Merge branch 'releases/1.6.0' into ashwin/anomaly_geti_task_migration…

…_1_6

CHANGELOG.md

@@ -2,11 +2,21 @@
 
 All notable changes to this project will be documented in this file.
 
+## \[v1.6.5\]
+
+### Bug fixes
+
+- Fix labels names in hierarchical config (<https://github.com/openvinotoolkit/training_extensions/pull/3879>)
+- Fix Learning Rate and Loss Handling in Tile Classifier MaskRCNN EfficientNet (<https://github.com/openvinotoolkit/training_extensions/pull/3873>)
+- Disable Tile Classifier in Rotated Detection (<https://github.com/openvinotoolkit/training_extensions/pull/3894>)
+- Enhance Memeory Cache Handler with Complex Unique Keys (<https://github.com/openvinotoolkit/training_extensions/pull/3897>)
+
 ## \[v1.6.4\]
 
 ### Bug fixes
 
 - Pad input image to square when tiling is enabled (<https://github.com/openvinotoolkit/training_extensions/pull/3844>)
+- Fix confusion matrix calculation (<https://github.com/openvinotoolkit/training_extensions/pull/3857>)
 
 ## \[v1.6.3\]
 

README.md

@@ -98,11 +98,11 @@ You can find more details with examples in the [CLI command intro](https://openv
 
 ## Updates
 
-### v1.6.3 (3Q24)
+### v1.6.5 (3Q24)
 
-- Improve Vitens Kiemgetal OTX 1.6.x
-- Fix rotated detection data pipeline
-- Update dependencies to omit vulnerabilities
+- Fix labels names in hierarchical config
+- Fix Learning Rate and Loss Handling in Tile Classifier MaskRCNN EfficientNet
+- Enhance Memeory Cache Handler with Complex Unique Keys
 
 ### Release History
 

docs/source/guide/release_notes/index.rst

@@ -4,6 +4,19 @@ Releases
 .. toctree::
    :maxdepth: 1
 
+v1.6.5 (3Q24)
+-------------
+
+- Fix labels names in hierarchical config
+- Fix Learning Rate and Loss Handling in Tile Classifier MaskRCNN EfficientNet
+- Enhance Memeory Cache Handler with Complex Unique Keys
+
+v1.6.4 (3Q24)
+-------------
+
+- Pad input image to square when tiling is enabled
+- Fix confusion matrix calculation
+
 v1.6.3 (3Q24)
 -------------
 

src/otx/__init__.py

@@ -3,5 +3,5 @@
 # Copyright (C) 2021-2023 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
-__version__ = "1.6.4rc0"
+__version__ = "1.6.5"
 # NOTE: Sync w/ src/otx/api/usecases/exportable_code/demo/requirements.txt on release

src/otx/algorithms/classification/adapters/openvino/task.py

@@ -150,6 +150,7 @@ def _async_callback(self, request: Any, callback_args: tuple) -> None:
             result_handler(id, annotation, aux_data)
 
         except Exception as e:
+            logger.exception(e)
             self.callback_exceptions.append(e)
 
     def predict(self, image: np.ndarray) -> Tuple[ClassificationResult, AnnotationSceneEntity]:
@@ -280,6 +281,9 @@ def add_prediction(id: int, predicted_scene: AnnotationSceneEntity, aux_data: tu
 
         self.inferencer.await_all()
 
+        if self.inferencer.callback_exceptions:
+            raise RuntimeError("Inference failed, check the exceptions log.")
+
         self._avg_time_per_image = total_time / len(dataset)
         logger.info(f"Avg time per image: {self._avg_time_per_image} secs")
         logger.info(f"Total time: {total_time} secs")

src/otx/algorithms/classification/utils/cls_utils.py

@@ -23,14 +23,20 @@
 from otx.api.entities.label import LabelEntity
 from otx.api.entities.label_schema import LabelSchemaEntity
 from otx.api.serialization.label_mapper import LabelSchemaMapper
+from otx.api.utils.labels_utils import get_normalized_label_name
 
 
-def get_multihead_class_info(label_schema: LabelSchemaEntity):  # pylint: disable=too-many-locals
+def get_multihead_class_info(
+    label_schema: LabelSchemaEntity, normalize_labels: bool = False
+):  # pylint: disable=too-many-locals
     """Get multihead info by label schema."""
     all_groups = label_schema.get_groups(include_empty=False)
     all_groups_str = []
     for g in all_groups:
-        group_labels_str = [lbl.name for lbl in g.labels]
+        if normalize_labels:
+            group_labels_str = [get_normalized_label_name(lbl) for lbl in g.labels]
+        else:
+            group_labels_str = [lbl.name for lbl in g.labels]
         all_groups_str.append(group_labels_str)
 
     single_label_groups = [g for g in all_groups_str if len(g) == 1]
@@ -76,7 +82,7 @@ def get_cls_inferencer_configuration(label_schema: LabelSchemaEntity):
     hierarchical = not multilabel and len(label_schema.get_groups(False)) > 1
     multihead_class_info = {}
     if hierarchical:
-        multihead_class_info = get_multihead_class_info(label_schema)
+        multihead_class_info = get_multihead_class_info(label_schema, normalize_labels=True)
     return {
         "multilabel": multilabel,
         "hierarchical": hierarchical,
@@ -112,17 +118,19 @@ def get_cls_model_api_configuration(label_schema: LabelSchemaEntity, inference_c
     all_labels = ""
     all_label_ids = ""
     for lbl in label_entities:
-        all_labels += lbl.name.replace(" ", "_") + " "
+        all_labels += get_normalized_label_name(lbl) + " "
         all_label_ids += f"{lbl.id_} "
 
     mapi_config[("model_info", "labels")] = all_labels.strip()
     mapi_config[("model_info", "label_ids")] = all_label_ids.strip()
 
     hierarchical_config = {}
-    hierarchical_config["cls_heads_info"] = get_multihead_class_info(label_schema)
+    hierarchical_config["cls_heads_info"] = get_multihead_class_info(label_schema, normalize_labels=True)
     hierarchical_config["label_tree_edges"] = []
     for edge in label_schema.label_tree.edges:  # (child, parent)
-        hierarchical_config["label_tree_edges"].append((edge[0].name, edge[1].name))
+        hierarchical_config["label_tree_edges"].append(
+            (get_normalized_label_name(edge[0]), get_normalized_label_name(edge[1]))
+        )
 
     mapi_config[("model_info", "hierarchical_config")] = json.dumps(hierarchical_config)
     return mapi_config
@@ -137,7 +145,7 @@ def get_hierarchical_label_list(hierarchical_cls_heads_info: Dict, labels: List)
     hierarchical_labels = []
     for label_str, _ in label_to_idx.items():
         for label_entity in labels:
-            if label_entity.name == label_str:
+            if get_normalized_label_name(label_entity) == label_str:
                 hierarchical_labels.append(label_entity)
                 break
     return hierarchical_labels

src/otx/algorithms/common/adapters/mmcv/pipelines/load_image_from_otx_dataset.py

@@ -42,7 +42,12 @@ def _get_unique_key(results: Dict[str, Any]) -> Tuple:
         if "cache_key" in results:
             return results["cache_key"]
         d_item = results["dataset_item"]
-        results["cache_key"] = d_item.media.path, d_item.roi.id
+        if d_item.media.path:  # when video extracted frames come, media.path is None
+            results["cache_key"] = d_item.media.path, d_item.roi.id
+        elif len(d_item.annotation_scene.annotations) > 0:
+            results["cache_key"] = d_item.roi.id, d_item.annotation_scene.annotations[0].id
+        else:
+            results["cache_key"] = d_item.roi.id
         return results["cache_key"]
 
     def _get_memcache_handler(self):

src/otx/algorithms/detection/adapters/mmdet/models/detectors/custom_maskrcnn_tile_optimized.py

@@ -149,6 +149,8 @@ def forward_train(
             img, img_metas, gt_bboxes, gt_labels, gt_bboxes_ignore, gt_masks, proposals, **kwargs
         )
         losses.update(rcnn_loss)
+        if "acc" in losses:
+            losses.pop("acc")
         return losses
 
     @staticmethod

src/otx/algorithms/detection/adapters/mmdet/utils/config_utils.py

@@ -122,11 +122,8 @@ def patch_tiling(config, hparams, dataset=None):
 
             if config.model.backbone.type == "efficientnet_b2b":
                 learning_rate = 0.002
-                logger.info(
-                    f"Patched {config.model.backbone.type} LR: "
-                    f"{hparams.learning_parameters.learning_rate} -> {learning_rate}"
-                )
-                hparams.learning_parameters.learning_rate = learning_rate
+                logger.info(f"Patched {config.model.backbone.type} LR: " f"{config.optimizer.lr} -> {learning_rate}")
+                config.optimizer.lr = learning_rate
 
             config.data.train.filter_empty_gt = False
 

src/otx/algorithms/detection/adapters/openvino/task.py

@@ -155,6 +155,7 @@ def _async_callback(self, request: Any, callback_args: tuple) -> None:
             result_handler(id, processed_prediciton, features)
 
         except Exception as e:
+            logger.exception(e)
             self.callback_exceptions.append(e)
 
     def enqueue_prediction(self, image: np.ndarray, id: int, result_handler: Any) -> None:
@@ -557,6 +558,9 @@ def add_prediction(id: int, predicted_scene: AnnotationSceneEntity, aux_data: tu
 
         self.inferencer.await_all()
 
+        if self.inferencer.callback_exceptions:
+            raise RuntimeError("Inference failed, check the exceptions log.")
+
         self._avg_time_per_image = total_time / len(dataset)
         logger.info(f"Avg time per image: {self._avg_time_per_image} secs")
         logger.info(f"Total time: {total_time} secs")

src/otx/algorithms/detection/configs/base/data/tiling/base_iseg_tile_pipeline.py

@@ -17,7 +17,7 @@
     dict(type="Resize", img_scale=img_size, keep_ratio=True),
     dict(type="RandomFlip", flip_ratio=0.5),
     dict(type="Normalize", **img_norm_cfg),
-    dict(type="Pad", pad_to_square=True),
+    dict(type="Pad", size=img_size),
     dict(type="DefaultFormatBundle"),
     dict(
         type="Collect",
@@ -45,7 +45,7 @@
             dict(type="Resize", keep_ratio=True),
             dict(type="RandomFlip"),
             dict(type="Normalize", **img_norm_cfg),
-            dict(type="Pad", pad_to_square=True),
+            dict(type="Pad", size=img_size),
             dict(type="ImageToTensor", keys=["img"]),
             dict(type="Collect", keys=["img"]),
         ],

src/otx/algorithms/detection/configs/base/data/tiling/efficientnet_iseg_tile_pipeline.py

@@ -17,7 +17,7 @@
     dict(type="Resize", img_scale=img_size, keep_ratio=True),
     dict(type="RandomFlip", flip_ratio=0.5),
     dict(type="Normalize", **img_norm_cfg),
-    dict(type="Pad", pad_to_square=True),
+    dict(type="Pad", size=img_size),
     dict(type="DefaultFormatBundle"),
     dict(
         type="Collect",
@@ -45,7 +45,7 @@
             dict(type="Resize", keep_ratio=True),
             dict(type="RandomFlip"),
             dict(type="Normalize", **img_norm_cfg),
-            dict(type="Pad", pad_to_square=True),
+            dict(type="Pad", size=img_size),
             dict(type="ImageToTensor", keys=["img"]),
             dict(type="Collect", keys=["img"]),
         ],

src/otx/algorithms/detection/configs/rotated_detection/configuration.yaml

@@ -565,16 +565,16 @@ tiling_parameters:
     header: Enable tile classifier
     description: Enabling tile classifier enhances the speed of tiling inference by incorporating a tile classifier into the instance segmentation model. This feature prevents the detector from making predictions on tiles that do not contain any objects, thus optimizing its speed performance.
     default_value: false
-    editable: true
+    editable: false
     affects_outcome_of: TRAINING
     type: BOOLEAN
     ui_rules:
       action: DISABLE_EDITING
       operator: AND
       rules: []
       type: UI_RULES
-    value: true
-    visible_in_ui: true
+    value: false
+    visible_in_ui: false
     warning: The tile classifier prioritizes inference speed over training speed, it requires more training in order to achieve its optimized performance.
 
   enable_adaptive_params:

src/otx/algorithms/detection/configs/rotated_detection/efficientnetb2b_maskrcnn/tile_pipeline.py

@@ -17,7 +17,7 @@
     dict(type="Resize", img_scale=img_size, keep_ratio=False),
     dict(type="RandomFlip", flip_ratio=0.5),
     dict(type="Normalize", **img_norm_cfg),
-    dict(type="Pad", pad_to_square=True),
+    dict(type="Pad", size=img_size),
     dict(type="DefaultFormatBundle"),
     dict(
         type="Collect",
@@ -45,7 +45,7 @@
             dict(type="Resize", keep_ratio=False),
             dict(type="RandomFlip"),
             dict(type="Normalize", **img_norm_cfg),
-            dict(type="Pad", pad_to_square=True),
+            dict(type="Pad", size=img_size),
             dict(type="ImageToTensor", keys=["img"]),
             dict(type="Collect", keys=["img"]),
         ],

src/otx/algorithms/detection/configs/rotated_detection/resnet50_maskrcnn/tile_pipeline.py

@@ -17,7 +17,7 @@
     dict(type="Resize", img_scale=img_size, keep_ratio=True),
     dict(type="RandomFlip", flip_ratio=0.5),
     dict(type="Normalize", **img_norm_cfg),
-    dict(type="Pad", pad_to_square=True),
+    dict(type="Pad", size=img_size),
     dict(type="DefaultFormatBundle"),
     dict(
         type="Collect",
@@ -45,7 +45,7 @@
             dict(type="Resize", keep_ratio=True),
             dict(type="RandomFlip"),
             dict(type="Normalize", **img_norm_cfg),
-            dict(type="Pad", pad_to_square=True),
+            dict(type="Pad", size=img_size),
             dict(type="ImageToTensor", keys=["img"]),
             dict(type="Collect", keys=["img"]),
         ],

src/otx/algorithms/segmentation/adapters/openvino/task.py

@@ -151,6 +151,7 @@ def _async_callback(self, request: Any, callback_args: tuple) -> None:
             result_handler(id, annotation, processed_prediciton.feature_vector, processed_prediciton.saliency_map)
 
         except Exception as e:
+            logger.exception(e)
             self.callback_exceptions.append(e)
 
 
@@ -254,6 +255,9 @@ def add_prediction(
 
         self.inferencer.await_all()
 
+        if self.inferencer.callback_exceptions:
+            raise RuntimeError("Inference failed, check the exceptions log.")
+
         self._avg_time_per_image = total_time / len(dataset)
         logger.info(f"Avg time per image: {self._avg_time_per_image} secs")
         logger.info(f"Total time: {total_time} secs")

src/otx/api/usecases/exportable_code/demo/requirements.txt

@@ -1,4 +1,4 @@
 openvino==2023.3.0
 openvino-model-api==0.1.9.1
-otx==1.6.3
+otx==1.6.5
 numpy>=1.21.0,<=1.23.5  # np.bool was removed in 1.24.0 which was used in openvino runtime

src/otx/api/utils/labels_utils.py

@@ -18,3 +18,8 @@ def get_empty_label(label_schema: LabelSchemaEntity) -> Optional[LabelEntity]:
     if empty_candidates:
         return empty_candidates[0]
     return None
+
+
+def get_normalized_label_name(label: LabelEntity) -> str:
+    """Gets a nomalized label name"""
+    return label.name.replace(" ", "_")

tests/unit/algorithms/detection/tiling/test_tiling_tile_classifier.py

@@ -211,3 +211,4 @@ def test_patch_tiling_func(self):
         hyper_parameters.tiling_parameters.enable_tiling = True
         hyper_parameters.tiling_parameters.enable_tile_classifier = True
         patch_tiling(cfg, hyper_parameters, self.dataset)
+        assert cfg.optimizer.lr == 0.002, "Learning rate should be 0.002 when MRCNN EfficientNet is used"