Seeking Advice on Designing an Invertible Neural Network for Fission about normalizing-flows HOT 2 CLOSED

Hi @wileewang,

I'm glad that you find the normflows package useful.

From what you said about your problem, (discrete) normalizing flows seem to be the ideal model for $z$. They typically are tractable and invertible by design. As an introduction to this class of models, I recommend reading this review. It is very comprehensive and you certainly don't have to know all that's in there, so you could just start with section 2.1 and 3.1.
I recommend you to start with a simple, yet expressive normalizing flow such as Real NVP. It is tractable and has an explicit and fast to evaluate inverse. The only requirement for using it is that $z$ is at least two dimensional, see #58. This notebook can help you get started, and you can even run it in Colab.

If you run into any problems or have more specific questions, feel free to reach out.

Best regards,
Vincent

P.S. Depending on whom you ask, invertible neural networks is just a different term for (discrete) normalizing flows (not continuous normalizing flows or neural ODEs), or they are a subclass of them, i.e. existing neural network architectures such as MLPs or residual networks that are made invertible. The latter often do not have an explicit formula for the inverse, see e.g. residual flows, so what you probably want to use for your application are normalizing flows such as Real NVP.

from normalizing-flows.

Hi all,

I'm closing this issue due to inactivity, so I assume my reply solved it for you.

Best regards,
Vincent

from normalizing-flows.