hiwonjoon / tf-vqvae

Tensorflow Implementation of the paper [Neural Discrete Representation Learning](https://arxiv.org/abs/1711.00937) (VQ-VAE).
258 stars 41 forks source link

Do you have any progress ? #1

Closed gaopeng-eugene closed 6 years ago

hiwonjoon commented 6 years ago

I have no plan yet to reproduce their results; first, there is no available pixel CNN model for ImageNet, and second, I cannot currently figure out what they mean by using the pre-trained network as a prior for z.

Nevertheless, any contributions or discussions are welcome! Thanks.

hiwonjoon commented 6 years ago

Hello,

I implemented sampling procedure by training prior Z using pixelCNN. Please check the code and results. The result for MNIST seems good (as always, everyone is happy with MNIST) while the result for CIFAR10 seems (only) acceptable.

Further comments and contributions are still welcome! Thanks.