Yonghye Kwon

@stigma0617 Actually... I guessed that the results with modified code may have the better results in terms of speed and memory. But, when I checked the execution time, there are...

@lmarval Thanks luis!

It is related to #246.

> @jalaldev1980 > I try to guess, where did you take this _calc_anchors_ flag in your command line? I didn't find it in YOLO-2, > may be, it is in...

Hi @heilaw! How does it make the training more stable?

@capricornfati Refer to https://github.com/developer0hye/Simple-PyTorch-Deformable-Convolution-v2 This code doesn't need any compilation step. However, I didn't test my DCNv2 with this project.

Refer to [developer0hye/Simple-CenterNet](https://github.com/developer0hye/Simple-CenterNet)

@linlinmu If you have some problem with DCNv2, refer to [developer0hye/Simple-PyTorch-Deformable-Convolution-v2](https://github.com/developer0hye/Simple-PyTorch-Deformable-Convolution-v2). You don't need to compile c++ and cuda code for DCNv2.

Should I use `with autocast(False)` here? https://github.com/Megvii-BaseDetection/YOLOX/blob/d78fe47802d3987a915645179d87f5c1a96e646f/yolox/models/yolo_head.py#L394-L401

@blackCmd Sorry, I can't implement this great function **now** because I am too busy recently, but I really want to do it someday. I want someone send a PR related...