Code for a Multi-Hinge Loss with K+1 GANs

This is the implementation for the multi-hinge loss of this paper for K+1 GANs. For the main MultiHingeGAN repository please see:https://github.com/ilyakava/gan

This repository is forked from: https://github.com/ajbrock/BigGAN-PyTorch

Please cite:

@inproceedings{kavalerov2020study,
  title={A study of quality and diversity in K+ 1 GANs},
  author={Kavalerov, Ilya and Czaja, Wojciech and Chellappa, Rama},
  booktitle={''I Can't Believe It's Not Better!''NeurIPS 2020 workshop},
  year={2020}
}

Summary

CIFAR10 best IS & FID are 9.58 & 6.40, CIFAR100 best IS & FID are 14.36 & 13.32, and STL10 best IS & FID are 12.16 & 17.44.

Installation

Tested on

Python 3.7.3, pytorch '1.0.1.post2', tensorflow '1.13.1', cuda 10.0.130, cudnn 7.5.0, on rhel linux 7.7 (Maipo).

Also tested on this docker image.

Additional packages

pip install scipy h5py tqdm

Running

CIFAR100

Train with:

sh scripts/final/launch_cifar100_mhingegan.sh

After adjusting read dir --data_root and write dirs --weights_root, --logs_root, --samples_root. Samples will be created automatically in --samples_root.

Using the historical saves run:

sh scripts/final/sample_cifar100_mhingegan.sh

Which saves IS/FID numbers into scoring_hist.npy in the --weights_root + --experiment_name directory. You can make plots like:

Get the same numbers for the BigGAN baseline via:

sh scripts/final/launch_cifar100_baseline.sh
sh scripts/final/sample_cifar100_baseline.sh

CIFAR10

Train with:

sh scripts/final/launch_cifar_mhingegan.sh

Samples will be created automatically in --samples_root.

Using the historical saves run:

sh scripts/final/sample_cifar_mhingegan.sh

Get the same numbers for the BigGAN baseline via:

sh scripts/final/launch_cifar_baseline.sh
sh scripts/final/sample_cifar_baseline.sh

STL10 48x48

Train with:

sh scripts/final/launch_stl48_mhingegan.sh

Samples will be created automatically in --samples_root.

Using the historical saves run:

sh scripts/final/sample_stl48_mhingegan.sh

Get the same numbers for the BigGAN baseline via:

sh scripts/final/launch_stl48_baseline.sh
sh scripts/final/sample_stl48_baseline.sh

Accuracy plots

In scripts/final/sample*.py change:

--sample_np_mem \
--official_IS \
--official_FID \

to --get_train_error or --get_test_error or --get_self_error or --get_generator_error. For --get_generator_error adjust your paths in sample.py to one of the pretrained models. Or add your own. See here for explanations of these metrics.

Training classification networks:

• Cifar10/100
• STL

Links to pretrained models

Google Drive

Attribution

Forked from: https://github.com/ajbrock/BigGAN-PyTorch Official IS code copied from: https://github.com/openai/improved-gan Official FID code copied from: https://github.com/bioinf-jku/TTUR Classification densenet copied from: https://github.com/kuangliu/pytorch-cifar Classification wideresnet copied from: https://github.com/YU1ut/MixMatch-pytorch