Wang Xinyu

最后一层在pytorch是adaptive pooling，这边用的固定kernel，需要根据input size改一下

不是，是改tensorrtx里面最后的pooling层的kernel size

对，pool3输出是1x1，所以要根据前一层的shape改

yes, yolov4.pt should also work. Follow these steps https://github.com/wang-xinyu/tensorrtx/tree/master/yolov4#excute, not using this repo, but using https://github.com/ultralytics/yolov3

Follow these steps https://github.com/wang-xinyu/tensorrtx/tree/master/yolov4#excute, to convert yolov4.pt file into tensorrt

gen_wts.py can also convert yolov4.pt to yolov4.wts. Can you try it ?

@YonghaoHe Thanks a lot. For the range[64, 512], does it mean the plate width in the image? For example: - the original image size is 2048x1536, while plate is around...

@YonghaoHe Yes, I understand, thanks a lot!

Code changes is required, you need to modify the build_engine() function in yolov5.cpp.

We don't have such plan to add this. But PR is welcomed.