On the Effectiveness of Gradient Normalized Generative Adversarial Networks

This is the official implementation of Faster Gradient Normalized GAN (Faster GN-GAN) by the authors.

Requirements

• CUDA 11.3
• Python packages

  pip install -U pip setuptools
  pip install -r requirements.txt

Datasets

• CIFAR-10 and STL-10

  We use the PyTorch built-in dataset for CIFAR-10 and STL-10.

• CelebA-HQ

  We obtain CelebA-HQ from this repository and preprocess them into lmdb format using the following command:

  python -m training.datasets --dataset celebahq/images --out ./data/celebahq
  

• LSUN Church

  We obtain LSUN Church from official website.

Folder Structure

./data
├── celebahq
│   ├── data.mdb
│   └── lock.mdb
├── cifar10 (created by pytorch)
├── lsun
│   └── church_outdoor_train_lmdb
│       ├── data.mdb
│       └── lock.mdb
└── stl10 (created by pytorch)

Preprocessing Datasets for FID

• Download pre-calculated statistic from here to calculating FID.

• The folder structure should be as follows:

  ./stats
  ├── celebahq.all.256.npz
  ├── church.train.256.npz
  ├── cifar10.test.npz
  ├── cifar10.train.npz
  └── stl10.unlabeled.48.npz
  

NOTE

All the values reported in our paper are calculated using the official implementation of Inception Score and FID.

Training

All the configurations can be found in ./configs.

• To train GN-GAN from scratch:

  CUDA_VISIBLE_DEVICES=0 python main.py \
      --config ./config/GN_cifar10_resnet.json \
      --normalize_G training.gn.normalize_D \
      --logdir ./logs/GN_cifar10_resnet_0

• To train Faster GN-GAN from scratch:

  CUDA_VISIBLE_DEVICES=0 python main.py \
      --config ./config/GN_cifar10_resnet.json \
      --normalize_G training.gn.normalize_G \
      --logdir ./logs/GN_cifar10_resnet_0

• To train Faster GN-GAN with rescaling from scratch:

  CUDA_VISIBLE_DEVICES=0 python main.py \
      --config ./config/GN_cifar10_resnet.json \
      --normalize_G training.gn.normalize_G \
      --scale 0 \
      --logdir ./logs/GN_cifar10_resnet_0

• To train GN-GAN with multi-GPU:

  CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py \
      ...