Implementation of Conditional Variational Auto-encoder

[Cicilalal]

Hi, when I want to implement CVAE, I set use_mmd=False and beta = 0 in Class TRVAE. Is this right? Is there anything else that needs to be changed when loading and training query data?
I have another question. According to the paper "Mapping single-cell data to reference atlases by transfer learning", shouldn't MMD in trvae model be added to the first layer of decoder? Why is MMD added to latent space by default in source code?
When I add MMD to the first layer of the decoder, I don't know why I can't get a good integration result.
Thank you in advance.

[M0hammadL]

Hi, mmd =0 is mimicks a normal cvae. For your second question, we found mmd on Z results in better integration and easier to optmize thus we changed it but for the paper, mmd on first decoder layer were used but as said that is harder to converge and optimize.

[Cicilalal]

I am not sure when I set “use_mmd=False“ and ”beta = 0“（beta is the scaling factor for MMD loss)，is mmd equal to 0？

[M0hammadL]

That is correct

[Cicilalal]

It seems the paper does not explain the reconstruction methods for CVAE(MSE) and CVAE(NB). "MSE" seems to refer to the mean square error between intput and the reconstruction of input, but I don't know how to understand "NB loss". Does it have something in common with the "NB" loss of SCVI?