github-pengge/PyTorch-progressive_growing_of_gans

resize_activation function

Paseam opened this issue · 1 comments

hi,as you say you replace repeat by torch.nn.functional.upsample, but i read the source codes from author, it seems that author uses repeat, do you notice it ?

However, using repeat suffers from severe perturbation when fading in. Besides, the author used repeat to implement nearest neighbor upsampling. You can verify that via numpy.repeat(PyTorch can not achieve this through repeat, since it only repeat at size 1 dim.).