h2k/dcgan-lsgan-wgan-gp-dragan-tensorflow-2

DCGAN LSGAN WGAN-GP DRAGAN Tensorflow 2

Recommendation

• Our GAN based work for facial attribute editing - AttGAN.

New

• 8 April 2019: We re-implement these GANs by Tensorflow 2! The old version is here: v1 or in the "v1" directory.
• PyTorch Version

---

---

GANs - Tensorflow 2

Tensorflow 2 implementations of DCGAN, LSGAN, WGAN-GP and DRAGAN.

Exemplar results

Fashion-MNIST

DCGAN	LSGAN	WGAN-GP	DRAGAN

CelebA

DCGAN	LSGAN
WGAN-GP	DRAGAN

Anime

WGAN-GP	DRAGAN

Usage

• Prerequisites

  • Tensorflow 2.0 pip install tensorflow-gpu
  • Tensorflow Addons pip install tensorflow-addons
  • (if you meet "tf.summary.histogram fails with TypeError" pip install --upgrade tb-nightly)
  • scikit-image, oyaml, tqdm
  • Python 3.6

• Datasets

  • Fashion-MNIST will be automatically downloaded
  • CelebA should be prepared by yourself in ./data/img_align_celeba/*.jpg
    • dataset link: Dropbox
    • the above link might be inaccessible, the alternatives are (find "img_align_celeba.zip")
      • Baidu Netdisk or
      • Google Drive
  • the Anime dataset should be prepared by yourself in ./data/faces/*.jpg
    • dataset link: https://pan.baidu.com/s/1eSifHcA, password: g5qa
    • reference: https://zhuanlan.zhihu.com/p/24767059

• Examples of training

  • Fashion-MNIST DCGAN

    CUDA_VISIBLE_DEVICES=0 python train.py --dataset=fashion_mnist --epoch=25 --adversarial_loss_mode=gan

  • CelebA DRAGAN

    CUDA_VISIBLE_DEVICES=0 python train.py --dataset=celeba --epoch=25 --adversarial_loss_mode=gan --gradient_penalty_mode=dragan

  • Anime WGAN-GP

    CUDA_VISIBLE_DEVICES=0 python train.py --dataset=anime --epoch=200 --adversarial_loss_mode=wgan --gradient_penalty_mode=wgan-gp --n_d=5

  • see more training exampls in commands.sh

  • tensorboard for loss visualization

    tensorboard --logdir ./output/fashion_mnist_gan/summaries --port 6006