support functions that return scalars #51
Conversation
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## main #51 +/- ##
==========================================
+ Coverage 93.58% 93.67% +0.08%
==========================================
Files 5 5
Lines 78 79 +1
==========================================
+ Hits 73 74 +1
Misses 5 5
☔ View full report in Codecov by Sentry. |
|
Thanks for the contribution! Not gonna merge it right away because I still struggle to make it work for reverse mode. |
|
The empty getindex |
|
@mohamed82008 this PR is a bit of a mess and I could use some help. Basically, scalar outputs work in forward mode but it's much harder in reverse mode to juggle the arrays. |
|
let me take a look |
|
Let's take advantage of |
|
even better struct ScalarImplicitFunction{F}
f::F
end
function ScalarImplicitFunction(forward, conditions, linear_solver)
_forward = x -> begin
y, z = forward(only(x))
return @SVector([y]), z
end
_conditions = (x, y, z) -> @SVector [conditions(only(x), only(y), z)]
f = ImplicitFunction(_forward, _conditions, linear_solver)
return ScalarImplicitFunction(f)
end
function (f::ScalarImplicitFunction)(x::Real;_ kwargs...)
return only(f.f(@SVector([x]); kwargs...))
end |
|
this shouldn't require touching any line of code in ImplicitDifferentiation unless we don't support static arrays. |
|
Gonna close this and start a new PR for the issue at hand |
| pbmA = PullbackMul!(pbA, size(y)) | ||
| pbmB = PullbackMul!(pbB, size(y)) | ||
| Aᵀ_op = LinearOperator(R, m, m, false, false, pbmA) | ||
| Aᵀ_op = LinearOperator(R, m, m, false, false, pbmA) # TODO: can it accept a vector if y is scalar? |
There was a problem hiding this comment.
@baggepinnen the bug that made me give up was here, so if you find an elegant solution we'll be happy to implement it!
| @@ -1,3 +1,6 @@ | |||
| make_array(a::AbstractArray) = a | |||
| make_array(a::Number) = [a] | |||
There was a problem hiding this comment.
For now it is not static but it should be
| @@ -39,7 +43,9 @@ function (implicit::ImplicitFunction)( | |||
| reshape(dₖy_vec, size(y)) | |||
| end | |||
|
|
|||
| y_and_dy = let y = y, dy = dy | |||
| y_and_dy = if y isa Number | |||
There was a problem hiding this comment.
Is this type-stable or should it be two methods?
There was a problem hiding this comment.
If the type of y is inferred correctly the code will be type stable, if it isn't type stable, you will have a problem either case
example:
julia> foo(x) = x isa Int ? 2 : 2.0
foo (generic function with 1 method)
julia> @code_warntype foo(3)
MethodInstance for foo(::Int64)
from foo(x) @ Main REPL[15]:1
Arguments
#self#::Core.Const(foo)
x::Int64
Body::Int64
1 ─ %1 = (x isa Main.Int)::Core.Const(true)
└── goto #3 if not %1
2 ─ return 2
3 ─ Core.Const(:(return 2.0))the interesting part is (x isa Main.Int)::Core.Const(true)
|
I think @mohamed82008 is right and this will be a bit easier once #57 is merged, because we'll have custom types for |
No description provided.