yolov5 release tag v5 , main_yolov5.py run error in line 88: “bs, _, ny, nx, _ = outs[i].shape“ ，perhaps need change to “bs, _, ny, nx, _ = outs[i].shape ” ?

Other, detect decoding is too slow, how too ?

OpenCV(4.1.2) /usr/local/src/opencv/modules/dnn/include/opencv2/dnn/dnn.inl.hpp:349: error: (-204:Requested object was not found) Required argument "starts" not found into dictionary in function 'get'
完全按照大佬说的做的，但是加载失败。

你好，项目需要，但不是很懂得深度学习。转onnx文件出现这个错误。

您好，我刚刚接触C++，请问这个代码如何测试
我在网上查

g++ -o test main_yolo.cpp

可是报错

In file included from main_yolo.cpp:1:
yolo.h:7:10: fatal error: opencv2/dnn.hpp: 没有那个文件或目录
 #include <opencv2/dnn.hpp>
          ^~~~~~~~~~~~~~~~~
compilation terminated.

但是在我电脑里面终端输入

whereis opencv

会得到

opencv: /usr/include/opencv /usr/share/opencv

请问是哪儿存在了问题

祝好
tongchangD

/yolov5-dnn-cpp-python/build$ ./opencv_test_project
Net use yolov5s
terminate called after throwing an instance of 'cv::Exception'
what(): OpenCV(4.5.3-pre) /home/code/opencv/modules/dnn/src/onnx/onnx_importer.cpp:78: error: (-5:Bad argument) Can't read ONNX file: yolov5s.onnx in function 'ONNXImporter'

Aborted (core dumped)

cmake .. make 都没报错
opencv4.5.0

./opencv_test_project: symbol lookup error: ./opencv_test_project: undefined symbol: _ZN2cv3dnn14dnn4_v202009083Net7forwardERKNS_12_OutputArrayERKSt6vectorINSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEESaISC_EE

Traceback (most recent call last):
File "C:/Users/xtuye/yolov5-dnn-cpp-python/main_yolov5.py", line 114, in
dets = yolonet.detect(srcimg)
File "C:/Users/xtuye/yolov5-dnn-cpp-python/main_yolov5.py", line 90, in detect
outs[i] = outs[i].reshape(bs, self.na, self.no, ny, nx).transpose(0, 1, 3, 4, 2)
TypeError: 'tuple' object does not support item assignment

Hi @hpc203
can you please add the conversion of the model
https://github.com/ultralytics/yolov5/blob/master/models/hub/yolov5-fpn.yaml
It's not clear to me how to set the dimensions for BottleneckCSP

Many thanks

看了您的博文，明白了要想将导出的onnx模型供cv2.dnn调用，需要改写官方源码。但作为小白，怎么改写还是不太懂。能将这个详细过程再讲一遍吗？十分感谢。

How to run code? arguments?
where is onnx file?

Then when I infer the *.onnx by above, I get :
cannot reshape array of size 1632000 into shape (1,3,6,80,80)
cannot reshape array of size 408000 into shape (1,3,6,40,40)
cannot reshape array of size 102000 into shape (1,3,6,20,20)
Use YOLOV5 v3.1 yolov5s.pt file.

First of all, I really appreciate your work!
While I was trying to work with your code, I found it hard to understand how you enccoded the raw outputs of the onnx model -
(tx,ty,tw,th) to (bx,by,bh,bw). Can you explain how you did it?

Thanks

C++版本的
boxes.size()是0，检测不到任何物体是怎么回事呢？

在推理阶段 读取模型就可以 ，请问为什么需要原始的9个anchors呢？有什么用呢

你好，你的代码跑通了，模型是yolov5哪个版本。
ultralytics/yolov5-5.0的模型能跑吗，需要改代码吗

Greetings! First off this is great work and I'm happy to maybe finally find a solution to my problem.

Does your code work for own weights as well, which I've gained by training on a custom dataset with YOLOv5? I need to export the .pt to .onnx files, so I can use my own wiehgts with OpenCV DNN.

you mean convert yolov5s.pt to yolov5_param.pth?

is there a difference beteween .pt and .pth?
wow, i hope to recieve your answer。

找了一些原因，可能是由于在原版yolov5中：
# Compute conf
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
但是我发现我这边得到的obj_conf(detection[5])比原版的要低很多，乘下来就只有0.2左右了。
我不是太清楚，这个是怎么计算的，想请教一下您😂
在您的代码中也有类似的描述：
###其实只需要对x,y,w,h做sigmoid变换的， 不过全做sigmoid变换对结果影响不大，因为sigmoid是单调递增函数，那么就不影响类别置信度的排序关系，因此不影响后面的NMS
###不过设断点查看类别置信度，都是负数，看来有必要做sigmoid变换把概率值强行拉回到0到1的区间内

谢谢！

用DNN加载，模型得到的置信度都接近于1.

运行convert_onnx.py时出现下列警告：
Namespace(net_type='yolov5s')
360 360
360 360
H:\yuanbaoxi\ybx_gitee\yolov5-dnn-cpp-py\yolov5s.py:36: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
route = F.interpolate(xRt2, size=(int(xRt2.shape[2] * 2), int(xRt2.shape[3] * 2)), mode='nearest')
H:\yuanbaoxi\ybx_gitee\yolov5-dnn-cpp-py\yolov5s.py:40: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
route = F.interpolate(xRt3, size=(int(xRt3.shape[2] * 2), int(xRt3.shape[3] * 2)), mode='nearest')
convert yolov5s.onnx to onnx finish!!!
然后出现下面错误：
read failed

您好，opencv的cuda加速在yolact上很明显，大概有5倍的提升，但是在yolov5s上，几乎没有变化，想请教您可能是什么原因呢？非常感谢！

string modelFile = this->netname;
modelFile += ".onnx";
cout << "load modelFile : " << modelFile << endl; 
this->net = readNet(modelFile);
// 增加这两行用来加速 
this->net.setPreferableBackend(DNN_BACKEND_CUDA);
this->net.setPreferableTarget(DNN_TARGET_CUDA);

OpenCV(4.0.0) Error: Requested object was not found (Blob 363 not found in const blobs) in cv::dnn::dnn4_v20180917::ONNXImporter::getBlob, file c:\build\master_winpack-build-win64-vc15\opencv\modules\dnn\src\onnx\onnx_importer.cpp, line 272

pytorch1.7.1 cv2 4.3.0 win10
I download the yolov5s.pth then I run the convert_onnx.py,the onnx file ran normally,but in the end of convert_onnx.py dnnnet = cv2.dnn.readNet(output_onnx) failed:error: (-212:Parsing error) Input node with name 873 not found in function 'cv::dnn::Subgraph::getInputNodeId'

1.修改，因为提示参数不多，返回的只有三个参数
#bs, _, ny, nx = outs[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
bs, ny, nx = outs[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
修改参数之后，包这个错误：
ValueError: cannot reshape array of size 151200 into shape (1,3,6,25200,6)

[INFO]blob shape: (1, 3, 640, 640)
0 (1, 25200, 6)
Traceback (most recent call last):
File "main_yolov5.py", line 124, in
dets = yolonet.detect(srcimg)
File "main_yolov5.py", line 99, in detect
outs[i] = outs[i].reshape(bs, self.na, self.no, ny, nx).transpose(0, 1, 3, 4, 2)
ValueError: cannot reshape array of size 151200 into shape (1,3,6,25200,6)

/media/yzl/C4A8E8FBE4769FDE/sr/cpp/opencv-4.2.0/modules/dnn/src/onnx/onnx_importer.cpp:327: error: (-215:Assertion failed) model_proto.has_graph() in function 'populateNet'

为什么会报这样的错误？