search

kwotsin / mimicry

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

Plans on adding SAGAN? #3

Closed nivha closed 4 years ago

nivha commented 4 years ago

Great repository! any plans on adding SAGAN (https://arxiv.org/pdf/1805.08318.pdf)? Thanks

kwotsin commented 4 years ago

Yes indeed! It's one of the GANs I'm considering to add in the future, will keep this issue open and update it when it's done!

zhangmozhe commented 4 years ago

Also Biggan if possible?

kwotsin commented 4 years ago

@zhangmozhe Thanks for the suggestion! I'll look into BigGAN as well although I might not have enough resources to train it using the hyperparameters suggested (will check again)

nivha commented 4 years ago

@zhangmozhe in the case of BigGAN there's an "official unofficial" pytorch implementation by the authors written for reproducing their results

kwotsin commented 4 years ago

Hi everyone, I've implemented SAGAN (https://github.com/kwotsin/mimicry/blob/master/torch_mimicry/nets/sagan/sagan_128.py) but it seems since the original paper only trained on imagenet 128x128 that requires about 2 months to train using 1 GPU (1M iterations), I might not have the resources to reasonably train it on time. However, I provided the implementation in the master branch which you can install to check it out, and if you have the resources to train the model, please feel free to share the checkpoint weights!

I also included an extra 32x32 version of SAGAN which I think is easier to train, although of course the implementation is "unofficial" since the placement of the self-attention block may not be optimal, and there's no reference implementation.

Will keep this issue closed for now but will update/reopen if there are any changes. Thanks everyone!