JuliaDiff / JuliaDiff/ChainRulesCore.jl

Explicit examples on how to derive pushfowards and pullbacks

Open
#338 0 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Julia
Stars
267
Forks
66
PR merge metrics
No merged PRs in 30d

Description

It was mentioned on Slack that the docs could really use some explicit examples on how to derive a pushforward or pullback, so here is an example I cooked up that I'll try to polish up and turn into a docs PR.


Consider an R^2 -> R^2 function (2 real inputs and 2 real outputs)

 f((x, y)::Vector) = [u(x, y)
                      v(x, y)]

This thing has a Jacobian matrix

J(f)((x, y)) = [∂_u_x(x, y) ∂_u_y(x, y)
                ∂_v_x(x, y) ∂_v_y(x, y)]

where ∂_u_x is meant to be the partial derivative of u with respect to x and so on.

The pushforward is then defined as

Pf(f)((x, y)) = ((Δx, Δy),) -> J(f)((x, y)) * [Δx, Δy]

where Δx, Δy should be interpreted as 'wiggles in the input space'. This simplifies to

Pf(f)((x, y)) = ((Δx, Δy),) -> = [∂_x_u(x, y) * Δx + ∂_y_u(x, y) * Δy
                                  ∂_x_v(x, y) * Δx + ∂_y_v(x, y) * Δy]

Similarly, the pullback is defined as

Pb(f)((x, y)) = ((Δu, Δv),) -> transpose(J(f)((x, y))) * [Δu, Δv]

where Δu, Δv are to be interpreted at 'wobbles in the output space'. This simplifies to

Pb(f)((x, y)) = ((Δu, Δv),) -> [∂_x_u(x, y) * Δu + ∂_x_v(x, y) * Δv
                                ∂_y_u(x, y) * Δu + ∂_y_v(x, y) * Δv]

Okay, so suppose for concreteness that we have u(x, y) = sin(x) * y^2 and v(x, y) = exp(2x) + y then

J(f)((x, y)) = [cos(x)*y^2   2sin(x)*y
                2exp(2x) + y exp(2x) + 1]

and the pushforward is

Pf(f)((x, y)) = ((Δx, Δy),) -> [cos(x)*y^2 * Δx   + 2sin(x)*y * Δy
                                (2exp(2x)+y) * Δx + (exp(2x)+1) * Δy]

whereas the pullback is

Pb(f)((x, y)) = ((Δu, Δv),) -> [cos(x)*y^2 * Δu + (2exp(2x)+y) * Δv
                                2sin(x)*y * Δu  + (exp(2x)+1) * Δv]

We are now ready to write an frule and rrule for our function f.

function frule((_, (Δx, Δy)::Vector{<:Real}), ::typeof(f), r::Vector{<:Real})
    (x, y) = r
    z = f(r)
    sinx, cosx = sincos(x)
    exp2x = exp(2x)
    ∂z = [cosx*y^2 * Δx   + 2sinx*y * Δy
          (2exp2x+y) * Δx + (exp2x+1) * Δy]
    return z, ∂z
end

function rrule(::typeof(f), r::Vector{<:Real})
    (x, y) = r
    sinx, cosx = sincos(x)
    exp2x = exp(2x)
    z = f(r)
    Pb((Δu, Δv)::Vector{<:Real}) = [cosx*y^2 * Δu + (2exp2x+y) * Δv
                                    2sinx*y * Δu  + (exp2x+1) * Δv]
    return z, Pb
end

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reviewing the worked Julia example in this issue, including its Jacobian, pushforward, pullback, frule, and rrule sections. Locate the repository's documentation entry point and adapt the example there. Done means the docs contain explicit, polished derivations and corresponding rule examples that readers can follow.

Written by the indexing model from the issue text.

Assessment

Tech stack
julia
Domain
documentation
Issue type
Documentation
Difficulty
3/5
Estimated time
1-2 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.