Han Zhang, Ian Goodfellow, Dimitris Metaxas and Augustus Odena, "Self-Attention Generative Adversarial Networks." arXiv preprint arXiv:1805.08318 (2018).

This repository provides a PyTorch implementation of SAGAN. Both wgan-gp and wgan-hinge loss are ready, but note that wgan-gp is somehow not compatible with the spectral normalization. Remove all the spectral normalization at the model for the adoption of wgan-gp.

Self-attentions are applied to later two layers of both discriminator and generator.

• Supervised setting
• Tensorboard loggings
• [20180608] updated the self-attention module. Thanks to my colleague Cheonbok Park! see 'sagan_models.py' for the update. Should be efficient, and run on large sized images
• Attention visualization (LSUN Church-outdoor)
• Unsupervised setting (use no label yet)
• Applied: Spectral Normalization, code from here
• Implemented: self-attention module, two-timescale update rule (TTUR), wgan-hinge loss, wgan-gp loss

• Python 3.5+
• PyTorch 0.3.0

$ git clone https://github.com/heykeetae/Self-Attention-GAN.git
$ cd Self-Attention-GAN

$ bash download.sh CelebA
or
$ bash download.sh LSUN

$ python python main.py --batch_size 64 --imsize 64 --dataset celeb --adv_loss hinge --version sagan_celeb
or
$ python python main.py --batch_size 64 --imsize 64 --dataset lsun --adv_loss hinge --version sagan_lsun

$ cd samples/sagan_celeb
or
$ cd samples/sagan_lsun

Samples generated every 100 iterations are located. The rate of sampling could be controlled via --sample_step (ex, --sample_step 100).