RCF-GAN

This code is for the implementation of RCF-GAN proposed in the paper "Reciprocal Adversarial Learning via Characteristic Functions" that has been accepted as Spotlight in NeurIPS 2020. This repo is a simple implementation on the RCF-GAN, for the ease of understanding basics. We also provide an advanced implementation upon the BigGAN architecture at https://github.com/ShengxiLi/bigrcf_gan. A further application in the graph link prediction task can be found here https://github.com/ShengxiLi/grcf.

RCF-GAN, with a reciprocal theorey, intrinsically combines an auto-encoder and GAN by only two neat modules, together with a simple yet stable training strategy:

The RCF-GAN under the BigGAN architecture at this repo: https://github.com/ShengxiLi/bigrcf_gan

What can the characteristic function control in RCF-GAN?

Requirements

Python, Pytorch and then pip install -r requirements.txt

How to Use

You can run the following example command for (roughly) replicating results in the paper (due to randomness in sampling):

CUDA_VISIBLE_DEVICES=0 python main_parse.py --dataset cifar_train --img_size 32 --mark rcf-gan

Results for other dataset and settings can use the following relationships (* means the default and # for the required things):

#Commands	--dataset#	--img_size#	--z_dim	--net_model	--epochs
cifar-10 dcgan	cifar_train	32	128*	dcgan*	300*
celeba dcgan	celeba	64	64	dcgan*	300*
lsun_bedroom dcgan	lsunb_train	64	128*	dcgan*	500
celeba resnet 64	celeba	64	64	resnet	300*
lsun_bedroom resnet 64	lsunb_train	64	128*	resnet	300*
celeba resnet 128	celeba	128	64	resnet	300*
lsun_bedroom resnet 128	lsunb_train	128	128*	resnet	300*

Other commands that you might use

    '--t_net', type=int, default=64, help='0 to disable t_net.'
    '--bs', type=int, default=64, help='Batch sizes for all.'
    '--alpha', type=float, default=0.5, help='Alpha weight of the amplitude in CF loss.'
    '--lr', type=float, default=2e-4, help='Learning rate.'
    '--resume_training', type=int, default=0, help='Whether to resume training; positive numbers for resuming epochs; negative number for evaluation only.')
    '--save_period', type=int, default=10, help='A period to save checkpoint.'

You could find more adjustable parameters in main_parse.py for your own experiments.

OR you can directly run main_xxx_xxx.py in ./main_files for each setting we used in the paper. You just need to copy them in the root directory and run those.

Directories

• Dataset preparation:
  • For cifar-10 dataset (automatically downloaded): ./data/cifar-10
  • For celeba: ./data/img_align_celeba/img_align_celeba/ *.png
  • For lsun_bedroom (download by refering to this): ./data/bedroom_train_lmdb/ *.mdb
• Automatically created output directories:
  • ./model: checkpoints and final model
  • ./results: middle visualisations and final images
  • ./runs: information for tensorboardX

To Do

• cfgan_model.py: Add data parallel module

• clean everything up and standardise code comments

Terms of Use

The code is provided for research purposes only and without any warranty. Any commercial use is prohibited. When using the code in your research work, you should cite the following paper:

@inproceedings{shengxi2020reciprocal,
 author = {Li, Shengxi and Yu, Zeyang and Xiang, Min and Mandic, Danilo},
 booktitle = {Proceedings of the Advances in Neural Information Processing Systems},
 pages = {217--228},
 publisher = {Curran Associates, Inc.},
 title = {Reciprocal Adversarial Learning via Characteristic Functions},
 volume = {33},
 year = {2020}
}

Special Thanks

• Our evaluation codes in ./metrics on FID and KID scores were almost fully adapted from abdulfatir/gan-metrics-pytorch;
• Our ResNet.py were developed from jalola/improved-wgan-pytorch;
• We also appreciated igul222/improved_wgan_training for providing net structure details on ResNet structures.