Open xiaoyanLi629 opened 3 years ago
I am a user just like you, but I can answer your question. It is all explained in the referenced paper. There is the usual VAE loss function involving reconstruction and KLD for the variational vs z prior distribution. Then there is the MMD loss, having to do with the desire to separate the hidden distribution for samples with different conditions.
Hello,.
I have a question about how the model was trained? I saw two losses in the compile_models function in the _trave.py file. I also saw two outputs in the model, reconstruction_output and mmd_output. It looks like to me these two outputs are parallel.
Could you explain a little bit how the training works? Are these losses alternated minimized? Are these two losses correlated to each other (minimize one loss will affect the other one)?
Thank you!