opencv python dnn reading exported ONNX problem

sonOfAnton29 commented 2 years ago

Search before asking

[X] I have searched the YOLOv5 issues and found no similar bug report.

YOLOv5 Component

Detection, Export

Bug

hello. I recently faced an issue while working on yolov5. I trained a model in yolov5 small weight format, using yolov5s.pt as initial weights. but after I exported that weight file to ONNX, cv2.dnn.readNetFromONNX() can't read it and raises this exception:

cv2.error: OpenCV(4.6.0) D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\onnx\onnx_importer.cpp:1040: error: (-2:Unspecified error) in function 'cv::dnn::dnn4_v20220524::ONNXImporter::handleNode'

Node [Identity@ai.onnx]:(onnx_node!Identity_0) parse error: OpenCV(4.6.0) D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\layer.cpp:246: error: (-215:Assertion failed) inputs.size() in function 'cv::dnn::dnn4_v20220524::Layer::getMemoryShapes'

I tried yolov5m.pt and some other options such as fp16 format export but it does not work. I would appreciate your help

Environment

windows 10 cuda available python 3.8.8 opencv-python 4.6.0

Minimal Reproducible Example

No response

Additional

No response

Are you willing to submit a PR?

[ ] Yes I'd like to help by submitting a PR!

JustasBart commented 2 years ago

Hi, I've only done this on C++, but just by reading that I believe that your input shape might be incorrect. Please be aware that unless if you export your .onnx with the --dynamic flag it will have a fixed input shape of the model unlike the .pt models. Please use the Netron viewer and investigate your (.onnx) model to see what the input size is.

For example:

python3 export.py \
--weights yolov5s.pt \
--img 640 \
--simplify \
--optimize \
--include onnx

Netron viewer:

So now I know that my OpenCV::dnn::fromONNX(myModel.onnx) will want an input size of 640x640

I hope that this helps you out!

wvalcke commented 2 years ago

Please check : #10304 You need to change export.py : do_constant_folding=True, If you do this, then the OpenCV 4.6.0 DNN reading of the onnx model will work.

sonOfAnton29 commented 2 years ago

Please check : #10304 You need to change export.py : do_constant_folding=True, If you do this, then the OpenCV 4.6.0 DNN reading of the onnx model will work.

it was already set to true but thanks for reply

sonOfAnton29 commented 2 years ago

@JustasBart thank you. the input size wasn't the problem I used your export code and it worked. I think optimize or simplify did the trick but I don't know how

JustasBart commented 2 years ago

@sonOfAnton29 I'm glad to hear it, cheers! :rocket:

deannazhang20 commented 1 year ago

I have the same problem you meet. Who can help to solve the problem

wvalcke commented 1 year ago

Just a suggestion. Using the export to onnx of the standard yolov5 implementation gives a working onnx model if you use the onnxruntime software. I tested this with CPU & GPU and the performance is very good. Onnxruntime supports python & C++ inference, i tested the python version at this moment. What i did was combining opencv for image preprocessing & display, but used onnxruntime for the model inference. The yolov5s model runs about 15FPS on a corei7 10850H cpu. I did the same test using a RTX 2080Ti GPU, and the raw inference performance is 211FPS! Personally i would just avoid dnn from OpenCv and use onnxruntime for model inference.

glenn-jocher commented 1 year ago

@wvalcke agreed. Avoid DNN and use ONNX directly.

github-actions[bot] commented 1 year ago

👋 Hello, this issue has been automatically marked as stale because it has not had recent activity. Please note it will be closed if no further activity occurs.

Access additional YOLOv5 🚀 resources:

Wiki – https://github.com/ultralytics/yolov5/wiki
Tutorials – https://docs.ultralytics.com/yolov5
Docs – https://docs.ultralytics.com

Access additional Ultralytics ⚡ resources:

Ultralytics HUB – https://ultralytics.com/hub
Vision API – https://ultralytics.com/yolov5
About Us – https://ultralytics.com/about
Join Our Team – https://ultralytics.com/work
Contact Us – https://ultralytics.com/contact

Feel free to inform us of any other issues you discover or feature requests that come to mind in the future. Pull Requests (PRs) are also always welcomed!

Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐!

ultralytics / yolov5