search

DavidRuhe / simple-variational-diffusion-models

A simple implementation of Kingma et al., 2021
MIT License
34 stars 4 forks source link

How to calculate the density p(x) of an example x? #2

Open ZouXinn opened 2 years ago

ZouXinn commented 2 years ago

Hi, I just read the paper "Variational Diffusion Models" but the paper just shows an integral to calculate p(x), which is impossible. I wonder how to calculate the density p(x) of example x in practice? Is it right to do the forward and reverse process for N times (where N is large enough) and use the average of the N examples to estimate Equation (6) in "Variational Diffusion Models"?

I would be appreciated it if you can show me your understanding of this question.

DavidRuhe commented 2 years ago

I think only a lower bound is computed by estimating the integrals through sampling. That's why some models in Table 1 are annotated with "variational bound". The lower bound simply amounts to the lower bound that is also used for training. So yes, I think what you are suggesting is correct.

ZouXinn commented 2 years ago

I think only a lower bound is computed by estimating the integrals through sampling. That's why some models in Table 1 are annotated with "variational bound". The lower bound simply amounts to the lower bound that is also used for training. So yes, I think what you are suggesting is correct.

Oh, I just ignored the "variational bound" term in Table 1. Thank you for your advice!