pm.sample() can optionally return a trace as an arviz InferenceData object instead of a MultiTrace, and soon that will be in the default behavior. InferenceData is better than MultiTrace in several ways. But it seems to be missing a key feature. When sampling in NUTS encounters divergent transition…

InferenceData is missing details about divergent transitions?

bacalfa March 12, 2022, 6:06pm 5

Thanks! Added a comment there.

2 Likes