NUTS adapatation samples appear in full chain

[cpfiffer]

It looks like the samples from the NUTS adaptation phase are being included in the full chain. Is this expected behavior? If not, we need some kind of infrastructure in place to separate out the adaptation samples from the chains so that parameter estimates are more appropriate.

Let me know if this needs to be moved over to MCMCChains.

[goedman]

The links are no longer correct.

[trappmartin]

I think it's not expected behaviour. I'm currently looking into the NUTS code. I'll create a PR on this and on some minor performance improvements.

[yebai]

@kai @cpfiffer Has this been fixed by recent changes in AHMC and MCMCChains?

[cpfiffer]

No, MCMCChains still does not have any internals to catch adaptation samples. I'm wondering if we should fix this on the Turing side by defaulting to only return the post-warmup samples, and discarding the adaptation samples unless the user specifically requests them (using a keyword like NUTS(. . . , keep_warmup=true) or something). Does that seem reasonable? Is there any reason to keep the warmup samples around if most people just get rid of them?

[robsmith11]

As a beginner attempting to reproduce the resuts of a simple linear model from Stan, I was surprised to the see the warmup samples included in the summary stats. I expected them to have been removed.

[yebai]

Excluding warmup samples from chains sounds reasonable. In the longer term, we might need an argument for the sample interface (#746) to decide to retain/discard warmup samples.

[cpfiffer]

I've opened up a PR Turing-side to fix this: #784