isjakewong / infodiffusion Goto Github PK

Hi,

Could you please give me some insights about the reconstruction loss in the model? Which part of the loss in the article does this
recon loss correspond to? Besides, why do you utilize x0 and epsilon_theta(xt) to obtain x0 in this loss?

Thank you for your attention.

Hi yingheng, thanks for sharing the code, I have two questions regarding 'Custom prior and Discrete latents.'

First, regarding the Custom prior mentioned in the article, does it mean that different datasets correspond to different priors? If so, could you please specify which prior was used for the dataset tested in the article? Also, when choosing different priors (Gaussian, Gaussian mixture, Swiss roll), we should be able to sample directly from these different priors, as stated in your Appendix D. However, this is not reflected in the code, where it seems to directly use Normal distribution for sampling.

Second, concerning Discrete latent, Appendix G describes how to handle discrete hidden features, but I didn't find any mention of Gumbel Softmax in the code.

Lastly, would it be possible for you to publish the hyperparameter settings for several datasets, including both training and testing (FID and DCI)?