incremental-learning / mergan

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

Dependences

• Python2.7, NumPy, SciPy, NVIDIA GPU, Tensorflow 1.4
• Dataset:
  • MNIST
  • SVHN(http://ufldl.stanford.edu/housenumbers/)
  • LSUN(bedroom, kitchen, church outdoor, tower)(http://lsun.cs.princeton.edu/2017/)

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

LSUN

References

• [1] 'Improved Training of Wasserstein GANs' by Ishaan Gulrajani et. al, https://arxiv.org/abs/1704.00028, (https://github.com/igul222/improved_wgan_training)[code]
• [2] 'Transferring GANs generating images from limited data' by Yaxing Wang et. al, https://arxiv.org/abs/1805.01677, (https://github.com/yaxingwang/Transferring-GANs)[code]

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}
}