josephdviviano/wae

ift6135 final project

Wasserstein Auto-Encoders (WAE) and Variational Auto-Encoders (VAE) for Celeb A and MNIST Datasets

code built on top of this implementation and this one.

• preprocess.py -- generates preprocessed data (from raw/, to data/).
• wae.py -- trains WAE model.
• inception.py -- InceptionV3 model used for Frechet Inception Distance measure (FID).
• fid_score.py -- Methods for calculating FID score on pairs of images.

Pytorch replication of the results presented in Tolstikhin, Bousquet, Gelly, Schoelkopf (2017).

resources

• https://github.com/tolstikhin/wae
  • Official tensorflow implementation.
• https://github.com/paruby/Wasserstein-Auto-Encoders
  • clean WIP code for WAE (being used for some new paper).
• https://github.com/schelotto/Wasserstein_Autoencoders
  • Another simple implementation.
• https://github.com/maitek/waae-pytorch/blob/master/WAAE.py
  • this is WAE in an adversarial setting -- not quite what we need but some aspects of this code are very clean.
• https://github.com/wsnedy/WAE_Pytorch/blob/master/wae_for_mnist.py
  • simple WAE implementation in pytorch for MNIST.
• https://github.com/wohlert/semi-supervised-pytorch/blob/master/examples/notebooks/Variational%20Autoencoder.ipynb
  • nice explaination of normal VAE.
• https://github.com/sbarratt/inception-score-pytorch
  • inception score
• https://github.com/mseitzer/pytorch-fid
  • frechet inception distance

** todo **

• FID measure
• Blur measure