[OTE][Release][XAI] Detection fix two stage bbox_head error

.gitignore

@@ -8,6 +8,7 @@ __pycache__
 env
 .env
 .tox
+results/
 
 data/*
 .coverage

external/model-preparation-algorithm/configs/detection/mobilenetv2_ssd_cls_incr/data_pipeline.py

@@ -37,10 +37,10 @@
     samples_per_gpu=10,
     workers_per_gpu=4,
     train=dict(
-          type=dataset_type,
-          ann_file="data/coco/annotations/instances_train2017.json",
-          img_prefix="data/coco/train2017",
-          pipeline=train_pipeline,
+        type=dataset_type,
+        ann_file="data/coco/annotations/instances_train2017.json",
+        img_prefix="data/coco/train2017",
+        pipeline=train_pipeline,
     ),
     val=dict(
         type=dataset_type,

external/model-preparation-algorithm/mpa_tasks/apis/detection/task.py

@@ -309,15 +309,13 @@ def _add_predictions_to_dataset(self, prediction_results, dataset, confidence_th
             if saliency_map is not None:
                 if saliency_map.ndim == 2:
                     # Single saliency map per image, support e.g. EigenCAM use case
-                    actmap = get_actmap(
-                        saliency_map, (dataset_item.width, dataset_item.height)
-                    )
+                    actmap = get_actmap(saliency_map, (dataset_item.width, dataset_item.height))
                     saliency_media = ResultMediaEntity(
                         name="Saliency Map",
                         type="saliency_map",
                         annotation_scene=dataset_item.annotation_scene,
                         numpy=actmap,
-                        roi=dataset_item.roi
+                        roi=dataset_item.roi,
                     )
                     dataset_item.append_metadata_item(saliency_media, model=self._task_environment.model)
                 elif saliency_map.ndim == 3:
@@ -326,22 +324,22 @@ def _add_predictions_to_dataset(self, prediction_results, dataset, confidence_th
                     num_saliency_maps = saliency_map.shape[0]
                     if num_saliency_maps == len(labels) + 1:
                         # Include the background as the last category
-                        labels.append(LabelEntity('background', Domain.DETECTION))
+                        labels.append(LabelEntity("background", Domain.DETECTION))
                     for class_id, class_wise_saliency_map in enumerate(saliency_map):
-                        actmap = get_actmap(
-                            class_wise_saliency_map, (dataset_item.width, dataset_item.height)
-                        )
+                        actmap = get_actmap(class_wise_saliency_map, (dataset_item.width, dataset_item.height))
                         class_name_str = labels[class_id].name
                         saliency_media = ResultMediaEntity(
                             name=class_name_str,
                             type="saliency_map",
                             annotation_scene=dataset_item.annotation_scene,
-                            numpy=actmap, roi=dataset_item.roi
+                            numpy=actmap,
+                            roi=dataset_item.roi,
                         )
                         dataset_item.append_metadata_item(saliency_media, model=self._task_environment.model)
                 else:
-                    raise RuntimeError(f'Single saliency map has to be 2 or 3-dimensional, '
-                                       f'but got {saliency_map.ndim} dims')
+                    raise RuntimeError(
+                        f"Single saliency map has to be 2 or 3-dimensional, but got {saliency_map.ndim} dims"
+                    )
 
     def _patch_data_pipeline(self):
         base_dir = os.path.abspath(os.path.dirname(self.template_file_path))
@@ -380,9 +378,7 @@ def patch_data_pipeline(cfg):
                         pipeline_step.type = "LoadAnnotationFromOTEDataset"
                         pipeline_step.domain = domain
                         pipeline_step.min_size = cfg.pop("min_size", -1)
-                    if (subset == "train" and
-                        pipeline_step.type == "Collect" and
-                        cfg.type not in ["ImageTilingDataset"]):
+                    if subset == "train" and pipeline_step.type == "Collect" and cfg.type not in ["ImageTilingDataset"]:
                         pipeline_step = BaseTask._get_meta_keys(pipeline_step)
                 patch_color_conversion(cfg.pipeline)
 

external/model-preparation-algorithm/mpa_tasks/extensions/datasets/mpa_cls_dataset.py

@@ -131,9 +131,9 @@ def evaluate(self, results, metric="accuracy", metric_options=None, logger=None)
 
         # compute top-k metrics from mmcls and align them in [0,1] range
         eval_results = super().evaluate(results, metrics, metric_options, logger)
-        for k in metric_options['topk']:
-            eval_results[f'accuracy_top-{k}'] /= 100
-            assert 0 <= eval_results[f'accuracy_top-{k}'] <= 1
+        for k in metric_options["topk"]:
+            eval_results[f"accuracy_top-{k}"] /= 100
+            assert 0 <= eval_results[f"accuracy_top-{k}"] <= 1
 
         # Add Evaluation Accuracy score per Class - it can be used only for multi-class dataset.
         if self.class_acc:

external/model-preparation-algorithm/tests/test_xai.py

@@ -8,8 +8,8 @@
 from mmcls.models import build_classifier
 from mmdet.models import build_detector
 
+from mpa.det.stage import DetectionStage  # noqa
 from mpa.modules.hooks.auxiliary_hooks import ReciproCAMHook, DetSaliencyMapHook
-from mpa.det.stage import DetectionStage
 
 
 torch.manual_seed(0)
@@ -37,31 +37,31 @@ def get_classification_model():
     @staticmethod
     def get_detection_model():
         model_cfg = dict(
-            type='CustomSingleStageDetector',
-            backbone=dict(type='mobilenetv2_w1',
-                          out_indices=(4, 5),
-                          frozen_stages=-1,
-                          norm_eval=False,
-                          pretrained=True),
+            type="CustomSingleStageDetector",
+            backbone=dict(
+                type="mobilenetv2_w1", out_indices=(4, 5), frozen_stages=-1, norm_eval=False, pretrained=True
+            ),
             bbox_head=dict(
-                type='CustomSSDHead',
+                type="CustomSSDHead",
                 in_channels=(96, 320),
                 num_classes=20,
                 anchor_generator=dict(
-                    type='SSDAnchorGeneratorClustered',
+                    type="SSDAnchorGeneratorClustered",
                     reclustering_anchors=True,
                     strides=[16, 32],
-                    widths=[np.array([70.93408016, 132.06659281, 189.56180207, 349.90057837]),
-                            np.array([272.31733885, 448.52200666, 740.63350023, 530.78990182,
-                                      790.99297377])],
-                    heights=[np.array([93.83759764, 235.21261441, 432.6029996, 250.08979657]),
-                             np.array([672.8829653, 474.84783528, 420.18291446, 741.02592293,
-                                       766.45636125])]),
+                    widths=[
+                        np.array([70.93408016, 132.06659281, 189.56180207, 349.90057837]),
+                        np.array([272.31733885, 448.52200666, 740.63350023, 530.78990182, 790.99297377]),
+                    ],
+                    heights=[
+                        np.array([93.83759764, 235.21261441, 432.6029996, 250.08979657]),
+                        np.array([672.8829653, 474.84783528, 420.18291446, 741.02592293, 766.45636125]),
+                    ],
+                ),
                 bbox_coder=dict(
-                    type='DeltaXYWHBBoxCoder',
-                    target_means=[.0, .0, .0, .0],
-                    target_stds=[0.1, 0.1, 0.2, 0.2]),
-            )
+                    type="DeltaXYWHBBoxCoder", target_means=[0.0, 0.0, 0.0, 0.0], target_stds=[0.1, 0.1, 0.2, 0.2]
+                ),
+            ),
         )
         model = build_detector(model_cfg)
         return model.eval()

ote_sdk/ote_sdk/utils/tiler.py

@@ -6,6 +6,7 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
+import copy
 from itertools import product
 from typing import Any, List, Tuple, Union
 
@@ -79,7 +80,7 @@ def predict(self, image: np.ndarray):
             detections = np.append(detections, output, axis=0)
             # cache full image feature vector and saliency map at 0 index
             if i == 0:
-                features = feats
+                features = copy.deepcopy(feats)
 
         if np.prod(detections.shape):
             dets, keep = multiclass_nms(detections, max_num=self.max_number)
@@ -128,7 +129,7 @@ def predict_tile(
             if "feature_vector" in raw_predictions or "saliency_map" in raw_predictions:
                 features = [
                     raw_predictions["feature_vector"].reshape(-1),
-                    raw_predictions["saliency_map"],
+                    raw_predictions["saliency_map"][0],
                 ]
         return features, output