+  task: ROTATED_DETECTION
+  device: auto
+
+callback_monitor: val/map_50
+
+data: ../_base_/data/instance_segmentation.yaml
+overrides:
+  task: ROTATED_DETECTION
+  max_epochs: 100
+  data:
+    input_size:
+      - 512
+      - 512
+    tile_config:
+      enable_tiler: true
+      enable_adaptive_tiling: true
+
+    train_subset:
+      batch_size: 4
+      num_workers: 8
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Resize
+          init_args:
+            keep_ratio: false
+            scale: $(input_size)
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            pad_to_square: false
+            size_divisor: 32
+        - class_path: torchvision.transforms.v2.Normalize
+          init_args:
+            std: [1.0, 1.0, 1.0]
+      sampler:
+        class_path: otx.algo.samplers.balanced_sampler.BalancedSampler
+
+    val_subset:
+      num_workers: 4
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Resize
+          init_args:
+            keep_ratio: false
+            scale: $(input_size)
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            pad_to_square: false
+            size_divisor: 32
+        - class_path: torchvision.transforms.v2.Normalize
+          init_args:
+            std: [1.0, 1.0, 1.0]
+
+    test_subset:
+      num_workers: 4
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Resize
+          init_args:
+            keep_ratio: false
+            scale: $(input_size)
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            pad_to_square: false
+            size_divisor: 32
+        - class_path: torchvision.transforms.v2.Normalize
+          init_args:
+            std: [1.0, 1.0, 1.0]

src/otx/recipe/rotated_detection/maskrcnn_r50.yaml

@@ -1,5 +1,5 @@
 model:
-  class_path: otx.algo.instance_segmentation.maskrcnn.MaskRCNNResNet50
+  class_path: otx.core.model.rotated_detection.RotatedMaskRCNNResNet50
   init_args:
     label_info: 80
 
@@ -19,7 +19,7 @@ model:
           init_args:
             mode: max
             factor: 0.1
-            patience: 9
+            patience: 4
             monitor: val/map_50
 
 engine:
@@ -28,15 +28,30 @@ engine:
 
 callback_monitor: val/map_50
 
-data: ../_base_/data/rotated_detection.yaml
+data: ../_base_/data/instance_segmentation.yaml
 overrides:
+  task: ROTATED_DETECTION
   max_epochs: 100
+  gradient_clip_val: 35.0
   data:
     train_subset:
       batch_size: 4
+      num_workers: 8
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            size_divisor: 32
 
     val_subset:
-      batch_size: 1
+      num_workers: 4
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            size_divisor: 32
 
     test_subset:
-      batch_size: 1
+      num_workers: 4
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            size_divisor: 32

src/otx/recipe/rotated_detection/maskrcnn_r50_tile.yaml

@@ -0,0 +1,79 @@
+model:
+  class_path: otx.core.model.rotated_detection.RotatedMaskRCNNResNet50
+  init_args:
+    label_info: 80
+
+    optimizer:
+      class_path: torch.optim.SGD
+      init_args:
+        lr: 0.007
+        momentum: 0.9
+        weight_decay: 0.001
+
+    scheduler:
+      class_path: otx.core.schedulers.LinearWarmupSchedulerCallable
+      init_args:
+        num_warmup_steps: 100
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.1
+            patience: 4
+            monitor: val/map_50
+
+engine:
+  task: ROTATED_DETECTION
+  device: auto
+
+callback_monitor: val/map_50
+
+data: ../_base_/data/instance_segmentation.yaml
+overrides:
+  task: ROTATED_DETECTION
+  max_epochs: 100
+  gradient_clip_val: 35.0
+  data:
+    input_size:
+      - 512
+      - 512
+    tile_config:
+      enable_tiler: true
+      enable_adaptive_tiling: true
+
+    train_subset:
+      batch_size: 4
+      num_workers: 8
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Resize
+          init_args:
+            keep_ratio: false
+            scale: $(input_size)
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            pad_to_square: false
+            size_divisor: 32
+
+    val_subset:
+      num_workers: 4
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Resize
+          init_args:
+            keep_ratio: false
+            scale: $(input_size)
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            pad_to_square: false
+            size_divisor: 32
+
+    test_subset:
+      num_workers: 4
+      transforms:
+        - class_path: otx.core.data.transform_libs.torchvision.Resize
+          init_args:
+            keep_ratio: false
+            scale: $(input_size)
+        - class_path: otx.core.data.transform_libs.torchvision.Pad
+          init_args:
+            pad_to_square: false
+            size_divisor: 32