Can I use a GAN-based network to replace the flow-based prior P(Z|X)?

[seekerzz]

If I understand this paper and FlowSeq correctly, the normalizing flow is used to model the dependence of text X (from the posterior P(Z|X, Y)). As GAN can also model the distribution, can I use a GAN-based network to replace the flow-based prior P(Z|X)?

[light1726]

Hi @seekerzz! The issue is that you need the prior to sample z's and inference the probs of z's efficiently. While GAN can do the sampling, it cannot do the inference.

[seekerzz]

Thank you so much for the quick reply!😁 Here is my rough understanding: For P(Z|X, Y), we can use another predicted prob P(Z|X) to get close to it.

1. If we want to model its distribution explicitly, we need to calculate the prob of P(Z|X,Y) and also the prob of P(Z|X) and use KL to make them close. Thus, P(Z|X, Y) is modelled as Gaussian for a simply prob calculation and P(Z|X) calculated by the reversed normalizing flow.
2. Maybe I misunderstand your thought, but I suppose inference the probs of z's is used to do the aforementioned idea (to get close to P(Z|X,Y)). What I think is that we can also use a GAN to get close to P(Z|X,Y): Sampling from a certain noise, combined with X, generating G(Z|X) for fooling the discriminator between P(Z|X,Y) and the generated G(Z|X) such that their distributions are close. Will this be OK or I made some mistakes? Thank you!

[light1726]

Hmm, I was too concerned about the computation of the KL at first glance of your question.

I think your idea is doable. GAN is good at sampling high-quality samples. However, if you sample from GAN and close the distance with P(Z|X, Y), it's would be a reverse KL computation.

But it doesn't mean that it's a bad choice as we care more about the generative quality instead of NLL or ELBO for TTS.

Anyway, I think you can give it a shot. Good luck!

[seekerzz]

Many thanks to you😁😊