genericize CGLS #53
genericize CGLS #53
Conversation
| } | ||
| } | ||
|
|
||
| /// Errors + scalar | ||
| static func + (_ lhs: Self, _ rhs: Double) -> Self { |
There was a problem hiding this comment.
Could you explain briefly how these are automatically implemented? Curious on the workings...
There was a problem hiding this comment.
Yeah! We have automatic synthesis for AdditiveArithmetic and VectorProtocol that looks at all the fields of the struct. If all the fields conform to AdditiveArithmetic or VectorProtocol, it automatically implements the requirements by applying the functions to all the members.
There was a problem hiding this comment.
The single field Array<Error>.DifferentiableView conforms to AdditveArithmetic and VectorProtocol because:
Error(akaVector) conforms to them.Array.DifferentiableViewconforms toAdditiveArithmeticwhen itsElementdoes: https://github.com/apple/swift/blob/4bc72aedb2b32c33ed8e2ec241615fc890b60002/stdlib/public/Differentiation/ArrayDifferentiation.swift#L98Array.DifferentiableViewconforms toVectorProtocolwhen itsElementdoes: https://github.com/tensorflow/swift-apis/blob/a8a24c46e478ce50c1c9a7718a41eec453e5b670/Sources/TensorFlow/StdlibExtensions.swift#L224
|
See my comment in #42 - might impact above |
|
Based on discussions in #42 and with Dave, I have improved the protocol naming. |
|
We really need to talk :-) I sent an invite for a hangout at 11am your time. |
|
|
||
| // MARK: - Additional EuclideanVectorSpace requirements. | ||
|
|
||
| public var squaredNorm: Double { |
There was a problem hiding this comment.
Do we still need the .differentiableMap methods?
There was a problem hiding this comment.
.differentiableMap is still necessary for writing a differentiable function that does a map over an array.
This function currently isn't declared @differentiable, so it does not need to use .differentiableMap.
I like to avoid adding the @differentiable annotation until we actually need to use a function in a differentiable way, to keep things as simple as possible.
|
Closing this to clean up, since it won't be relevant with the new generic factor graph data structure. |
This PR introduces a
DecomposedAffineFunctionprotocol abstracting the concept of an affine function whose linear and bias components are accessible. Also, its linear component can be applied in the forwards or adjoint direction.This PR also changes CGLS to work in terms of an abstract
DecomposedAffineFunction. Finally, this PR conformsGaussianFactorGraphtoDecomposedAffineFunctionso that we can run CGLS on it.This allows us to run CGLS on anything, as long as it supports the required operations. For example:
GaussianFactorGraphas it is today.GaussianFactorGraph!DistributedGaussianFactorGraphthat distributes its state over multiple computers, and run the existing CGLS algorithm without any modifications. (I haven't thought too hard about this, so there may be some practical obstacles.)I wrote more justifications for why genericizing is good at #42 (comment) and #42 (comment).