Cherry-pick from 2.3 (#4242)

* Bump MAPI to 0.2.5.1 (#4211)

* Bump MAPI to 0.2.5.1

* Updatge changelog

* Update changelog

* Revert h-cls head to linear (#4221)

* Revert h-cls head to linear

* Update changelog

* Update third-party-programs.txt (#4225)

update third-party-programs.txt

---------

Co-authored-by: Prokofiev Kirill <[email protected]>

Author: sovrasov

CHANGELOG.md

@@ -75,6 +75,8 @@ All notable changes to this project will be documented in this file.
   (<https://github.com/openvinotoolkit/training_extensions/pull/3985>)
 - Bump OV to 2024.6, update empty label handling
   (<https://github.com/openvinotoolkit/training_extensions/pull/4162>)
+- Bump MAPI to 0.2.5.1
+  (<https://github.com/openvinotoolkit/training_extensions/pull/4211>)
 
 ### Bug fixes
 
@@ -96,6 +98,8 @@ All notable changes to this project will be documented in this file.
   (<https://github.com/openvinotoolkit/training_extensions/pull/4199>)
 - Fix label info on loading checkpoint
   (<https://github.com/openvinotoolkit/training_extensions/pull/4200>)
+- Revert h-cls head to linear one
+  (<https://github.com/openvinotoolkit/training_extensions/pull/4221>)
 
 ## \[2.2.2\]
 

pyproject.toml

@@ -81,7 +81,7 @@ base = [
     "timm==1.0.3",
     "openvino==2024.6",
     "openvino-dev==2024.6",
-    "openvino-model-api==0.2.5",
+    "openvino-model-api==0.2.5.1",
     "onnx==1.17.0",
     "onnxconverter-common==1.14.0",
     "nncf==2.14.1",

src/otx/algo/classification/efficientnet.py

@@ -13,7 +13,7 @@
 
 from otx.algo.classification.backbones.efficientnet import EFFICIENTNET_VERSION, EfficientNetBackbone
 from otx.algo.classification.classifier import HLabelClassifier, ImageClassifier
-from otx.algo.classification.heads import HierarchicalCBAMClsHead, LinearClsHead, MultiLabelLinearClsHead
+from otx.algo.classification.heads import HierarchicalLinearClsHead, LinearClsHead, MultiLabelLinearClsHead
 from otx.algo.classification.losses.asymmetric_angular_loss_with_ignore import AsymmetricAngularLossWithIgnore
 from otx.algo.classification.necks.gap import GlobalAveragePooling
 from otx.algo.classification.utils import get_classification_layers
@@ -225,11 +225,8 @@ def _build_model(self, head_config: dict) -> nn.Module:
 
         return HLabelClassifier(
             backbone=backbone,
-            neck=nn.Identity(),
-            head=HierarchicalCBAMClsHead(
-                in_channels=backbone.num_features,
-                **copied_head_config,
-            ),
+            neck=GlobalAveragePooling(dim=2),
+            head=HierarchicalLinearClsHead(**copied_head_config, in_channels=backbone.num_features),
             multiclass_loss=nn.CrossEntropyLoss(),
             multilabel_loss=AsymmetricAngularLossWithIgnore(gamma_pos=0.0, gamma_neg=1.0, reduction="sum"),
         )

src/otx/algo/classification/mobilenet_v3.py

@@ -14,7 +14,7 @@
 
 from otx.algo.classification.backbones import MobileNetV3Backbone
 from otx.algo.classification.classifier import HLabelClassifier, ImageClassifier
-from otx.algo.classification.heads import HierarchicalCBAMClsHead, LinearClsHead, MultiLabelNonLinearClsHead
+from otx.algo.classification.heads import HierarchicalLinearClsHead, LinearClsHead, MultiLabelNonLinearClsHead
 from otx.algo.classification.losses.asymmetric_angular_loss_with_ignore import AsymmetricAngularLossWithIgnore
 from otx.algo.classification.necks.gap import GlobalAveragePooling
 from otx.algo.utils.support_otx_v1 import OTXv1Helper
@@ -250,15 +250,13 @@ def _build_model(self, head_config: dict) -> nn.Module:
 
         copied_head_config = copy(head_config)
         copied_head_config["step_size"] = (ceil(self.input_size[0] / 32), ceil(self.input_size[1] / 32))
-
+        in_channels = MobileNetV3Backbone.MV3_CFG[self.mode]["out_channels"]
         backbone = MobileNetV3Backbone(mode=self.mode, input_size=self.input_size)
+
         return HLabelClassifier(
             backbone=backbone,
-            neck=nn.Identity(),
-            head=HierarchicalCBAMClsHead(
-                in_channels=MobileNetV3Backbone.MV3_CFG[self.mode]["out_channels"],
-                **copied_head_config,
-            ),
+            neck=GlobalAveragePooling(dim=2),
+            head=HierarchicalLinearClsHead(**copied_head_config, in_channels=in_channels),
             multiclass_loss=nn.CrossEntropyLoss(),
             multilabel_loss=AsymmetricAngularLossWithIgnore(gamma_pos=0.0, gamma_neg=1.0, reduction="sum"),
         )

src/otx/algo/classification/timm_model.py

@@ -14,7 +14,7 @@
 
 from otx.algo.classification.backbones.timm import TimmBackbone
 from otx.algo.classification.classifier import HLabelClassifier, ImageClassifier
-from otx.algo.classification.heads import HierarchicalCBAMClsHead, LinearClsHead, MultiLabelLinearClsHead
+from otx.algo.classification.heads import HierarchicalLinearClsHead, LinearClsHead, MultiLabelLinearClsHead
 from otx.algo.classification.losses.asymmetric_angular_loss_with_ignore import AsymmetricAngularLossWithIgnore
 from otx.algo.classification.necks.gap import GlobalAveragePooling
 from otx.algo.utils.support_otx_v1 import OTXv1Helper
@@ -283,11 +283,8 @@ def _build_model(self, head_config: dict) -> nn.Module:
         copied_head_config["step_size"] = (ceil(self.input_size[0] / 32), ceil(self.input_size[1] / 32))
         return HLabelClassifier(
             backbone=backbone,
-            neck=nn.Identity(),
-            head=HierarchicalCBAMClsHead(
-                in_channels=backbone.num_features,
-                **copied_head_config,
-            ),
+            neck=GlobalAveragePooling(dim=2),
+            head=HierarchicalLinearClsHead(**copied_head_config, in_channels=backbone.num_features),
             multiclass_loss=nn.CrossEntropyLoss(),
             multilabel_loss=AsymmetricAngularLossWithIgnore(gamma_pos=0.0, gamma_neg=1.0, reduction="sum"),
         )

src/otx/algo/classification/torchvision_model.py

@@ -13,7 +13,7 @@
 from otx.algo.classification.backbones.torchvision import TorchvisionBackbone, TVModelType
 from otx.algo.classification.classifier import HLabelClassifier, ImageClassifier
 from otx.algo.classification.heads import (
-    HierarchicalCBAMClsHead,
+    HierarchicalLinearClsHead,
     LinearClsHead,
     MultiLabelLinearClsHead,
 )
@@ -252,11 +252,8 @@ def _build_model(self, head_config: dict) -> nn.Module:
         backbone = TorchvisionBackbone(backbone=self.backbone, pretrained=self.pretrained)
         return HLabelClassifier(
             backbone=backbone,
-            neck=nn.Identity(),
-            head=HierarchicalCBAMClsHead(
-                in_channels=backbone.in_features,
-                **head_config,
-            ),
+            neck=GlobalAveragePooling(dim=2),
+            head=HierarchicalLinearClsHead(**head_config, in_channels=backbone.in_features),
             multiclass_loss=nn.CrossEntropyLoss(),
             multilabel_loss=AsymmetricAngularLossWithIgnore(gamma_pos=0.0, gamma_neg=1.0, reduction="sum"),
         )

src/otx/algo/classification/vit.py

@@ -19,7 +19,7 @@
 from otx.algo.classification.backbones.vision_transformer import VIT_ARCH_TYPE, VisionTransformer
 from otx.algo.classification.classifier import HLabelClassifier, ImageClassifier
 from otx.algo.classification.heads import (
-    HierarchicalCBAMClsHead,
+    HierarchicalLinearClsHead,
     MultiLabelLinearClsHead,
     VisionTransformerClsHead,
 )
@@ -452,11 +452,7 @@ def _build_model(self, head_config: dict) -> nn.Module:
         return HLabelClassifier(
             backbone=vit_backbone,
             neck=None,
-            head=HierarchicalCBAMClsHead(
-                in_channels=vit_backbone.embed_dim,
-                step_size=1,
-                **head_config,
-            ),
+            head=HierarchicalLinearClsHead(**head_config, in_channels=vit_backbone.embed_dim),
             multiclass_loss=nn.CrossEntropyLoss(),
             multilabel_loss=AsymmetricAngularLossWithIgnore(gamma_pos=0.0, gamma_neg=1.0, reduction="sum"),
             init_cfg=init_cfg,