search

EperLuo / scDiffusion

A model developed for the generation of scRNA-seq data
MIT License
27 stars 4 forks source link

Regarding the training of the classifier #3

Open jingfengou opened 5 months ago

jingfengou commented 5 months ago

Did you incorporate the noise from the diffusion forward process during its training? After I added this noise, the classifier became difficult to train. I suspect that the noise is too large arnd masks the features of the original data.

EperLuo commented 5 months ago

Hi! I did add the noise from the diffusion forward process during the training of classifier, but I only added noise from the first 500 steps (totally 1000 steps), which is relatively small and do not masks to much of the data feature. And the guidance will start from the 500th step during inference.

Sbs12 commented 2 months ago

Hello, can your unconditional generation effect achieve the effect described in the article?