GAN Memory for Lifelong learning

This is a pytorch implementation of the NeurIPS paper GAN Memory with No Forgetting.

Please consider citing our paper if you refer to this code in your research.

@article{cong2020gan,
  title={GAN Memory with No Forgetting},
  author={Cong, Yulai and Zhao, Miaoyun and Li, Jianqiao and Wang, Sijia and Carin, Lawrence},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Requirement

python=3.7.3
pytorch=1.2.0

Notes

The source model is based on the GP-GAN.

GANMemory_Flowers.py is the implementation of the model in Figure1(a).

classConditionGANMemory.py is the class-conditional generalization of GAN memory, which is used as pseudo rehearsal for a lifelong classification as shown in Section 5.2.

Lifelong_classification.py is the code for the lifelong classification part as shown in Section 5.2.

Usage

First, download the pretrained GP-GAN model by running download_pretrainedGAN.py. Note please change the path therein.

Second, download the training data to the folder ./data/. For example, download the Flowers dataset from: https://www.robots.ox.ac.uk/~vgg/data/flowers/102/ to the folder ./data/102flowers/.

Dataset preparation

data
├──102flowers
           ├──all8189images
├── CelebA
...

Finally, run GANMemory_Flowers.py.

The FID scores of our method shown in Figure 1(b) are summerized in the following table.

Dataset	5K	10K	15K	20K	25K	30K	35K	40K	45K	50K	55K	60K
Flowers	29.26	23.25	19.73	17.98	17.04	16.10	15.93	15.38	15.33	14.96	15.19	14.75
Cathedrals	19.78	18.32	17.10	16.47	16.15	16.33	16.08	15.94	15.78	15.60	15.64	15.67
Cats	38.56	25.74	23.14	21.15	20.80	20.89	19.73	19.88	18.69	18.57	17.57	18.18

For lifelong classification

1. run classConditionGANMemory.py for each task until the whole sequeence of tasks are remembered and save the generators;

2. run Lifelong_classification.py to get the classification results.

3. run Compression_low_rank_six_butterfly.py to get the compression results.

Note, for the sake of simplicity, we devide the pseudo rehearsal based lifelong classification processes into above two stages, one can of course find a way to merge these two stages to form a learning process along task sequence.

Acknowledgement

Our code is based on GAN_stability: https://github.com/LMescheder/GAN_stability from the paper Which Training Methods for GANs do actually Converge?.