How to compute negative log-likelihood (nll) or bits per dimension (bpd)?

lucidrains / denoising-diffusion-pytorch

Implementation of Denoising Diffusion Probabilistic Model in Pytorch

MIT License

8.05k stars 1.01k forks source link

Closed davitpapikyan closed 3 days ago

davitpapikyan commented 1 year ago

I wonder is there a way to estimate negative log-likelihood or bpd? For reference see Table 1 in Improved Denoising Diffusion Probabilistic Models.

DevashishPrasad commented 3 days ago

Were able to find a good way to compute negative log-likelihood or bpd?

Thank you!

davitpapikyan commented 3 days ago

Sure, check this function.