Skip to content

[NeurIPS‘2021] "TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up", Yifan Jiang, Shiyu Chang, Zhangyang Wang

License

Notifications You must be signed in to change notification settings

VITA-Group/TransGAN

Repository files navigation

TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up

Code used for TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up.

Implementation

  • checkpoint gradient using torch.utils.checkpoint
  • 16bit precision training
  • Distributed Training (Faster!)
  • IS/FID Evaluation
  • Gradient Accumulation
  • Stronger Data Augmentation
  • Self-Modulation

Guidance

Cifar training script

python exp/cifar_train.py

I disabled the evaluation during training job as it causes strange bug. Please launch another evaluation job simultaneously by copying the path to test script.

Cifar test

First download the cifar checkpoint and put it on ./cifar_checkpoint. Then run the following script.

python exp/cifar_test.py

Main Pipeline

Main Pipeline

Representative Visual Results

Cifar Visual Results Visual Results

README waits for updated

Acknowledgement

Codebase from AutoGAN, pytorch-image-models

Citation

if you find this repo is helpful, please cite

@article{jiang2021transgan,
  title={Transgan: Two pure transformers can make one strong gan, and that can scale up},
  author={Jiang, Yifan and Chang, Shiyu and Wang, Zhangyang},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

About

[NeurIPS‘2021] "TransGAN: Two Pure Transformers Can Make One Strong GAN, and That Can Scale Up", Yifan Jiang, Shiyu Chang, Zhangyang Wang

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published