Update ModelAPI in 1.4 release (#2347)

* Upgrade model API

* Update otx in exportable code

* Fix unit tests

* Fix black

* Fix detection inference

* Fix det tiling

* Fix mypy

* Fix demo

* Fix visualizer in demo

* Fix black

Author: sovrasov

requirements/openvino.txt

@@ -2,7 +2,7 @@
 # OpenVINO Requirements.                                                      #
 nncf==2.5.0
 onnx==1.13.0
-openvino-model-api==0.1.2
+openvino-model-api==0.1.3
 openvino==2023.0
 openvino-dev==2023.0
 openvino-telemetry>=2022.1.0

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

@@ -235,7 +235,7 @@ def __init__(
     def post_process(self, prediction: Dict[str, np.ndarray], metadata: Dict[str, Any]) -> AnnotationSceneEntity:
         """Detection specific post-process."""
         detections = self.model.postprocess(prediction, metadata)
-        detections = detection2array(detections)
+        detections = detection2array(detections.objects)
         return self.converter.convert_to_annotation(detections, metadata)
 
 

src/otx/api/usecases/exportable_code/demo/demo_package/executors/asynchronous.py

@@ -78,7 +78,9 @@ def render_result(self, results: Tuple[Any, dict]) -> np.ndarray:
         predictions, frame_meta = results
         if isinstance(self.converter, DetectionToAnnotationConverter):
             # Predictions for the detection task
-            predictions = np.array([[pred.id, pred.score, *pred.get_coords()] for pred in predictions])
+            predictions = np.array(
+                [[pred.id, pred.score, *[pred.xmin, pred.ymin, pred.xmax, pred.ymax]] for pred in predictions.objects]
+            )
             predictions.shape = len(predictions), 6
         annotation_scene = self.converter.convert_to_annotation(predictions, frame_meta)
         current_frame = frame_meta["frame"]

src/otx/api/usecases/exportable_code/demo/demo_package/model_container.py

@@ -121,7 +121,7 @@ def infer(self, frame):
 
         # MaskRCNN returns tuple so no need to process
         if self._task_type == TaskType.DETECTION:
-            predictions = detection2array(predictions)
+            predictions = detection2array(predictions.objects)
         return predictions, frame_meta
 
     def infer_tile(self, frame):

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

@@ -1,4 +1,4 @@
 openvino==2023.0
-openvino-model-api==0.1.2
-otx==1.4.0rc2
+openvino-model-api==0.1.3
+otx @ git+https://github.com/openvinotoolkit/training_extensions/@e4269e035bcaa3903c6b99044f46c42fcbf98f25#egg=otx
 numpy>=1.21.0,<=1.23.5  # np.bool was removed in 1.24.0 which was used in openvino runtime

src/otx/api/usecases/exportable_code/prediction_to_annotation_converter.py

@@ -214,7 +214,7 @@ def convert_to_annotation(
         image_size = metadata["original_shape"][1::-1]
         for box in predictions:
             scored_label = ScoredLabel(self.labels[int(box.id)], float(box.score))
-            coords = np.array(box.get_coords(), dtype=float)
+            coords = np.array([box.xmin, box.ymin, box.xmax, box.ymax], dtype=float)
             if (coords[2] - coords[0]) * (coords[3] - coords[1]) < 1.0:
                 continue
             coords /= np.tile(image_size, 2)

src/otx/api/utils/tiler.py

@@ -4,14 +4,15 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-import copy
 import cv2
 from itertools import product
 from typing import Dict, List, Optional, Tuple, Union
 
 import numpy as np
 from openvino.model_api.models import Model, ImageModel
 
+from openvino.model_api.models.utils import DetectionResult
+
 from otx.algorithms.common.utils.logger import get_logger
 from otx.api.utils.async_pipeline import OTXDetectionAsyncPipeline
 from otx.api.utils.detection_utils import detection2array
@@ -196,7 +197,7 @@ def predict_async(self, image: np.ndarray, tile_coords: List[List[int]]):
         merged_features = self.merge_features(features, merged_results)
         return merged_results, merged_features
 
-    def postprocess_tile(self, predictions: Union[List, Tuple], offset_x: int, offset_y: int) -> Dict[str, List]:
+    def postprocess_tile(self, predictions: DetectionResult, offset_x: int, offset_y: int) -> Dict[str, List]:
         """Postprocess single tile prediction.
 
         Args:
@@ -221,8 +222,8 @@ def postprocess_tile(self, predictions: Union[List, Tuple], offset_x: int, offse
             )
             output_dict["masks"] = tile_masks
         else:
-            assert isinstance(predictions, list)
-            out = detection2array(predictions)
+            assert isinstance(predictions.objects, list)
+            out = detection2array(predictions.objects)
             out[:, 2:] += np.tile([offset_x, offset_y], 2)
         output_dict["bboxes"] = out
         return output_dict

tests/unit/algorithms/action/adapters/openvino/test_action_openvino_models.py

@@ -221,4 +221,4 @@ def test_postprocess(self) -> None:
         # argmax index is 2 because first index is for background
         assert out[0].id == 2
         assert out[0].score == 0.7
-        assert out[0].get_coords() == (0, 0, 256, 256)
+        assert (out[0].xmin, out[0].ymin, out[0].xmax, out[0].ymax) == (0, 0, 256, 256)

tests/unit/algorithms/detection/adapters/openvino/model_wrappers/test_detection_openvino_models.py

@@ -78,6 +78,8 @@ def __init__(self, *args):
         self.resize_type = "standard"
         self.output_parser = MockBatchBoxesLabelsParser()
         self.labels = []
+        self.w = 10
+        self.h = 10
         super().__init__(MockOpenvinoAdapter)
 
 
@@ -142,7 +144,7 @@ def test_postprocess(self) -> None:
         }
         sample_meta = {"original_shape": (10, 10, 3), "resized_shape": (5, 5, 3)}
         out = self.model.postprocess(sample_output, meta=sample_meta)
-        assert len(out) <= 1
+        assert len(out.objects) <= 1
 
 
 class TestBatchBoxesLabelsParser: