The paper has been accepted in NeurIPS 2018. An arXiv pre-print version, a blog and a video are available.
For training:
python mergan.py --dataset mnist --result_path mnist_SFT/
Sequential Fine Tuningpython mergan.py --dataset mnist --RA --RA_factor 1e-3 --result_path mnist_RA_1e_3/
MeRGAN Replay Alignmentpython mergan.py --dataset mnist --JTR --result_path mnist_JTR/
MeRGAN Joint Training with Replaypython joint.py --dataset mnist --result_path mnist_joint/
Joint TrainingFor 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/
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}
}