Should NaN's display as errors?

[penelopeysm]

Some of the FiniteDifferences 'wrong' entries are not really numerically wrong, but rather NaN's.

It appears to me that it would be fairer to mark NaN as 'error' rather than 'wrong', since attempting to use this in sampling will fail (rather than proceeding silently with the wrong result, which is the scenario we really want to be careful about).

[yebai]

I suggest we distinguish between two types of wrongs

• Wrong (incorrect answer)
• Wrong (NaN)

[penelopeysm]

Distinguishing between outright incorrect answers and NaN's will be possible once DynamicPPL 0.36 (and specifically this PR TuringLang/DynamicPPL.jl#890) is released!

Not closing yet as there is still some work to be done in this repo, although it will just be a fairly trivial any(isnan) check.

(Incidentally, that release will also make the AD run with linked varinfo by default, so it should actually get rid of the NaN's)