search

enazoe / yolo-tensorrt

TensorRT8.Support Yolov5n,s,m,l,x .darknet -> tensorrt. Yolov4 Yolov3 use raw darknet *.weights and *.cfg fils. If the wrapper is useful to you,please Star it.
MIT License
1.18k stars 313 forks source link

Correct cv4 act for bottleneck_csp #70

Closed junhua-zhang closed 3 years ago

junhua-zhang commented 3 years ago

Activation is changed from hardswish to leakyrelu. The inference precision problem would be solved.

enazoe commented 3 years ago

是的，我知道这个情况，的确是用leakyrelu效果好，但是用leakyrelu是yolov5 v2.0版本用的，v3.0 使用的hardswish效果的确不好，性能有问题 https://github.com/enazoe/yolo-tensorrt/blob/master/modules/trt_utils.cpp#L819 https://github.com/enazoe/yolo-tensorrt/blob/master/modules/trt_utils.h#L221 https://github.com/enazoe/yolo-tensorrt/blob/yolov5-2.0/modules/trt_utils.h#L218

enazoe commented 3 years ago

@junhua-zhang 你可以加群，进一步讨论，效率高点，群号在readme

junhua-zhang commented 3 years ago

@enazoe 看你改好了