convert yolov5s.pt to openvino IR format success, predict stuff failed

Question

This is begin from curious....I download the pretrained yolov5s.pt from public google drive, and convert it as yolov5s.onnx file with input shape [1,3,640,640] by using yolov5's models/export.py. Then I use openvino's deployment tools/mo.py to convert the yolov5s.onnx into openvino inference engines file (.xml+.bin). The conversion is success without error. At last, I try to run the predict by using these files by using openvino's demo program the prediction is successfully return the result. But all the result return negative numbers, and the array level is wrong. Is it impossible to convert the yolov5 weights for openvino?

Answer 1

YOLOv5 is currently not an officially supported topology by the OpenVINO toolkit. Please see the list of validated ONNX and PyTorch topologies here https://docs.openvinotoolkit.org/latest/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_ONNX.html

However, we have one recommendation for you to try, but it was no guaranteed it will succeed. You can do it by change export.py to include the Detect layer: yolov5/models/export.py

Line 28 in a1c8406

model.model[-1].export = True  # set Detect() layer export=True 

The above needs to be changed from True to False. For more detail, you can follow this thread here.

Answer 2

Try this :python mo.py --input_model yolov5s.onnx -s 255 --reverse_input_channels --output Conv_245,Conv_261,Conv_277 (use Netron to check your architecture)