Zherlock

U can try to create a folder model in the corresponding position.

My implementation is according to detect.py in [https://github.com/ultralytics/yolov5/blob/master/detect.py](url). I'll test batching and reply in days.

@james77777778 Did u replicate the results of this paper? Is it good after quantization

@sudapure hey I add nms.h and nms.cpp in the repo.

@hmmlencat not sure about this, maybe u can try latest libtorch, mine is 1.5. Or u can check ur cmakelist

export.py没有修改，导出时，把detect layer的输出改成输入的x,是个列表。 anchor grid是把python里面的anchor grid打印出来。

@JohnTang93 images in different shapes can be fed. 可以用不同尺寸输入。

@pengyang1225 hey, sorry that I did't try onnx, but I think it is included in export.py. when u export, let the detect layer output the inputed list 'x'.

Like what said in the error, cmake found cuda9.1 but u use cuda10.1. I think u need to fix ur code in cmakelist and set path for cuda.

hi @molyswu Sorry I do not have the environment to build that. I think maybe u could refer to https://github.com/yasenh/libtorch-yolov5.