This repository contains the code for reproducing the paper: Training GANs with Stronger Augmentations via Contrastive Discriminator by Jongheon Jeong and Jinwoo Shin.

TL;DR: We propose a novel discriminator of GAN showing that contrastive representation learning, e.g., SimCLR, and GAN can benefit each other when they are jointly trained.

Qualitative comparison of unconditional generations from GANs on high-resoultion, yet limited-sized datasets of AFHQ-Dog (4739 samples), AFHQ-Cat (5153 samples) and AFHQ-Wild (4738 samples) datasets.

An overview of Contrastive Discriminator (ContraD). The representation of ContraD is not learned from the discriminator loss (L_dis), but from two contrastive losses (L _con and L-_con), each is for the real and fake samples, respectively. The actual discriminator that minimizes L_dis is simply a 2-layer MLP head upon the learned contrastive representation.

Currently, the following environment has been confirmed to run the code:

• python >= 3.6
• pytorch >= 1.6.0 (See https://pytorch.org/ for the detailed installation)
• tensorflow-gpu == 1.14.0 to run test_tf_inception.py for FID/IS evaluations
• Other requirements can be found in environment.yml (for conda users) or environment_pip.txt (for pip users)

#### Install dependencies via conda.
# The file also includes `pytorch`, `tensorflow-gpu=1.14`, and `cudatoolkit=10.1`.
# You may have to set the correct version of `cudatoolkit` compatible to your system.
# This command creates a new conda environment named `contrad`.
conda env create -f environment.yml

#### Install dependencies via pip.
# It assumes `pytorch` and `tensorflow-gpu` are already installed in the current environment.
pip install -r environment_pip.txt

By default, the code assumes that all the datasets are placed under data/. You can change this path by setting the $DATA_DIR environment variable.

CIFAR-10/100 can be automatically downloaded by running any of the provided training scripts.

CelebA-HQ-128:

1. Download the CelebA-HQ dataset and extract it under $DATA_DIR.
2. Run third_party/preprocess_celeba_hq.py to resize and split the 1024x1024 images in $DATA_DIR/CelebAMask-HQ/CelebA-HQ-img:

   python third_party/preprocess_celeba_hq.py
   

AFHQ datasets:

1. Download the AFHQ dataset and extract it under $DATA_DIR.
2. One has to reorganize the directories in $DATA_DIR/afhq to make it compatible with torchvision.datasets.ImageFolder. Please refer the detailed file structure provided in below.

The structure of $DATA_DIR should be roughly like as follows:

$DATA_DIR
├── cifar-10-batches-py   # CIFAR-10
├── cifar-100-python      # CIFAR-100
├── CelebAMask-HQ         # CelebA-HQ-128
│   ├── CelebA-128-split  # Resized to 128x128 from `CelebA-HQ-img`
│   │   ├── train
│   │   │   └── images
│   │   │       ├── 0.jpg
│   │   │       └── ...
│   │   └── test
│   ├── CelebA-HQ-img     # Original 1024x1024 images
│   ├── CelebA-HQ-to-CelebA-mapping.txt
│   └── README.txt
└── afhq                  # AFHQ datasets
    ├── cat
    │   ├── train
    │   │   └── images
    │   │       ├── flickr_cat_00xxxx.jpg
    │   │       └── ...
    │   └── val
    ├── dog
    └── wild

We provide training scripts to reproduce the results in train_*.py, as listed in what follows:

The samples below demonstrate how to run each script to train GANs with ContraD. More instructions to reproduce our experiments, e.g., other baselines, can be found in EXPERIMENTS.md. One can modify CUDA_VISIBLE_DEVICES to further specify GPU number(s) to work on.

# SNDCGAN   ContraD on CIFAR-10
CUDA_VISIBLE_DEVICES=0 python train_gan.py configs/gan/cifar10/c10_b512.gin sndcgan \
--mode=contrad --aug=simclr --use_warmup

# StyleGAN2   ContraD on CIFAR-10 - it is OK to simply use `train_stylegan2.py` even with ContraD
python train_stylegan2.py configs/gan/stylegan2/c10_style64.gin stylegan2 \
--mode=contrad --aug=simclr --lbd_r1=0.1 --no_lazy --halflife_k=1000 --use_warmup

# Nevertheless, StyleGAN2   ContraD can be trained more efficiently with `train_stylegan2_contraD.py` 
python train_stylegan2_contraD.py configs/gan/stylegan2/afhq_dog_style64.gin stylegan2_512 \
--mode=contrad --aug=simclr_hq --lbd_r1=0.5 --halflife_k=20 --use_warmup \
--evaluate_every=5000 --n_eval_avg=1 --no_gif 

• The script test_gan_sample.py generates and saves random samples from a pre-trained generator model into *.jpg files. For example,

  CUDA_VISIBLE_DEVICES=0 python test_gan_sample.py PATH/TO/G.pt sndcgan --n_samples=10000
  

  will load the generator stored at PATH/TO/G.pt, generate n_samples=10000 samples from it, and save them under PATH/TO/samples_*/.

• The script test_gan_sample_cddls.py additionally takes the discriminator, and a linear evaluation head obtained from test_lineval.py to perform class-conditional cDDLS. For example,

  CUDA_VISIBLE_DEVICES=0 python test_gan_sample_cddls.py LOGDIR PATH/TO/LINEAR.pth.tar sndcgan
  

  will load G and D stored in LOGDIR, the linear head stored at PATH/TO/LINEAR.pth.tar, and save the generated samples from cDDLS under LOGDIR/samples_cDDLS_*/.

• The script test_lineval.py performs linear evaluation for a given pre-trained discriminator model stored at model_path:

  CUDA_VISIBLE_DEVICES=0 python test_lineval.py PATH/TO/D.pt sndcgan
  

• The script test_tf_inception.py computes Fréchet Inception distance (FID) and Inception score (IS) with TensorFlow backend using the original code of FID available at https://github.com/bioinf-jku/TTUR. tensorflow-gpu <= 1.14.0 is required to run this script. It takes a directory of generated samples (e.g., via test_gan_sample.py) and an .npz of pre-computed statistics:

  python test_tf_inception.py PATH/TO/GENERATED/IMAGES/ PATH/TO/STATS.npz --n_imgs=10000 --gpu=0 --verbose
  

  A pre-computed statistics file per dataset can be either found in http://bioinf.jku.at/research/ttur/, or manually computed - you can refer third_party/tf/examples for the sample scripts to this end.

@inproceedings{jeong2021contrad,
  title={Training {GAN}s with Stronger Augmentations via Contrastive Discriminator},
  author={Jongheon Jeong and Jinwoo Shin},
  booktitle={International Conference on Learning Representations},
  year={2021},
  url={https://openreview.net/forum?id=eo6U4CAwVmg}
}