Update to ChainRulesCore 0.10 #182
Conversation
| filldist_broken = if d.f(d.θ...) isa Skellam | ||
| (d.broken..., :ReverseDiff) | ||
| ((d.broken..., :ReverseDiff), (d.broken..., :Zygote, :ReverseDiff)) |
There was a problem hiding this comment.
Is the redundant :ReverseDiff in (d.broken..., :ReverseDiff) and (d.broken..., :Zygote, :ReverseDiff) intended?
There was a problem hiding this comment.
Yes, ReverseDiff still doesn't work for both tested sizes (single sample + two samples) whereas Zygote only fails if there are multiple samples. Upstream changes in SpecialFunctions and Zygote fixed most of the issues in #156 (and TuringLang/DistributionsAD.jl#145) but this one is still present.
|
Test errors are caused by NNlib it seems - I don't understand why we end up with such an old version (0.7.9), IIRC the problem is fixed in recent versions of NNlib. |
|
Ah OK, so 0.7.9 was the latest release without explicit ChainRules dependency: ChainRulesCore 0.10 is pulled in first and since no more recent version of NNlib is compatible with it (well, 0.7.21 is but it requires Julia >= 1.6) the old version with the bug is installed. I'll try to force NNlib 0.7.18 explicitly (latest release which supports ChainRulesCore 0.9 + Julia 1.3). |
|
Many thanks, @devmotion. |
Supercedes #180 and #181.