Haha, yes! We are working on adding support for numpyro and pymc next.

Kudos to @sethaxen.com for implementing the Pyro wrapper that makes this possible (shipped in sbi v0.25)!

And thanks to @juanitorduz.bsky.social sharing the cookie factory example—it's a great accessible example for hierarchical inference.

Everything runs in Colab 📊

thanks for the response! yes, that makes sense. I was wondering how exactly you sample from the augmented posterior as this was not clear to me from reading the paper. E.g., in the SBI case (DDM example), how do you sample p(theta | x_i, x_o) given that the conditions are not iid?

Really neat idea!
I had a quick look and did not quite understand how the „augmented“ posterior is obtained or how you can sample from it to calculate the conditional ranks.
Thanks 🙏

Actually, I think this can be useful for mixed-effects models as well: you define a hierarchical prior and simulator with your fixed effects model and then train NLE on single trial data. At inference time, you then condition on x and conditions for each subject, or for multiple subjects at once.

Good question, thanks! When you train NLE on single trials, you can now run inference given a set of iid trials and additionally condition on corresp. set of varying experimental condition (without retraining). Perm. inv. nets are used in NPE settings, which usually requires retraining. Clarified?

Would you mind adding me as well? I work on SBI and related methods. Thanks!

Great list! Thanks for initiating this. Can you please add me as well? I work on OSS for science, e.g., bsky.app/profile/sbi-...
Thanks!

Great, thanks for creating this! Please add me as well.