UNeedCryDear


 if you debug detect.py,you will find it concat the three outputs by the layer Detect(),P3 has shape[1,3,80,80,85],P4:[1.3.40.40.85] and P5:[1.3.20.20.85].The resoult of concat renamed "output" ,witch shape is [1,25200,85] but...

> You'd best train and eval in the same or approximate shape！ I know this situation, but as I said above, in many cases, I always hope to train with...

@sctrueew example by opencv-dnn: https://github.com/UNeedCryDear/yolov5-seg-opencv-dnn-cpp

you can try this pr https://github.com/ultralytics/yolov5/pull/9645