Jan

icon indicating a website link janfb.github.io

icon company @aai-institute icon location Munich icon location Researcher @aai-institute. Working on Bayesian Machine Learning and Simulation Based Inference

Results 241 comments of Jan

Implement SNPE-B

More context: At the moment, we only implement multi-round NPE via - SNPE-A with post-hoc correction, - and SNPE-C based on learning the proposal posterior. SNPE-B enables multi-round NPE using...

Implement SNPE-B

Thanks a lot for looking into this and adding the detailed report 👍 I am tagging @jan-matthis because he probably encountered similar implementation issues in [`delfi`](https://github.com/mackelab/delfi/blob/main/delfi/inference/SNPEB.py) back then. See also...

Runtime of some tests could be reduced by reducing num simulations

fixed through #923 and #924

sbi.diagnostics?

Most the of the methods are implemented by now: SBC, ArviZ, coverage.

Local sbc, issue #621

fix for #621

Local sbc, issue #621

closed in favor of #1109

allow sampling of posterior outside of prior bounds

fix: sample from the density estimator directly and set your own rejection bounds.

systematic tests for sampling methods

related to #908: we can add centralized tests when introducing new MCMC samplers.

Deprecate the simple interface.

More context: The simple interface is here: https://github.com/sbi-dev/sbi/blob/c3c2b6e142fb4c57d5599effc1f01f8222a37c57/sbi/inference/base.py#L35 i.e., it's a single function that takes simulator, prior, method and budget and runs SBI to return a posterior object. We created...

Flow training in unconstrained space

More context: With #952 #957 getting fixed, it will be possible to add new density estimators, e.g., those from [zuko](https://github.com/probabilists/zuko). Goal: some priors are bounded, but it can be beneficial...