One theoretical question.

[Mr-Nobody-dey]

In all GAN training, their is one generator and one discriminator.

What will happen if I train such model in two stages.

Note: The generator has two part, 1. encoder-decoder part, 2. Id injection part.

Stage 1: In first stage, I train the generator to recreate the same input with any simple loss say MAE. Add id of same person.
This is possible, since it is a supervised problem.

Once the first stage is complete.

Stage 2: In the second stage, we freeze the encoder-decoder part, and train the model as it is generally trained (such as here).

Is their anything fundamentally wrong, that I am missing?

By the way, in line 395 of train_adapter.py, I think id_emb will be the other one, "nn.utils.clip_grad_norm_(ID_emb.parameters(), 2.0)".

[bone-11]

I don't quite understand your point accurately. Generally, GAN training is divided into two stages. For example, in the first stage, the generator is fixed and only the discriminator is trained, while in the second stage, the discriminator is fixed and the generator is trained. The reason for dividing it into two stages is that the objective losses are different.

[Mr-Nobody-dey]

Never mind, I was just wandering.