How could we combine two trained model (deep learning network: GAN network with CNN and ResNet)

Question

I have a GAN network (Generative Adversarial Network), consisting of some CNN, ResNet as the structure. I was wondering if I could combine two trained models into one model that maintains functions of model1 and mode2 as before.

I have two training datasets 1 and 2 (denote them as T1 and T2). After I train the GAN with T1, I get model 1 (denoted as M1). Then, I continue to train M1 with T2 (In this case, T2 is new) to get model 2 (M2).

What should I do if I want to combine M1 and M2 ? Should I put 50% of T1 and 50% of T2 and continue to train M2 ? Is there a better way to combine both of them ?

Answer 1

What should I do if I want to combine M1 and M2 ? Should I put 50% of T1 and 50% of T2 and continue to train M2 ?

According to your statement, you have already trained M1 on T1, so you don't need to make 50% of T1 reappear in the training set for M2. Maybe you can achieve a sort of transfer learning after pre-training M1 on T1 and fine-tuning it on T2 to get M2.

Is there a better way to combine both of them ?

An other way is maybe to think about Two-Stream GANs. Check this paper.