Generalize flow based models #43
Description
So far (S)NLE and NPE ony support discrete NFs, while a separate class FMPE is needed for continuous NFs. Also there is currently no way to use CNFs with NLE.
It would be better in the long run to have a flexible NLE/NPE class that can take any density estimator. This would violate the current naming conventions in the literature, but this generalized view is a bit more reasonable imho.
We could then flexibly use
- Continuous flows using the rectified flow and flow matching losses
- Discrete flows such as the (block) neural autoregressive flows
- Consistency (flow) models and derivations