unofficialjuliamirror / difflinearalgebra.jl-b6128fcc-59fa-5305-9ee8-8a745badfb2b

Last mirrored from https://github.com/invenia/DiffLinearAlgebra.jl.git on 2019-10-07T21:39:12.128-04:00 by @UnofficialJuliaMirrorBot via Travis job 475.9 , triggered by Travis cron job on branch "master"

DiffLinearAlgebra

Note: the current version of this package is not intended for general consumption.

DiffLinearAlgebra can be (very loosely) thought of as DiffRules.jl for linear algebra. For every sensitivity, we provide a function which, when provided with the input and output from the forward pass and the reverse-mode sensitvity w.r.t the output from the forward pass, computes the sensitivity of the specified argument.

  A, B = randn(5, 3), randn(3, 4)
  C, C̄ = A * B, randn(5, 4)
  Ā = ∇(*, Val{1}, (), C, C̄, A, B)
  B̄ = ∇(*, Val{2}, (), C, C̄, A, B)

In the above example, the sensitivities of A and B are computed from C and a random seeding of C̄. (Note that the third argument is currently redundant; see this issue for motivation for its inclusion.)

We also expose some "metadata" for each implemented sensitivity. This is done via a set called ops contains DiffOp structs. These structs contain information regarding the arguments types supported by each sensitivity, and which arguments are differentiable.