https://github.com/JiaRenChang/PSMNet
finetune.py is used to run kitti, main is used to run Scene Flow, the --no-cuda parameter should be set to false, so that it can run on gpu
The original code batch_size is 12, which is too large. I adjusted it in TrainImgLoader. I have been adjusting it in main and found that it has no effect. It should be adjusted in finetune.
There is a problem with no_grad() of with torch.no_grad(), now you can only choose to log out first, without testing