ispamm / realnvp-demo-pytorch Goto Github PK

Thanks for your demo. It really helps a newbie to learn. I have a few questions:
I found that you introduce label information in training. In both foward and backward steps, the encoded labels are involved. I was wondering if all the input data as [one_hot_label, embedding] are mapped into a single predefined Gaussian distribution. Is it feasible to define different Gaussian distributions to map embeddings from different classes ?
Since you are able to generate image for different digits by [one_hot_label, sampled_data], how does the label information guide the model to generate class-specific embeddings?
Any hints or resources on learning NF with label information would be appreciated !