Arthur-zfc

Yes，I have the same problem in pytorch1.7.1 with cuda 10.2.I also want the solutions.

@shuuny-matrix No, I haven't solved it yet

> 我是用的是单机8卡，在general.py的output_to_target方法中 > 将这两句注释if isinstance(output, torch.Tensor): > output = output.cpu().numpy() > 并加上 output = [temp.cpu().numpy() for temp in output] Thank you!