This is the main folder to reproduce our paper (Disentangled Recurrent Wasserstein Autoencoder).

Currently, we mainly provide the classes to train and test our stochastic moving MNIST result. For Sprites and TIMIT, the training and testing are very similar. You can straightforwardly change the data loader with simple modification of some hyper params. For MUG facial dataset, it is also easy to modify our architecture to train the model by checking the MUG architecture we provided on appendix.

To generate the stochastic moving MNIST video dataset, you can check "moving_mnist.py" to generate the video dataset. The main function "main_MNIST_DRWAE.py" provides the class to train our D-RWAE. You can also visualize some test results inside its functions.

The evaluation code is provided in "Metrics" folder.

If you have any question about running our code, please email us. Thanks for your time!
