stylegan-t

stylegan-t copied to clipboard

Thanks for your great job. I believe that the zero-shot training is conducted using the CC3M datasets as the training data all the same and evaluated on different validation datasets...

landian60

In train.py, line 87 says @click.option('--workers', help='DataLoader worker processes', type=click.IntRange(min=1), default=3) However, line 123 uses num_workers=3 regardless of the opts.workers flag: c.data_loader_kwargs = dnnlib.EasyDict(pin_memory=True, num_workers=3, prefetch_factor=2)

RolandGao

Thanks for your excellent work! I am wonder why the accumulation steps are just multiplied on the one step loss, but not accumulate gradient with multiple steps. https://github.com/autonomousvision/stylegan-t/blob/36ab80ce76237fefe03e65e9b3161c040ae888e3/training/training_loop.py#L186

yijicheng