您好，我使用export.py将pt转为onnx后，替换Model文件和Classes文件并不生效，点击unity运行按钮会卡死。

[bro8D]

[Greatapikachu]

我也一样，你解决了吗？是不是onnx版本问题？

[EnoxSoftware]

I just tried the example with the latest assets and the onnx model worked without problems. And I got the same results with a newly created onnx model following the traditional export procedure.
YOLOv5_export_to_OpenCVDNN_ONNX.ipynb
The opset version of onnx must be 12.

[Greatapikachu]

好的，我明天试试

[Greatapikachu]

还是不行啊，我更换了我自己的.onnx also my own .names file, but it occurs this error：
dnn::forward_14() : OpenCV(4.8.0-dev) C:\Users\satoo\Desktop\opencv\modules\dnn\src\layers\reshape_layer.cpp:109: error: (-215:Assertion failed) total(srcShape, srcRange.start, srcRange.end) == maskTotal in function 'cv::dnn::computeShapeByReshapeMask'
opset version of onnx is 12, cvforunity is newest 2.5.5 , unity is 2020.3.37

[Greatapikachu]

是不是需要yolov5n.onnx，rather than yolov5.onnx??

[EnoxSoftware]

First of all, please tell me if the "yolov5n.onnx" model (Assets/StreamingAssets/YOLOv5WithOpenCVForUnityExample/yolov5n.onnx) included in this repository works without problems.
This model is the same as the onnx model I generated using the "YOLOv5_export_to_OpenCVDNN_ONNX.ipynb" command shown in my one previous answer.

It is possible to test if the generated onnx model works with python opencv, if it is ok with python opencv we believe it will also work with OpenCVForUnity.
https://github.com/ultralytics/yolov5/blob/master/detect.py

Inference with OpenCV's DnnModule:

python detect.py --weights yolov5n.onnx --dnn

[Greatapikachu]

最原来的yolov5n.onnx在unity中运行正常。
我自己生成的onnx文件best.onnx：

yolov5n.onnx：

这两个文件在python opencv运行都没问题，我查看输出图像也都没问题。but in cvforunity，我的best.onnx就会在DNN.forward这一步出现问题
please help me! Thanks so mush!

[Greatapikachu]

The infer function：
`Mat input_blob = preprocess(image);
Debug.Log(input_blob);

        // Forward
        object_detection_net.setInput(input_blob);

        List<Mat> output_blob = new List<Mat>();
        object_detection_net.forward(output_blob, object_detection_net.getUnconnectedOutLayersNames());

        // Postprocess
        Mat results = postprocess(output_blob[0], image.size());`

我发现在执行postprocess函数时，if (output_blob_0.size(2) < 5 + num_classes)程序直接return。
` protected virtual Mat postprocess(Mat output_blob, Size original_shape)
{
Mat output_blob_0 = output_blob;

        if (output_blob_0.size(2) < 5 + num_classes)
        {
            Debug.Log("output_blob不满足要求！");
            return new Mat(); 
        }`

所以我估计是forward出了问题。

[EnoxSoftware]

This is a Netron analysis of the "yolov5n.onnx" model settings. Compare with your "best.onnx". Especially the output tensor shape.

According to the YOLOv5ObjectDetector.cs code, num_classes should be equal to the number of classes in the model.
Is the shape of the Mat output by the forward method the same as the output shape of the model?

Debug.Log(output_blob_0); // 1*25200*85*CV_32FC1

[Greatapikachu]

The result of Debug.Log(output_blob_0):
Mat [ 12520056*CV_32FC1, isCont=True, isSubmat=False, nativeObj=0x2143584032736, dataAddr=0x2148075774080 ]
aha, Watson, you have discovered a blind spot. So how should I solve this problem?

[Greatapikachu]

this is yolov5n.onnx:

my best.onnx:

我的多了slice和concat，是不是这个原因？

[EnoxSoftware]

Hmmm, just to check the situation from the beginning, was "best.onnx" generated according to the contents of "YOLOv5_export_to_OpenCVDNN_ONNX.ipynb", i.e. using the

python export.py --weights yolov5n.pt --include onnx --opset 12

command line?
It is very strange that the model outputs different contents from "yolov5n.onnx".
I do all my work in Google Colaboratory and have never tested the custom trained models.

[Greatapikachu]

yeah, was generated by this command.
cvforunity.dnn不能处理非85通道数目的模型吗？有没有其他办法，有可能通过修改YOLOv5ObjectDetector.cs解决问题吗？

[EnoxSoftware]

It would be helpful if you could pass on the appropriate information to help us resolve the issue.
You seem to be consistently talking about custom trained models, but does inference testing on OpenCVForUnity of the newly exported onnx model with the preset "yolov5n.pt" model work?
Are you saying that the problem only occurs with custom training models converted to the onnx format?

This is my advice based on fragmented information,
If the output shape of the "best.onnx" model is [1,25600,56], then the number of classes is 51, so you need to rewrite num_classes in line 25 of YOLOv5ObjectDetector.cs to 51. Also, the classes value in the scene inspector should be changed from "coco.names" to empty because num_classes may be changed on line 56.

[Greatapikachu]

Inference testing on OpenCVForUnity of the newly exported onnx model with the preset "yolov5n.pt" model work well.
The problem only occurs with my custom training models converted to the onnx format.