Add second forward output and autodiff backend extensions #40
Conversation
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## main #40 +/- ##
==========================================
+ Coverage 94.82% 96.38% +1.55%
==========================================
Files 2 5 +3
Lines 58 83 +25
==========================================
+ Hits 55 80 +25
Misses 3 3
☔ View full report in Codecov by Sentry. |
| We compute the Jacobian-vector product `Jv` by solving `Au = -Bv` and setting `Jv = u`. | ||
| Keyword arguments are given to both `implicit.forward` and `implicit.conditions`. | ||
| """ | ||
| function ChainRulesCore.frule( |
There was a problem hiding this comment.
Mostly because
- it's only useful for Diffractor at the moment
- I cannot rewrite it using AbstractDifferentiation because of add
ForwardRuleConfigBackend(to matchReverseRuleConfigBackend) JuliaDiff/AbstractDifferentiation.jl#80
| - version: 'nightly' | ||
| os: ubuntu-latest | ||
| arch: x64 | ||
| allow_failure: true |
There was a problem hiding this comment.
Added so that failures on nightly (currently due to Zygote precompilation) do not impact the "CI passing" badge
| ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4" | ||
| Krylov = "ba0b0d4f-ebba-5204-a429-3ac8c609bfb7" | ||
| LinearOperators = "5c8ed15e-5a4c-59e4-a42b-c7e8811fb125" | ||
| Requires = "ae029012-a4dd-5104-9daa-d747884805df" |
There was a problem hiding this comment.
For ForwardDiff.jl compatibility in Julia versions < 1.9
|
|
||
| The same trick works for multiple outputs. | ||
|
|
||
| ## Constrained optimization modeling |
There was a problem hiding this comment.
I made the tutorial more simple on this aspect, and added details in the FAQ
| @test ForwardDiff.jacobian(first ∘ implicit, X) ≈ JJ #src | ||
| @test Zygote.jacobian(first ∘ implicit, X)[1] ≈ JJ #src | ||
|
|
||
| # Skipped because of https://github.com/JuliaDiff/ChainRulesTestUtils.jl/issues/232 #src |
There was a problem hiding this comment.
Any idea what to do here?
|
|
||
| #= | ||
| Unsurprisingly, the Jacobian is the identity. | ||
| In this instance, we could use ForwardDiff.jl directly on the solver, but it returns the wrong result (not sure why). |
There was a problem hiding this comment.
Still no clue why ForwardDiff fails here
| When we call it as a function, it just falls back on `implicit.forward`, so unsurprisingly we get the same tuple $(y(x), z(x))$. | ||
| =# | ||
|
|
||
| (first ∘ implicit)(x) ≈ sqrt.(x) |
There was a problem hiding this comment.
Always using first is cumbersome but it is the price to pay to return y, z
| Dual{T}(y[i], Partials(Tuple(dy[k][i] for k in 1:N))) | ||
| end | ||
|
|
||
| z_and_dz = Dual{T}(z, Partials(Tuple(zero(z) for k in 1:N))) |
There was a problem hiding this comment.
This requires z to support zero, which is less than ideal, but I don't know how to indicate to ForwardDiff that a derivative is zero. We need something like ZeroTangent() that is dimension-agnostic
| end | ||
|
|
||
| """ | ||
| PushforwardMul!{P,N} |
There was a problem hiding this comment.
Replaced lazy_jacobian with pushforward/pullback_function due to JuliaDiff/AbstractDifferentiation.jl#83
Main changes planned:
f(x) = y, zwhereyis the actual output andzis any additional information used in the optimality conditions.Link to preview of new docs: https://gdalle.github.io/ImplicitDifferentiation.jl/previews/PR40/