Skip to content

Self-Attention Generative Adversarial Networks Implementation in PyTorch

License

Notifications You must be signed in to change notification settings

DanyangYue/sagan-pytorch

 
 

Repository files navigation

sagan-pytorch

Self-Attention Generative Adversarial Networks (SAGAN, https://arxiv.org/abs/1805.08318) in PyTorch

Usage:

python train.py PATH

Input directory should be structured like this (as with torchvision.datasets.ImageFolder):

PATH/class1
PATH/class2
...

Code for evaulate FID score came from https://github.com/bioinf-jku/TTUR

Notes

Sample

Sample from DCGAN generator (without residual connection) at 120k iteration. Seems like that model size is insufficient. FID is about 120. After this model is collapsed.

ResNet Sample

Sample from ResNet generator 290k iteration. FID is about 64.8. Used imbalanced learning schedule for generator vs discriminator, namely, 1:5 update schedule. Sample quality seems like improved, might because increased model size and more stable learning schedule. Train resnet model with 1:1 update schedule was hard and unstable.

About

Self-Attention Generative Adversarial Networks Implementation in PyTorch

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%