Default Training Config

[tsunghan-wu]

Hi,

First, thanks for your excellent work. I am a master student major in CS who is now working on adversarial training. I would like to ask whether your provided config is able to reach the comparable result as your paper report.

Specifically,

1. In ./config/cityscapes/cityscapes_aspp.yaml, is it reasonable to set test_h = 449, test_w = 449 while testing?
2. In your README, you stated that "For multiprocessing training, we use apex, tested with pytorch 1.0.1.". However, it seems that your config set opt_level: 'O0' and does not use float16 during training.

I would be very grateful if you could provide me any suggestions or further explanation. Thank you very much.

Best, Tsung-Han Wu.