roller44 / vq-vae

Minimalist implementation of VQ-VAE in Pytorch

CVAE and VQ-VAE

This is an implementation of the VQ-VAE (Vector Quantized Variational Autoencoder) and Convolutional Varational Autoencoder. from Neural Discrete representation learning for compressing MNIST and Cifar10. The code is based upon pytorch/examples/vae.

pip install -r requirements.txt
python main.py

requirements

• Python 3.6 (maybe 3.5 will work as well)
• PyTorch 0.4
• Additional requirements in requirements.txt

Usage

# For example
python3 main.py --dataset=cifar10 --model=vqvae --data-dir=~/.datasets --epochs=3

Results

All images are taken from the test set. Top row is the original image. Bottom row is the reconstruction.

k - number of elements in the dictionary. d - dimension of elements in the dictionary (number of channels in bottleneck).

• MNIST (k=10, d=64)

• CIFAR10 (k=128, d=256)

• Imagenet (k=512, d=128)

TODO:

• Implement Continuous Relaxation Training of Discrete Latent Variable Image Models

• Sample using PixelCNN prior

• Improve results on cifar - nearest neighbor should be performed to 10 dictionaries rather than 1

• Improve results on cifar - replace MSE with NLL

• Improve results on cifar - measure bits/dim

• Compare architecture with the offical one

• Merge VAE and VQ-VAE for MNIST and Cifar to one script

Acknowledgement

tf-vaevae for a good reference.