chain rules for DCT #273
chain rules for DCT #273
Conversation
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## master #273 +/- ##
==========================================
- Coverage 73.08% 70.70% -2.38%
==========================================
Files 5 6 +1
Lines 535 553 +18
==========================================
Hits 391 391
- Misses 144 162 +18
☔ View full report in Codecov by Sentry. |
|
This error is only happening with the MKL provider. With MKL, FFTW.jl doesn't even compile on my machine. Could be due to 008bc5b? test_frule: idct on Array{Float64, 3},Int64: Error During Test at /home/runner/.julia/packages/ChainRulesTestUtils/C9L2i/src/testers.jl:123
Got exception outside of a @test
FFTW could not create plan
Stacktrace:
[1] error(s::String)
@ Base ./error.jl:35
[2] macro expansion
@ FFTW ~/work/FFTW.jl/FFTW.jl/src/fft.jl:722 [inlined] |
|
@devmotion, could you please review this? LMK if you want me to remove the |
|
Only the MKL tests are failing. Is there a regression with MKL? |
More concretely, #249 outlines the intended approach. FFTW was not supposed to define custom rules. |
|
This PR defines Chain Rules for |
|
With #249 , gradient computation would error for DCT/IDCT. This is because julia> using FFTW, Zygote
julia>
julia> using LinearAlgebra, FFTW, Zygote
julia> x = rand(4)
4-element Vector{Float64}:
0.8692266334693106
0.6938635624794242
0.552208368655668
0.9197557963740512
julia> f(x) = x |> dct |> idct |> norm
f (generic function with 1 method)
julia> f(x)
1.5452787421840921
julia> Zygote.gradient(f, x)
ERROR: Compiling Tuple{Type{FFTW.r2rFFTWPlan{Float64, Any, false, 1}}, Vector{Float64}, FFTW.FakeArray{Float64, 1}, UnitRange{Int64}, Int64, UInt32, Float64}: try/catch is not supported.
Refer to the Zygote documentation for fixes.
https://fluxml.ai/Zygote.jl/latest/limitations |
ext/FFTWChainRulesCoreExt.jl
Outdated
|
|
||
| # R2R | ||
|
|
||
| function ChainRulesCore.frule(Δ, ::typeof(r2r), x::AbstractArray, region...) |
There was a problem hiding this comment.
It seems the rrule for r2r is missing?
There was a problem hiding this comment.
The R2R transforms are not unitary. There is some scaling involved that depends on the kind of R2R transform. Because it looks like an involved task, I chose to skip that for now. I am happy to look into that in a separate PR
Co-authored-by: David Widmann <[email protected]>
Co-authored-by: David Widmann <[email protected]>
|
@devmotion I've addressed all your comments. LMK if you have more questions :D |
|
@devmotion ping :) |
address #272