/MeRGAN

Primary LanguagePythonMIT LicenseMIT

Memory Replay GANs: learning to generate images from new categories without forgetting

The paper has been accepted in NeurIPS 2018. An arXiv pre-print version, a blog and a video are available.

Method

MerGAN

Dependences

Usage

For training:

  • python mergan.py --dataset mnist --result_path mnist_SFT/ Sequential Fine Tuning
  • python mergan.py --dataset mnist --RA --RA_factor 1e-3 --result_path mnist_RA_1e_3/ MeRGAN Replay Alignment
  • python mergan.py --dataset mnist --JTR --result_path mnist_JTR/ MeRGAN Joint Training with Replay
  • python joint.py --dataset mnist --result_path mnist_joint/ Joint Training

For testing:

  • python mergan.py --dataset mnist --test --result_path result/mnist_RA_1e_3/
  • python joint.py --dataset mnist --test --result_path result/mnist_joint/

Results

MNIST

MerGAN

LSUN

MerGAN

References

Citation

Please cite our paper if you are inspired by the idea.

@inproceedings{chenshen2018mergan,
title={Memory Replay GANs: learning to generate images from new categories without forgetting},
author={Wu, Chenshe and Herranz, Luis and Liu, Xialei and Wang, Yaxing and van de Weijer, Joost and Raducanu, Bogdan},
booktitle={Conference on Neural Information Processing Systems (NIPS)},
year={2018}
}