WAE-pytorch

Pytorch implementation of WAE-MMD(paper).

Dependencies

python 3.6.4
pytorch 0.3.1.post2
visdom

Usage

  1. download img_align_celeba.zip and list_eval_partition.txt files from here, make data directory, put downloaded files into data, and then run ./preprocess_celeba.sh. for example,
.
└── data
  └── img_align_celeba.zip
  └── list_eval_partition.txt
  1. initialize visdom
python -m visdom.server
  1. run by scripts
sh run_celeba_wae_mmd.sh
  1. check training process on the visdom server
localhost:8097

Results - CelebA

training plots

curves

train data reconstruction

train_recon

test data reconstruction

test_recon

random data generation via sampling z from P(z)

References

  1. Wasserstein Auto-Encoders, Tolstikhin et al, ICLR, 2018
  2. Code repos : official, re-implementation, both in Tensorflow