Closed willtebbutt closed 2 years ago
Merging #466 (aef58b9) into master (b5af459) will decrease coverage by
0.07%
. The diff coverage is66.66%
.
@@ Coverage Diff @@
## master #466 +/- ##
==========================================
- Coverage 93.16% 93.09% -0.08%
==========================================
Files 52 52
Lines 1259 1275 +16
==========================================
+ Hits 1173 1187 +14
- Misses 86 88 +2
Impacted Files | Coverage Δ | |
---|---|---|
src/transform/chaintransform.jl | 80.00% <66.66%> (+1.73%) |
:arrow_up: |
src/matrix/kernelpdmat.jl | 75.00% <0.00%> (-6.82%) |
:arrow_down: |
src/kernels/normalizedkernel.jl | 80.00% <0.00%> (-2.36%) |
:arrow_down: |
src/mokernels/lmm.jl | 100.00% <0.00%> (ø) |
|
src/kernels/kernelsum.jl | 100.00% <0.00%> (ø) |
|
src/kernels/kernelproduct.jl | 100.00% <0.00%> (ø) |
|
src/kernels/kerneltensorproduct.jl | 98.85% <0.00%> (+0.08%) |
:arrow_up: |
src/approximations/nystrom.jl | 92.68% <0.00%> (+0.18%) |
:arrow_up: |
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.
@theogf let me know whether my explanation of the testing is sufficient, and I'll add a docstring + merge
Will squash + merge when CI passes
Summary
The
ChainTransform
has some performance issues on master.Evidence:
master:
This branch:
Proposed changes
_map
rather thanmap
, because that's the APINote that the way I'm testing that this change has been successful is by checking that the number of allocations required to compute the
kernelmatrix
, its forwards-pass and pullback (using Zygote) is invariant to the size of input vector considered. I plan to roll this out more widely in the coming days.What alternatives have you considered?
None
Breaking changes
This only widens the set of permissible types in the
ChainTransform
, and which one gets used by default. On the basis of this, my inclination is to suggest that we shouldn't consider this breaking, but I might have missed something obvious.