Erratic errors in typed Hessians about forwarddiff.jl HOT 4 CLOSED

Good catch!

The FADHessian code in master has a bunch of little bugs in both utility functions (e.g. constructors/conversion) and a couple in the mathematical definitions as well. A large number of tiny bugfixes are included in #27, and the state of testing, functional coverage, and performance is greatly improved. A lot of this has to do with the move to automating the definitions of univariate mathematical functions - I'm not even sure how many bugs got fixed by 548be29, since past test coverage was intermittent, but I suspect quite a few. It's likely that by the time #27 lands (in a week or two), this could issue will be fixed (sorry for the wait in the meantime).

I don't have time tonight to check out for sure that it does fix the above, but I'll check tomorrow and report back.

from forwarddiff.jl.

That sounds fine, thanks for the quick response. I tried a handful of other trig functions and did not observe the problem, and I can implement atan with exp by hand in the meantime, so no rush on my account.

Thanks!

from forwarddiff.jl.

I ran your examples, and I can say with more certainty that this is fixed by #27 - I'm seeing a bad_versions/max_iter ratio of 0.0, even when cranking up max_iter to 10e6.

The code I used is here; it's basically the code example you gave, but modified to use the new API exposed in #27.

Unless you find anything else, I think this issue can be closed once #27 lands.

from forwarddiff.jl.

This now fixed on master as of the merging of #27.

from forwarddiff.jl.