Why can my conditional Masked Autoregressive Flow not learn multimodal posteriors?

[Elia Zonta]

Hello everyone,

I am trying to learn posterior conditional density distributions of the form p(x|y) with conditional Masked Autoregressive Flows (MAFs). My problem is that I cannot seem to get the model to estimate multimodal posteriors, since it always returns a Gaussian.

I am unsure if this is a misconception on my part, or if I am doing anything wrong in the model architecture / training / etc. If anyone here could take a look at the issue on stack overflow (tensorflow - Why can my conditional Masked Autoregressive Flow not learn multimodal posteriors? - Stack Overflow) I'd really really appreciate it!

Thank you so much in advance!