This is a re-implementation in Tensorflow 2 (2.2.0) of the WAE-GAN model proposed in 1 using the MNIST dataset.

These are a few samples decoded from the latent space. The quality of the generation is inferior to the original paper's implementation.

Even rows contain original samples while odd rows their relative reconstruction.

git clone https://github.com/w00zie/wae_mnist.git
cd wae_mnist

I provide a few (at the moment only one) trained models. You can explore the saved_model/ directory, where, for each experiment, I provide the weights (models/), the hyperparameters (config.json) and the Tensorboard logs (events.out.tfevents....) that you can load by running tensorboard --logdir saved_model/{experiment}/.

Take a look at test.py for inference and if you want to train a new model just open main.py, set your hparams and run

python main.py

tensorflow >= 2.2.0
matplotlib