search

kwotsin / mimicry

[CVPR 2020 Workshop] A PyTorch GAN library that reproduces research results for popular GANs.
MIT License
601 stars 62 forks source link

SSGAN missed Self-Modulation Normalization? #18

Open shimopino opened 4 years ago

shimopino commented 4 years ago

Hi! Thank you for creating this easy-to-use library!

I was using your implementation of SSGAN to reproduce the original paper https://arxiv.org/abs/1810.01365.

As I read the original paper, the highest performing model was the one that used Self-Modulation Normalization, not Batch Normalization like your GBlock https://github.com/kwotsin/mimicry/blob/master/torch_mimicry/modules/resblocks.py#L73.

image

image

Do you have any plans to add Self-Modulation Normalization?

kwotsin commented 4 years ago

Hi @KeisukeShimokawa, yes indeed I mostly compared to the version without sBN to keep the design simple as sBN is a general architectural improvement, but indeed it will be an interesting addition to have. Since it's suitable for most unconditional GAN I think it should be a good addition. Will keep this issue open and update in the future!

shimopino commented 4 years ago

Excellent! Thanks!