Merge remote-tracking branch 'origin/develop' into vs/fix_yolo_bfp16

Author: sovrasov

src/otx/algorithms/classification/configs/base/data/data_pipeline.py

@@ -1,18 +1,7 @@
 """Data Pipeline of Class-Incr model for Classification Task."""
 
-# Copyright (C) 2022 Intel Corporation
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions
-# and limitations under the License.
+# Copyright (C) 2022-2023 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
 
 # pylint: disable=invalid-name
 

src/otx/algorithms/classification/configs/base/data/selfsl/data_pipeline.py

@@ -1,26 +1,21 @@
 """Data Pipeline of Self-SL model for Classification Task."""
 
-# Copyright (C) 2022 Intel Corporation
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions
-# and limitations under the License.
+# Copyright (C) 2022-2023 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
 
 # pylint: disable=invalid-name
 
 __img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
 __img_size = 224
 
 __train_pipeline_v0 = [
-    dict(type="LoadImageFromOTXDataset"),
+    dict(
+        type="LoadResizeDataFromOTXDataset",
+        resize_cfg=dict(type="Resize", size=__img_size, downscale_only=True),
+        # To be resized in this op only if input is larger than expected size
+        # for speed & cache memory efficiency.
+        enable_memcache=True,  # Cache after resizing image
+    ),
     dict(type="RandomResizedCrop", size=__img_size),
     dict(type="RandomFlip"),
     dict(
@@ -37,7 +32,13 @@
     dict(type="Collect", keys=["img"]),
 ]
 __train_pipeline_v1 = [
-    dict(type="LoadImageFromOTXDataset"),
+    dict(
+        type="LoadResizeDataFromOTXDataset",
+        resize_cfg=dict(type="Resize", size=__img_size, downscale_only=True),
+        # To be resized in this op only if input is larger than expected size
+        # for speed & cache memory efficiency.
+        enable_memcache=True,  # Cache after resizing image
+    ),
     dict(type="RandomResizedCrop", size=__img_size),
     dict(type="RandomFlip"),
     dict(

src/otx/algorithms/classification/configs/base/data/semisl/data_pipeline.py

@@ -18,8 +18,11 @@
 __resize_target_size = 224
 
 __common_pipeline = [
-    dict(type="LoadImageFromOTXDataset"),
-    dict(type="Resize", size=__resize_target_size),
+    dict(
+        type="LoadResizeDataFromOTXDataset",
+        resize_cfg=dict(type="Resize", size=__resize_target_size, downscale_only=False),
+        enable_memcache=True,  # Cache after resizing image
+    ),
     dict(type="RandomFlip", flip_prob=0.5, direction="horizontal"),
     dict(type="AugMixAugment", config_str="augmix-m5-w3"),
     dict(type="RandomRotate", p=0.35, angle=(-10, 10)),
@@ -48,6 +51,17 @@
     dict(type="Collect", keys=["img", "img_strong"]),
 ]
 
+__val_pipeline = [
+    dict(
+        type="LoadResizeDataFromOTXDataset",
+        resize_cfg=dict(type="Resize", size=__resize_target_size, downscale_only=False),
+        enable_memcache=True,  # Cache after resizing image
+    ),
+    dict(type="Normalize", **__img_norm_cfg),
+    dict(type="ImageToTensor", keys=["img"]),
+    dict(type="Collect", keys=["img"]),
+]
+
 __test_pipeline = [
     dict(type="LoadImageFromOTXDataset"),
     dict(type="Resize", size=__resize_target_size),
@@ -64,6 +78,6 @@
         type=__dataset_type,
         pipeline=__unlabeled_pipeline,
     ),
-    val=dict(type=__dataset_type, test_mode=True, pipeline=__test_pipeline),
+    val=dict(type=__dataset_type, test_mode=True, pipeline=__val_pipeline),
     test=dict(type=__dataset_type, test_mode=True, pipeline=__test_pipeline),
 )

src/otx/algorithms/classification/configs/base/data/supcon/data_pipeline.py

@@ -1,18 +1,7 @@
 """Data Pipeline of SupCon model for Classification Task."""
 
 # Copyright (C) 2023 Intel Corporation
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions
-# and limitations under the License.
+# SPDX-License-Identifier: Apache-2.0
 
 # pylint: disable=invalid-name
 
@@ -21,11 +10,14 @@
 
 
 __train_pipeline = [
-    dict(type="LoadImageFromOTXDataset"),
+    dict(
+        type="LoadResizeDataFromOTXDataset",
+        resize_cfg=dict(type="Resize", size=__resize_target_size, downscale_only=False),
+        enable_memcache=True,  # Cache after resizing image
+    ),
     dict(
         type="TwoCropTransform",
         pipeline=[
-            dict(type="Resize", size=__resize_target_size),
             dict(type="RandomFlip", flip_prob=0.5, direction="horizontal"),
             dict(type="AugMixAugment", config_str="augmix-m5-w3"),
             dict(type="RandomRotate", p=0.35, angle=(-10, 10)),
@@ -38,6 +30,17 @@
     ),
 ]
 
+__val_pipeline = [
+    dict(
+        type="LoadResizeDataFromOTXDataset",
+        resize_cfg=dict(type="Resize", size=__resize_target_size, downscale_only=False),
+        enable_memcache=True,  # Cache after resizing image
+    ),
+    dict(type="Normalize", **__img_norm_cfg),
+    dict(type="ImageToTensor", keys=["img"]),
+    dict(type="Collect", keys=["img"]),
+]
+
 __test_pipeline = [
     dict(type="LoadImageFromOTXDataset"),
     dict(type="Resize", size=__resize_target_size),
@@ -50,6 +53,6 @@
 
 data = dict(
     train=dict(type=__dataset_type, pipeline=__train_pipeline),
-    val=dict(type=__dataset_type, test_mode=True, pipeline=__test_pipeline),
+    val=dict(type=__dataset_type, test_mode=True, pipeline=__val_pipeline),
     test=dict(type=__dataset_type, test_mode=True, pipeline=__test_pipeline),
 )

src/otx/algorithms/common/adapters/mmcv/hooks/dual_model_ema_hook.py

@@ -71,16 +71,17 @@ def before_run(self, runner):
         self.src_model = getattr(model, self.src_model_name, None)
         self.dst_model = getattr(model, self.dst_model_name, None)
         if self.src_model and self.dst_model:
+            self.enabled = True
             self.src_params = self.src_model.state_dict(keep_vars=True)
             self.dst_params = self.dst_model.state_dict(keep_vars=True)
-
-    def before_train_epoch(self, runner):
-        """Momentum update."""
-        if runner.epoch + 1 == self.start_epoch:
-            self._copy_model()
-            self.enabled = True
-
-        if self.epoch_momentum > 0.0 and self.enabled:
+            if runner.epoch == 0 and runner.iter == 0:
+                self._copy_model(sync_model=True)
+                logger.info("Initialized student model by teacher model")
+                logger.info(f"model_s model_t diff: {self._diff_model()}")
+
+    def before_epoch(self, runner):
+        """Compute adaptive EMA momentum."""
+        if self.epoch_momentum > 0.0:
             iter_per_epoch = len(runner.data_loader)
             epoch_decay = 1 - self.epoch_momentum
             iter_decay = math.pow(epoch_decay, self.interval / iter_per_epoch)
@@ -91,6 +92,12 @@ def before_train_epoch(self, runner):
     def after_train_iter(self, runner):
         """Update ema parameter every self.interval iterations."""
         if not self.enabled or (runner.iter % self.interval != 0):
+            # Skip update
+            return
+
+        if runner.epoch + 1 < self.start_epoch:
+            # Just copy parameters before start epoch
+            self._copy_model()
             return
 
         # EMA
@@ -107,12 +114,15 @@ def _get_model(self, runner):
             model = model.module
         return model
 
-    def _copy_model(self):
+    def _copy_model(self, sync_model=False):
         with torch.no_grad():
             for name, src_param in self.src_params.items():
                 if not name.startswith("ema_"):
                     dst_param = self.dst_params[name]
-                    dst_param.data.copy_(src_param.data)
+                    if sync_model:
+                        src_param.data.copy_(dst_param.data)
+                    else:
+                        dst_param.data.copy_(src_param.data)
 
     def _ema_model(self):
         momentum = min(self.momentum, 1.0)

src/otx/algorithms/common/adapters/mmcv/hooks/mean_teacher_hook.py

@@ -34,7 +34,7 @@ def before_train_epoch(self, runner):
         logger.info(f"avr_ps_ratio: {average_pseudo_label_ratio}")
         self._get_model(runner).enable_unlabeled_loss(True)
         self.unlabeled_loss_enabled = True
-        logger.info("---------- Enabled unlabeled loss and EMA smoothing")
+        logger.info("---------- Enabled unlabeled loss and EMA smoothing ----------")
 
     def after_train_iter(self, runner):
         """Update ema parameter every self.interval iterations."""
@@ -44,6 +44,7 @@ def after_train_iter(self, runner):
 
         if runner.epoch + 1 < self.start_epoch or self.unlabeled_loss_enabled is False:
             # Just copy parameters before enabled
+            self._copy_model()
             return
 
         # EMA