black

Author: Wovchena

model_api/cpp/models/src/detection_model_ssd.cpp

@@ -162,11 +162,11 @@ std::unique_ptr<ResultBase> ModelSSD::postprocessSingleOutput(InferenceResult& i
                 0.f,
                 floatInputImgHeight);
             desc.width = clamp(
-                round((detections[i * objectSize + 5] * netInputWidth - padLeft) * invertedScaleX),
+                round((detections[i * numAndStep.objectSize + 5] * netInputWidth - padLeft) * invertedScaleX),
                 0.f,
                 floatInputImgWidth) - desc.x;
             desc.height = clamp(
-                round((detections[i * objectSize + 6] * netInputHeight - padTop) * invertedScaleY),
+                round((detections[i * numAndStep.objectSize + 6] * netInputHeight - padTop) * invertedScaleY),
                 0.f,
                 floatInputImgHeight) - desc.y;
             result->objects.push_back(desc);
@@ -224,11 +224,11 @@ std::unique_ptr<ResultBase> ModelSSD::postprocessMultipleOutputs(InferenceResult
                 0.f,
                 floatInputImgHeight);
             desc.width = clamp(
-                round((boxes[i * objectSize + 2] * widthScale - padLeft) * invertedScaleX),
+                round((boxes[i * numAndStep.objectSize + 2] * widthScale - padLeft) * invertedScaleX),
                 0.f,
                 floatInputImgWidth) - desc.x;
             desc.height = clamp(
-                round((boxes[i * objectSize + 3] * heightScale - padTop) * invertedScaleY),
+                round((boxes[i * numAndStep.objectSize + 3] * heightScale - padTop) * invertedScaleY),
                 0.f,
                 floatInputImgHeight) - desc.y;
             result->objects.push_back(desc);

model_api/cpp/models/src/detection_model_yolo.cpp

@@ -544,7 +544,7 @@ void YoloV8::prepareInputsOutputs(std::shared_ptr<ov::Model>& model) {
         throw std::runtime_error("YoloV8 wrapper requires the output to be of rank 3");
     }
     if (!labels.empty() && labels.size() + 4 != out_shape[1]) {
-        throw std::runtime_error("YoloV8 wrapper number of labes must be smaller than output.shape[1] by 4");
+        throw std::runtime_error("YoloV8 wrapper number of labels must be smaller than output.shape[1] by 4");
     }
 }
 

model_api/cpp/utils/include/utils/nms.hpp

@@ -54,7 +54,10 @@ struct AnchorLabeled : public Anchor {
 
 template <typename Anchor>
 std::vector<size_t> nms(const std::vector<Anchor>& boxes, const std::vector<float>& scores,
-                     const float thresh, bool includeBoundaries=false, size_t maxNum=std::numeric_limits<size_t>::max()) {
+                     const float thresh, bool includeBoundaries=false, size_t keep_top_k=0) {
+    if (keep_top_k == 0) {
+        keep_top_k = boxes.size();
+    }
     std::vector<float> areas(boxes.size());
     for (size_t i = 0; i < boxes.size(); ++i) {
         areas[i] = (boxes[i].right - boxes[i].left + includeBoundaries) * (boxes[i].bottom - boxes[i].top + includeBoundaries);
@@ -64,7 +67,7 @@ std::vector<size_t> nms(const std::vector<Anchor>& boxes, const std::vector<floa
     std::sort(order.begin(), order.end(), [&scores](int o1, int o2) { return scores[o1] > scores[o2]; });
 
     size_t ordersNum = 0;
-    for (; ordersNum < order.size() && scores[order[ordersNum]] >= 0  && ordersNum < maxNum; ordersNum++);
+    for (; ordersNum < order.size() && scores[order[ordersNum]] >= 0  && ordersNum < keep_top_k; ordersNum++);
 
     std::vector<size_t> keep;
     bool shouldContinue = true;

model_api/python/openvino/model_api/models/__init__.py

@@ -118,7 +118,7 @@
     "YOLO",
     "YoloV3ONNX",
     "YoloV4",
-    "YOLOv5"
+    "YOLOv5",
     "YOLOv8",
     "YOLOF",
     "YOLOX",

model_api/python/openvino/model_api/models/model.py

@@ -126,7 +126,7 @@ def create_model(
         core=None,
         weights_path="",
         adaptor_parameters={},
-        device="CPU",
+        device="AUTO",
         nstreams="1",
         nthreads=None,
         max_num_requests=0,

model_api/python/openvino/model_api/models/utils.py

@@ -340,7 +340,7 @@ def crop_resize(image, size):
 }
 
 
-def nms(x1, y1, x2, y2, scores, thresh, include_boundaries=False, keep_top_k=None):
+def nms(x1, y1, x2, y2, scores, thresh, include_boundaries=False, keep_top_k=0):
     b = 1 if include_boundaries else 0
     areas = (x2 - x1 + b) * (y2 - y1 + b)
     order = scores.argsort()[::-1]
@@ -363,11 +363,10 @@ def nms(x1, y1, x2, y2, scores, thresh, include_boundaries=False, keep_top_k=Non
         intersection = w * h
 
         union = areas[i] + areas[order[1:]] - intersection
-        overlap = np.zeros_like(intersection, dtype=float)
         overlap = np.divide(
             intersection,
             union,
-            out=overlap,
+            out=np.zeros_like(intersection, dtype=float),
             where=union != 0,
         )
 

model_api/python/openvino/model_api/models/yolo.py

@@ -574,7 +574,6 @@ def set_strides_grids(self):
         self.expanded_strides = np.concatenate(expanded_strides, 1)
 
 
-
 class YoloV3ONNX(DetectionModel):
     __model__ = "YOLOv3-ONNX"
 
@@ -766,7 +765,9 @@ def non_max_suppression(
 
     # Detections matrix nx6 (xyxy, conf, cls)
     box, cls, mask = x[:, :4], x[:, 4 : nc + 4], x[:, nc + 4 :]
-    box = xywh2xyxy(box)  # center_x, center_y, width, height) to (x1, y1, x2, y2)  # TODO: first cut by conf_thres
+    box = xywh2xyxy(
+        box
+    )  # center_x, center_y, width, height) to (x1, y1, x2, y2)  # TODO: first cut by conf_thres
     if multi_label:
         i, j = (cls > conf_thres).nonzero(as_tuple=False).T
         x = torch.cat((box[i], x[i, 4 + j, None], j[:, None].float(), mask[i]), 1)
@@ -808,13 +809,9 @@ def __init__(self, inference_adapter, configuration, preload=False):
             self.raise_error("the output must be of precision f32")
         out_shape = output.shape
         if 3 != len(out_shape):
-            self.raise_error(
-                "the output must be of rank 3"
-            )
+            self.raise_error("the output must be of rank 3")
         if self.labels and len(self.labels) + 4 != out_shape[1]:
-            self.raise_error(
-                "number of labes must be smaller than out_shape[1] by 4"
-            )
+            self.raise_error("number of labels must be smaller than out_shape[1] by 4")
 
     @classmethod
     def parameters(cls):
@@ -841,7 +838,9 @@ def parameters(cls):
     def postprocess(self, outputs, meta):
         if 1 != len(outputs):
             raise RuntimeError("YoloV8 wrapper expects 1 output")
-        boxes = non_max_suppression(next(iter(outputs.values())), self.confidence_threshold, self.iou_threshold)
+        boxes = non_max_suppression(
+            next(iter(outputs.values())), self.confidence_threshold, self.iou_threshold
+        )
 
         inputImgWidth, inputImgHeight = (
             meta["original_shape"][1],
@@ -884,4 +883,5 @@ def postprocess(self, outputs, meta):
 
 class YOLOv8(YOLOv5):
     """YOLOv5 and YOLOv8 are identical in terms of inference"""
+
     __model__ = "YOLOv8"

model_api/python/setup.py

@@ -38,7 +38,12 @@
     install_requires=(SETUP_DIR / "requirements.txt").read_text(),
     extras_require={
         "ovms": (SETUP_DIR / "requirements_ovms.txt").read_text(),
-        "tests": ["httpx", "pytest", "openvino-dev[onnx,pytorch,tensorflow2]", "ultralytics>=8.0.114"],
+        "tests": [
+            "httpx",
+            "pytest",
+            "openvino-dev[onnx,pytorch,tensorflow2]",
+            "ultralytics>=8.0.114",
+        ],
     },
     long_description=(SETUP_DIR.parents[1] / "README.md").read_text(),
     long_description_content_type="text/markdown",

tests/python/accuracy/conftest.py

@@ -14,15 +14,36 @@ def pytest_addoption(parser):
 
 
 def _impaths(data):
-    impaths = sorted(file for file in (Path(data) / "coco128/images/train2017/").iterdir())
+    impaths = sorted(
+        file for file in (Path(data) / "coco128/images/train2017/").iterdir()
+    )
     if not impaths:
         raise RuntimeError(f"{Path(data) / 'coco128/images/train2017/'} is empty")
     return impaths
 
 
 def pytest_generate_tests(metafunc):
     if "pt" in metafunc.fixturenames:
-        metafunc.parametrize("pt", ("yolov5n6u.pt", "yolov5s6u.pt", "yolov5m6u.pt", "yolov5l6u.pt", "yolov5x6u.pt", "yolov8n.pt", "yolov8s.pt", "yolov8m.pt", "yolov8l.pt", "yolov8x.pt", "yolov5nu.pt", "yolov5su.pt", "yolov5mu.pt", "yolov5lu.pt", "yolov5xu.pt"))
+        metafunc.parametrize(
+            "pt",
+            (
+                "yolov5n6u.pt",
+                "yolov5s6u.pt",
+                "yolov5m6u.pt",
+                "yolov5l6u.pt",
+                "yolov5x6u.pt",
+                "yolov8n.pt",
+                "yolov8s.pt",
+                "yolov8m.pt",
+                "yolov8l.pt",
+                "yolov8x.pt",
+                "yolov5nu.pt",
+                "yolov5su.pt",
+                "yolov5mu.pt",
+                "yolov5lu.pt",
+                "yolov5xu.pt",
+            ),
+        )
     if "impath" in metafunc.fixturenames:
         metafunc.parametrize("impath", _impaths(metafunc.config.getoption("data")))
 

tests/python/accuracy/test_YoloV8.py

@@ -15,11 +15,14 @@
 from pathlib import Path
 import functools
 
+
 # TODO: update docs
 def patch_export(yolo):
     # TODO: move to https://github.com/ultralytics/ultralytics/
     if yolo.predictor is None:
-        yolo.predict(np.zeros([1, 1, 3], np.uint8))  # YOLO.predictor is initialized by predict
+        yolo.predict(
+            np.zeros([1, 1, 3], np.uint8)
+        )  # YOLO.predictor is initialized by predict
     export_dir = Path(yolo.export(format="openvino"))
     xml = [path for path in export_dir.iterdir() if path.suffix == ".xml"]
     if 1 != len(xml):
@@ -154,13 +157,17 @@ def cached_models(folder, pt):
     pt = Path(pt)
     yolo_folder = folder / "YOLOv8"
     yolo_folder.mkdir(exist_ok=True)  # TODO: maybe remove
-    export_dir = patch_export(YOLO(yolo_folder / pt))  # If there is no file it is downloaded
+    export_dir = patch_export(
+        YOLO(yolo_folder / pt)
+    )  # If there is no file it is downloaded
     copy_path = folder / "YOLOv8/detector" / pt.stem
     copy_tree(str(export_dir), str(copy_path))  # C++ tests expect a model here
-    xml = (copy_path / (pt.stem + ".xml"))
+    xml = copy_path / (pt.stem + ".xml")
     ref_dir = copy_path / "ref"
     ref_dir.mkdir(exist_ok=True)
-    impl_wrapper = YOLOv5.create_model(xml, device="CPU", model_type="YOLOv5")  # TODO: YOLOv5 vs v8
+    impl_wrapper = YOLOv5.create_model(
+        xml, device="CPU", model_type="YOLOv5"
+    )  # TODO: YOLOv5 vs v8
     ref_wrapper = YOLO(export_dir)
     ref_wrapper.overrides["imgsz"] = (impl_wrapper.w, impl_wrapper.h)
     compiled_model = ov.Core().compile_model(xml, "CPU")
@@ -243,10 +250,11 @@ def test_detector(impath, data, pt):
     )
     ref_preprocessed = ref_wrapper.predictor.preprocess([im]).numpy()
 
-    processed = resize_image_letterbox(im, (impl_wrapper.w, impl_wrapper.h), cv2.INTER_LINEAR, 114)
+    processed = resize_image_letterbox(
+        im, (impl_wrapper.w, impl_wrapper.h), cv2.INTER_LINEAR, 114
+    )
     processed = (
-        processed[None][..., ::-1].transpose((0, 3, 1, 2)).astype(np.float32)
-        / 255.0
+        processed[None][..., ::-1].transpose((0, 3, 1, 2)).astype(np.float32) / 255.0
     )
     assert (processed == ref_preprocessed).all()
     preds = next(iter(compiled_model({0: processed}).values()))