About the args.batch_size_val about semseg

[hszhao]

1. Thanks for the issue. I updated the default value of batch_size_val in the config files.
2. We follow the previous DeepLab in Caffe where the crop size need to be 8*n+1 (this is due to the implementation of the interp' layer that needs to do align corners). 8*n is also fine now in PyTorch interpolate' function.
3. At the time of the development of this repo, sync bn is not included in the official PyTorch. You can use PyTorch 1.1 or newer versions with sync bn incorporated.

[sunpeng1996]

Hi, hengshuang:
why input size in pspnet must be 8*n+1?

[sunpeng1996]

And why use Apex? Is it better than original pytorch-distributed？？