search

kazewong / flowMC

Normalizing-flow enhanced sampling package for probabilistic inference in Jax
https://flowmc.readthedocs.io/en/main/
MIT License
200 stars 23 forks source link

Zero Local Acceptance Rate but Good Global Acceptance #186

Closed zeeshan5885 closed 1 month ago

zeeshan5885 commented 1 month ago

Hi @kazewong , we are using flowMC for population inference of BBH and facing a strange issue. We are not sure what went wrong. We tried various configurations for the sampler but local acceptance is always zero. However, the global acceptance rate and training chains are going in the right direction.

We have plots and details in this issue https://github.com/gwkokab/gwkokab/issues/263 . We would appreciate any insights regarding such an issue.

Thanks

kazewong commented 1 month ago

Can you check whether your gradient is returning a non-nan value? Local sampler like MALA relies on the gradient to make a proposal.

zeeshan5885 commented 1 month ago

Thanks @kazewong, our issue has been resolved, we were getting the nan gradient for a single parameter.