search

kmzzhang / nbi

Package for Neural Posterior Estimation and Importance Sampling focused on Astronomical Applications
Other
33 stars 4 forks source link

SNPE importance resampling #4

Closed kmzzhang closed 1 year ago

kmzzhang commented 1 year ago

For rounds > 1, training parameters needs to be resampled to the prior to avoid bias.

kmzzhang commented 1 year ago

resampling proposal distribution for round > 1 to the prior prevents exact importance sampling to the posterior later on. need to rethink the algorithm

kmzzhang commented 1 year ago

doesn't seem to help with training anyways. deferred