Yolov5 OpenCV error: (-215:Assertion failed) whilst using .onnx

Question

I've just created a yolov5 model, and exported it in the onnx format so it is usable with opencv but I keep getting the error:

[ERROR:0] global D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\dnn.cpp (3554) cv::dnn::dnn4_v20211004::Net::Impl::getLayerShapesRecursively OPENCV/DNN: [Reshape]:(466): getMemoryShapes() throws exception. inputs=1 outputs=1/1 blobs=0 [ERROR:0] global D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\dnn.cpp (3557) cv::dnn::dnn4_v20211004::Net::Impl::getLayerShapesRecursively input[0] = [ 1 24 52 52 ] [ERROR:0] global D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\dnn.cpp (3561) cv::dnn::dnn4_v20211004::Net::Impl::getLayerShapesRecursively output[0] = [ ] [ERROR:0] global D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\dnn.cpp (3567) cv::dnn::dnn4_v20211004::Net::Impl::getLayerShapesRecursively Exception message: OpenCV(4.5.4) D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\layers\reshape_layer.cpp:106: error: (-215:Assertion failed) total(srcShape, srcRange.start, srcRange.end) == maskTotal in function 'cv::dnn::computeShapeByReshapeMask'

here is the code:

onnx_model = onnx.load('best.onnx')
onnx.checker.check_model(onnx_model)
net = cv.dnn.readNetFromONNX('best.onnx')


classes = []

with open("coco-dataset.labels", "r") as f:
    classes = [line.strip() for line in f.readlines()]

layerNames = net.getLayerNames()
outputLayers=[]
for i in net.getUnconnectedOutLayers():
    outputLayers.append(layerNames[i-1])




bounding_box = {'top': 0, 'left': 0, 'width': 1920, 'height': 1080}
sct = mss()

while True:
    screenShot = sct.grab(bounding_box)
    img_np = np.array(screenShot)
    frame = cv.cvtColor(img_np, cv.COLOR_RGBA2RGB)
    frame_id += 1
    height, width, channels= frame.shape


    blob = cv.dnn.blobFromImage(frame, 0.0039216, (416, 416), (0, 0, 0), True, crop=False)


    cv.imshow("Image", frame)

    net.setInput(blob)

(THIS IS THE ERROR LINE)

    outs = net.forward(outputLayers)

I'm not sure if it is just my understanding of onnx, or if there is a fix I haven't picked up on.

Any help is appreciated!

Answer 1

Until now latest version of OpenCV (4.5.4), OpenCV unable to load onnx file without simplification operation. You have exported yolov5 pt file to onnx file with below command.

 python export.py --weights .\best.pt

Try to export pt file to onnx file with below commands.

For Ubuntu:

python export.py --weights .\yolov5s.pt --include onnx --simplify

For Windows

pip install onnx
pip install onnx-simplifier 
python export.py --weights .\yolov5s.pt --include onnx --simplify

and then load created onnx file with OpenCV, OpenCV will read that onnx successfully.

Also change you model size to (640,640) in blobfromimage function.