DRAFT: Unconditional density estimation

dingo-gw / dingo

Dingo: Deep inference for gravitational-wave observations

MIT License

60 stars 20 forks source link

DRAFT: Unconditional density estimation #81

Closed max-dax closed 2 years ago

max-dax commented 2 years ago

Add code to recover posterior density by training an unconditional flow on dingo samples. This is necessary, since we lose access to the density with GNPE.

Beyond new scripts for this task, I needed to update dingo.core.models.PosteriorModel and some functions in dingo.core.nn.nsf to allow for unconditional sampling.

Still needs clean up!

max-dax commented 2 years ago

Todos for the future:

Improve output format of dingo samples
Use nflows sample_and_log_prob, which is faster than doing it separately
Check that model and model_init in inference use same reference time