Speed up and tune VQ-VAE hyperparameters for Hugo's binarized MNIST

[keyonvafa]

Currently, Hugo's binarized MNIST is slower and doesn't match the performance of MNIST thresholded at 0.5 in the VQ-VAE example. We should understand why this is happening and tune hyperparameters.

[KeerthanaS04]

I want to work on this issue. Kindly assign this to me