Replicating the bits / dim results in the original VQ-VAE paper

[pclucas14]

Hi,

thanks for your clean implementation ! I was wondering, have you ever tried to calculate the bits / dimension metric (as in the original paper) ? I've tried to do so using the provided code, and I'm still quite far from the results in the paper. I was hoping maybe you would have some insight to share as to why that is the case.

Thanks!
Lucas

[kashif]

@pclucas14 I haven't tried to do that but plan to... give me a few days to figure it out... I am trying to compile a table of bits/dim for MNIST/ Fashion-MNIST so I might as well start with this

[pclucas14]

great! let me know what you end up getting :)

[kashif]

Thanks! Here is where I am collecting the results: zalandoresearch/fashion-mnist#144

[pclucas14]

great! it's a very complete list, pretty cool :D

[pclucas14]

The best I got for VQ-VAE on CIFAR-10 was 5.00 BPD (using a unigram prior). This is still far from the results in the VQ-VAE paper. Hopefully you can manage to close the gap even more :)

[pvitoria]

Hi @kashif ,
could you provide the code you use to compute the bpd?

Thanks!

[podlipensky]

Wonder if there is code for proper bpd computation? If so, can you please share?