Learning disentangled representations with the Wasserstein Autoencoder