songweige/scdgn

Supervised Adversarial Alignment of Single-Cell RNA-seq Data

scdgn's Issues

After reading the paper, I have some questions.

Classification Loss and Domain Loss are being checked in the process of training the model. (I applied it to other datasets.)

Classification Loss is gradually decreased when reading the paper and looking at the code.

I think that Domain Loss should be increased gradually due to the Gradient Reverse Layer (GRL).

However, Domain Loss shows a form that increases to some extent and then continues to decrease.

I have two thoughts, so I left a question.

1. I misunderstood the code and used it.
2. Encoder (Feature Extractor) has lost information on domain, but loss is reduced because Discriminator learns well about its output.

By any chance, I wonder what kind of trend Domain Loss showed when you were learning.
Also, In case 2, how can I check if information according to the domain is lost?

Thank You!