search

FluxML / Tracker.jl

Flux's ex AD
Other
51 stars 36 forks source link

Fix tests on 1.8 #128

Closed mcabbott closed 2 years ago

mcabbott commented 2 years ago

For #125, this replaces the method definitions for cat functions with slightly less crazy loops, which have fewer ambiguities. In any cat(A,B,Cs...), one of the first two must be a tracked type, no longer one of the first 3, as this shortens the list.

The other failure was depthwiseconv. This is depends on the random number seed, it doesn't consistently fail, I have not investigated whether this is a numerical problem or a gradient problem.

coveralls commented 2 years ago

Pull Request Test Coverage Report for Build 2898925212

Changes Missing Coverage Covered Lines Changed/Added Lines %
src/lib/array.jl 13 16 81.25%
<!-- Total: 14 17 82.35% -->
Files with Coverage Reduction New Missed Lines %
src/lib/array.jl 9 66.67%
<!-- Total: 9 -->
Totals Coverage Status
Change from base Build 1896746788: 0.7%
Covered Lines: 469
Relevant Lines: 657
💛 - Coveralls
codecov-commenter commented 2 years ago

Codecov Report

Merging #128 (f885295) into master (84ff74d) will increase coverage by 0.70%. The diff coverage is 82.35%.

@@            Coverage Diff             @@
##           master     #128      +/-   ##
==========================================
+ Coverage   70.67%   71.38%   +0.70%     
==========================================
  Files           9        9              
  Lines         648      657       +9     
==========================================
+ Hits          458      469      +11     
+ Misses        190      188       -2
Impacted Files Coverage Δ
src/lib/array.jl 66.66% <81.25%> (+1.05%) :arrow_up:
src/numeric.jl 100.00% <100.00%> (ø)
src/back.jl 90.38% <0.00%> (+0.09%) :arrow_up:
src/forward.jl 85.71% <0.00%> (+0.86%) :arrow_up:

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

ToucheSir commented 2 years ago

If you're running tests with >1 threads, that's a long-standing NNlib issue and I'm not sure where it comes from :(