Dev

pth2pt_YOEO.py

@@ -0,0 +1,23 @@
+import torch
+from yoeo import models
+import torchvision
+from detectron2.modeling import build_model
+
+def inference_func(model, image):
+    inputs = [{"image": image}]
+    return model.inference(inputs, do_postprocess=False)[0]
+
+cfg = "/home/ss21mipt/Documents/starkit/DIPLOMA/YOEO/config/yoeo.cfg"
+# weights = "/home/ss21mipt/Documents/starkit/DIPLOMA/YOEO/weights/yoeo_mine_759.pth"
+weights = "/home/ss21mipt/Documents/starkit/DIPLOMA/YOEO/weights/yoeo.pth"
+# print("cfg.MODEL.WEIGHTS: ",cfg.MODEL.WEIGHTS)   ## RETURNS : cfg.MODEL.WEIGHTS:  drive/Detectron2/model_final.pth
+
+model = models.load_model(cfg, weights, "cpu")
+print(model)
+
+example = torch.rand(1, 3, 416, 416)
+# wrapper = TracingAdapter(module, example, inference_func)
+# model.eval()
+traced_script_module = torch.jit.trace(model, example)
+# traced_script_module = torch.jit.trace(wrapper, (example,))
+traced_script_module.save("/home/ss21mipt/Documents/starkit/DIPLOMA/YOEO/weights/model-final.pt")

pyproject.toml

@@ -34,3 +34,6 @@ build-backend = "poetry.core.masonry.api"
 yoeo-detect = "yoeo.detect:run"
 yoeo-train = "yoeo.train:run"
 yoeo-test = "yoeo.test:run"
+yoeo-detectcam = "yoeo.detectwebcam:run"
+yoeo-detectcam_IR = "yoeo.detectwebcam_IR:run"
+yoeo-detectcam_pt = "yoeo.detectwebcam_pt:run"

scripts/convertONNXModelToOpenVinoIR.py

@@ -0,0 +1,66 @@
+#! /usr/bin/env python3
+import argparse
+import os
+
+
+def convert_model(onnx_path: str, output_path: str) -> None:
+    command = assemble_command(onnx_path, output_path)
+    run_command(command)
+
+
+def assemble_command(onnx_path: str, output_path: str) -> str:
+    # https://docs.openvino.ai/latest/notebooks/102-pytorch-onnx-to-openvino-with-output.html (April 7, 2022)
+    mo_command = f"""mo
+		         --input_model "{onnx_path}"
+		         --output_dir "{output_path}"
+		         --input InputLayer
+		         --output Detections,Segmentations
+		         --framework onnx
+		         --static_shape
+		         --batch 1
+		         """
+    mo_command = " ".join(mo_command.split())
+
+    return mo_command
+
+
+def run_command(command: str) -> None:
+    # https://docs.openvino.ai/latest/notebooks/102-pytorch-onnx-to-openvino-with-output.html (April 7, 2022)
+    print("Exporting ONNX model to IR...")
+
+    mo_result = os.system(command)
+
+    print("=" * 30)
+    if mo_result == 0:
+        print("Model conversion was successful")
+    else:
+        print("Model conversion failed")
+
+
+def get_output_path(model_onnx: str) -> str:
+    return get_parent_dir(model_onnx)
+
+
+def get_parent_dir(path: str) -> str:
+    absolute_path = os.path.abspath(path)
+    parent_dir, filename = os.path.split(absolute_path)
+
+    return parent_dir
+
+
+def run():
+    parser = argparse.ArgumentParser(description='Convert ONNX Model to OpenVino IR')
+    parser.add_argument(
+        "model_onnx",
+        type=str,
+        help="full path to model file (.onnx)"
+    )
+
+    args = parser.parse_args()
+
+    output_path = get_output_path(args.model_onnx)
+    convert_model(args.model_onnx, output_path)
+
+
+if __name__ == "__main__":
+    run()

scripts/convertPyTorchModelToONNX.py

@@ -0,0 +1,100 @@
+#! /usr/bin/env python3
+import argparse
+import os.path
+from typing import Tuple
+
+import onnx
+import torch
+
+import yoeo.models
+
+
+def convert_model(model_cfg: str, weights_pth: str, output_path: str) -> None:
+    pytorch_model = yoeo.models.load_model(model_cfg, weights_pth)
+    convert_to_onnx(model=pytorch_model, output_path=output_path)
+
+
+def convert_to_onnx(model: yoeo.models.Darknet, output_path: str, image_size: int = 416, batch_size: int = 1) -> None:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    model.to(device)
+    dummy_input = torch.randn(batch_size, 3, image_size, image_size, device=device)
+
+    torch.onnx.export(
+        model,
+        dummy_input,
+        output_path,
+        verbose=False,
+        export_params=True,
+        input_names=["InputLayer"],
+        output_names=["Detections", "Segmentations"],
+        opset_version=11
+    )
+
+
+def check_model(model_path: str) -> None:
+    onnx_model = load_onnx(model_path)
+    check_onnx(onnx_model)
+
+
+def load_onnx(path: str) -> onnx.onnx_ml_pb2.ModelProto:
+    return onnx.load(path)
+
+
+def check_onnx(model: onnx.onnx_ml_pb2.ModelProto) -> None:
+    # https://github.com/onnx/onnx/blob/main/docs/PythonAPIOverview.md (April 7, 2022)
+    print("="*30)
+    try:
+        onnx.checker.check_model(model)
+    except onnx.checker.ValidationError as e:
+        print('The model is invalid: %s' % e)
+    else:
+        print('The model is valid!')
+
+
+def construct_path(model_cfg: str) -> str:
+    parent_dir = get_parent_dir(model_cfg)
+    filename = get_filename_wout_extension(model_cfg)
+
+    onnx_path = os.path.join(parent_dir, f"{filename}.onnx")
+
+    return onnx_path
+
+
+def get_parent_dir(path: str) -> str:
+    absolute_path = os.path.abspath(path)
+    parent_dir, filename = os.path.split(absolute_path)
+
+    return parent_dir
+
+
+def get_filename_wout_extension(path: str) -> str:
+    absolute_path = os.path.abspath(path)
+    parent_dir, filename = os.path.split(absolute_path)
+    filename, ext = os.path.splitext(filename)
+
+    return filename
+
+
+def run():
+    parser = argparse.ArgumentParser(description='Convert PyTorch Model to ONNX')
+    parser.add_argument(
+        "model_cfg",
+        type=str,
+        help="full path to model file (.cfg). ONNX model will be output with the same filename as well."
+    )
+    parser.add_argument(
+        "model_weights",
+        type=str,
+        help="full path to model weights file (.pth or .weights)."
+    )
+
+    args = parser.parse_args()
+
+    onnx_path = construct_path(args.model_cfg)
+    convert_model(args.model_cfg, args.model_weights, onnx_path)
+    check_model(onnx_path)
+
+
+if __name__ == "__main__":
+    run()

tezises/inputs&outputs

@@ -0,0 +1,38 @@
+PYTORCH .cfg + .pth
+
+num 1 DETECTIONS SHAPE IStorch.Size([1, 6000, 8]), <class 'torch.Tensor'>
+num 1 SEGMENTATIONS SHAPE IStorch.Size([1, 640, 640]), <class 'torch.Tensor'>
+
+torchscipt
+
+python .pt
+SHAPE OF EVERY ELEM: torch.Size([1, 3, 20, 20, 8])
+SHAPE OF EVERY ELEM: torch.Size([1, 3, 40, 40, 8])
+SHAPE OF EVERY ELEM: torch.Size([1, 3, 640, 640])
+
+cpp .pt
+Shape of every elem: [1, 3, 20, 20, 8]
+Shape of every elem: [1, 3, 40, 40, 8]
+Shape of every elem: [1, 3, 640, 640]
+
+IR
+
+python .xml
+outputs[
+<ConstOutput: names[Detections] shape[1,2535,8] type: f32>,
+<ConstOutput: names[Segmentations] shape[1,416,416] type: f32>
+]>
+
+cpp .xml
+[1,2535,8]
+[1,416,416]
+
+
+зрение робана
+
+Пути к весам и обработчику сеток:		~/env/common/vision_filters/ball_detection_wideangle.json
+Чето там еще (трешолды и другие параметры)	~/env/common/vision_filters/colors_wideangle.json
+пайплайн зрения прописан здесь		~/env/common/vision_filters/all.json
+
+
+

yoeo/detect.py

@@ -1,5 +1,4 @@
 #! /usr/bin/env python3
-
 from __future__ import division
 
 import os
@@ -59,6 +58,7 @@ def detect_directory(model_path, weights_path, img_path, classes, output_path,
         output_path,
         conf_thres,
         nms_thres)
+
     _draw_and_save_output_images(
         img_detections, segmentations, imgs, img_size, output_path, classes)
 
@@ -134,11 +134,15 @@ def detect(model, dataloader, output_path, conf_thres, nms_thres):
 
     for (img_paths, input_imgs) in tqdm.tqdm(dataloader, desc="Detecting"):
         # Configure input
+        print(f"VOT TAKOE 1 {input_imgs}")
         input_imgs = Variable(input_imgs.type(Tensor))
 
+        print(f"VOT TAKOE 2 {input_imgs}")
+
         # Get detections
         with torch.no_grad():
             detections, segmentations = model(input_imgs)
+            print(f"VOT TAKOE 3 {input_imgs}")
             detections = non_max_suppression(detections, conf_thres, nms_thres)
 
         # Store image and detections
@@ -195,7 +199,7 @@ def _draw_and_save_output_image(image_path, detections, seg, img_size, output_pa
     # Get segmentation
     seg = seg.cpu().detach().numpy().astype(np.uint8)
     # Draw all of it
-
+    print(f"ETO EST SEG {seg}")
     # The amount of padding that was added
     pad_x = max(img.shape[0] - img.shape[1], 0) * (img_size / max(img.shape[:2])) // 2
     pad_y = max(img.shape[1] - img.shape[0], 0) * (img_size / max(img.shape[:2])) // 2
@@ -206,13 +210,14 @@ def _draw_and_save_output_image(image_path, detections, seg, img_size, output_pa
                 int(pad_x) : int(img_size - pad_x),
                 ] * 255
 
-
+    print(f"MILLIARDNAYA {img}")
     ax.imshow(
         SegmentationMapsOnImage(
             seg[
                 int(pad_y) : int(img_size - pad_y),
                 int(pad_x) : int(img_size - pad_x),
                 ], shape=img.shape).draw_on_image(img)[0])
+    print("JJEEPPAAA")
     # Rescale boxes to original image
     detections = rescale_boxes(detections, img_size, img.shape[:2])
     unique_labels = detections[:, -1].cpu().unique()
@@ -275,6 +280,7 @@ def _create_data_loader(img_path, batch_size, img_size, n_cpu):
     :return: Returns DataLoader
     :rtype: DataLoader
     """
+
     dataset = ImageFolder(
         img_path,
         transform=transforms.Compose([DEFAULT_TRANSFORMS, Resize(img_size)]))
@@ -318,6 +324,5 @@ def run():
         conf_thres=args.conf_thres,
         nms_thres=args.nms_thres)
 
-
 if __name__ == '__main__':
     run()