Skip to content

[AutoDiff] [stdlib] Add operators for custom gradients. #23036

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Mar 2, 2019
Merged

[AutoDiff] [stdlib] Add operators for custom gradients. #23036

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 34 additions & 1 deletion stdlib/public/core/AutoDiff.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -170,6 +170,39 @@ public func differentiableFunction<T, U, R>(
return original
}

public extension Differentiable {
@differentiable(wrt: self, vjp: _vjpWithGrad)
func withGradient(_ body: @escaping (inout CotangentVector) -> Void) -> Self {
Copy link

@marcrasi marcrasi Mar 2, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

withGradient(_ body: @escaping (CotangentVector) -> CotangentVector) seems like a more natural API at first glance. Do you have a reason to do it this way?

Copy link

@marcrasi marcrasi Mar 2, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Discussed in person

Copy link
Contributor Author

@rxwei rxwei Mar 2, 2019 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think my core rationale is that withXXX APIs in Swift usually return whatever the closure returns, e.g. withUnsafePointer(_:) and withoutAcutallyEscaping(_:), so x.withGradient { $0 + 1 } would syntactically indicate to a Swift programmer that this is "returning the gradient plus one in the forward pass".

Forbidding a return value makes it slightly less misleading. We can totally change this later.

return self
}

@inlinable
internal func _vjpWithGrad(
_ body: @escaping (inout CotangentVector) -> Void
) -> (Self, (CotangentVector) -> CotangentVector) {
return (self, { grad in
var grad = grad
body(&grad)
return grad
})
}

@differentiable(wrt: self, vjp: _vjpWithGrad)
func withGradient(_ body: @escaping (CotangentVector) -> Void) -> Self {
return self
}

@inlinable
internal func _vjpWithGrad(
_ body: @escaping (CotangentVector) -> Void
) -> (Self, (CotangentVector) -> CotangentVector) {
return (self, { grad in
body(grad)
return grad
})
}
}

/// Make a function be recomputed in its pullback, known as "checkpointing" in
/// traditional automatic differentiation.
@inlinable
Expand All @@ -190,7 +223,7 @@ public extension Differentiable {
return body(self)
}

@usableFromInline
@inlinable
internal func _vjp_withRecomputationInPullbacks<Result : Differentiable>(
_ body: @escaping @differentiable (Self) -> Result
) -> (Result, (Result.CotangentVector) -> CotangentVector) {
Expand Down
17 changes: 17 additions & 0 deletions test/AutoDiff/custom_derivatives.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,4 +84,21 @@ CustomDerivativesTests.test("Retroactive") {
expectEqual(100, gradient(at: 3, in: functionWithRetroDeriv))
}

CustomDerivativesTests.test("SumOfGradPieces") {
var grad: Float = 0
let addToGrad = { grad += $0 }
_ = gradient(at: 4) { (x: Float) in
x.withGradient(addToGrad)
* x.withGradient(addToGrad)
* x.withGradient(addToGrad)
}
expectEqual(48, grad)
}

CustomDerivativesTests.test("ModifyGradientOfSum") {
expectEqual(30, gradient(at: 4) { (x: Float) in
x.withGradient { $0 *= 10 } + x.withGradient { $0 *= 20 }
})
}

runAllTests()