Basics

We present the main features of DifferentiationInterface.jl.

using DifferentiationInterface

Computing a gradient

A common use case of automatic differentiation (AD) is optimizing real-valued functions with first- or second-order methods. Let's define a simple objective (the squared norm) and a random input vector

function f(x::AbstractVector{T}) where {T}
    y = zero(T)
    for i in eachindex(x)
        y += abs2(x[i])
    end
    return y
end

x = collect(1.0:5.0)
5-element Vector{Float64}:
 1.0
 2.0
 3.0
 4.0
 5.0

To compute its gradient, we need to choose a "backend", i.e. an AD package to call under the hood. Most backend types are defined by ADTypes.jl and re-exported by DifferentiationInterface.jl.

ForwardDiff.jl is very generic and efficient for low-dimensional inputs, so it's a good starting point:

import ForwardDiff

backend = AutoForwardDiff()
AutoForwardDiff{nothing, Nothing}(nothing)
Tip

To avoid name conflicts, load AD packages with import instead of using. Indeed, most AD packages also export operators like gradient and jacobian, but you only want to use the ones from DifferentiationInterface.jl.

Now you can use the following syntax to compute the gradient:

gradient(f, backend, x)
5-element Vector{Float64}:
  2.0
  4.0
  6.0
  8.0
 10.0

Was that fast? BenchmarkTools.jl helps you answer that question.

using BenchmarkTools

@benchmark gradient($f, $backend, $x)
BenchmarkTools.Trial: 10000 samples with 154 evaluations.
 Range (minmax):  678.071 ns151.684 μs   GC (min … max): 0.00% … 99.21%
 Time  (median):     938.062 ns                GC (median):    0.00%
 Time  (mean ± σ):   968.250 ns ±   3.202 μs   GC (mean ± σ):  8.62% ±  2.62%

  █▇▄▃▃▂▂▁▁  ▁                ▃▅▆▇▇▆▆▆▅▄▃▂▂▂▂▂▂▁▁▁▁           ▃
  ███████████████▇▇████▇▇▇▆▆▆▇█████████████████████████▇▇▇▇▆▆ █
  678 ns        Histogram: log(frequency) by time       1.17 μs <

 Memory estimate: 848 bytes, allocs estimate: 4.

Not bad, but you can do better.

Overwriting a gradient

Since you know how much space your gradient will occupy (the same as your input x), you can pre-allocate that memory and offer it to AD. Some backends get a speed boost from this trick.

grad = similar(x)
gradient!(f, grad, backend, x)
grad  # has been mutated
5-element Vector{Float64}:
  2.0
  4.0
  6.0
  8.0
 10.0

The bang indicates that one of the arguments of gradient! might be mutated. More precisely, our convention is that every positional argument between the function and the backend is mutated (and the extras too, see below).

@benchmark gradient!($f, _grad, $backend, $x) evals=1 setup=(_grad=similar($x))
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (minmax):  680.000 ns 19.655 μs   GC (min … max): 0.00% … 0.00%
 Time  (median):     709.000 ns                GC (median):    0.00%
 Time  (mean ± σ):   737.869 ns ± 321.410 ns   GC (mean ± σ):  0.00% ± 0.00%

   ▁▆██▅▃         ▁▁▁▁▁▁▁                                      ▂
  ▅████████▆▇▇▆▇████████████▇▇█▇▇███▇▇▇▇▆▆▆▆▅▆▆▄▆▄▆▅▄▅▅▄▄▅▅▄▅ █
  680 ns        Histogram: log(frequency) by time       1.07 μs <

 Memory estimate: 752 bytes, allocs estimate: 3.

For some reason the in-place version is not much better than your first attempt. However, it makes fewer allocations, thanks to the gradient vector you provided. Don't worry, you can get even more performance.

Preparing for multiple gradients

Internally, ForwardDiff.jl creates some data structures to keep track of things. These objects can be reused between gradient computations, even on different input values. We abstract away the preparation step behind a backend-agnostic syntax:

extras = prepare_gradient(f, backend, randn(eltype(x), size(x)))
DifferentiationInterfaceForwardDiffExt.ForwardDiffGradientExtras{ForwardDiff.GradientConfig{ForwardDiff.Tag{typeof(Main.f), Float64}, Float64, 5, Vector{ForwardDiff.Dual{ForwardDiff.Tag{typeof(Main.f), Float64}, Float64, 5}}}}(ForwardDiff.GradientConfig{ForwardDiff.Tag{typeof(Main.f), Float64}, Float64, 5, Vector{ForwardDiff.Dual{ForwardDiff.Tag{typeof(Main.f), Float64}, Float64, 5}}}((Partials(1.0, 0.0, 0.0, 0.0, 0.0), Partials(0.0, 1.0, 0.0, 0.0, 0.0), Partials(0.0, 0.0, 1.0, 0.0, 0.0), Partials(0.0, 0.0, 0.0, 1.0, 0.0), Partials(0.0, 0.0, 0.0, 0.0, 1.0)), ForwardDiff.Dual{ForwardDiff.Tag{typeof(Main.f), Float64}, Float64, 5}[Dual{ForwardDiff.Tag{typeof(Main.f), Float64}}(0.0,6.90405776460003e-310,6.9039502772931e-310,0.0,0.0,0.0), Dual{ForwardDiff.Tag{typeof(Main.f), Float64}}(0.0,0.0,0.0,0.0,0.0,0.0), Dual{ForwardDiff.Tag{typeof(Main.f), Float64}}(0.0,0.0,0.0,0.0,0.0,0.0), Dual{ForwardDiff.Tag{typeof(Main.f), Float64}}(0.0,0.0,0.0,0.0,0.0,0.0), Dual{ForwardDiff.Tag{typeof(Main.f), Float64}}(0.0,0.0,0.0,0.0,0.0,0.0)]))

You don't need to know what this object is, you just need to pass it to the gradient operator. Note that preparation does not depend on the actual components of the vector x, just on its type and size. You can thus reuse the extras for different values of the input.

grad = similar(x)
gradient!(f, grad, backend, x, extras)
grad  # has been mutated
5-element Vector{Float64}:
  2.0
  4.0
  6.0
  8.0
 10.0

Preparation makes the gradient computation much faster, and (in this case) allocation-free.

@benchmark gradient!($f, _grad, $backend, $x, _extras) evals=1 setup=(
    _grad=similar($x);
    _extras=prepare_gradient($f, $backend, $x)
)
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (minmax):  181.000 ns 21.462 μs   GC (min … max): 0.00% … 0.00%
 Time  (median):     191.000 ns                GC (median):    0.00%
 Time  (mean ± σ):   199.762 ns ± 259.211 ns   GC (mean ± σ):  0.00% ± 0.00%

     ▃█                                                         
  ▂▂▄██▆▆▅▄▃▃▂▂▂▂▂▂▂▂▁▂▂▁▂▁▂▂▂▁▁▂▂▁▂▂▂▁▂▂▁▂▂▂▂▂▁▂▂▂▂▂▂▂▂▂▂▂▂▂ ▃
  181 ns           Histogram: frequency by time          294 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

Beware that the extras object is nearly always mutated by differentiation operators, even though it is given as the last positional argument.

Switching backends

The whole point of DifferentiationInterface.jl is that you can easily experiment with different AD solutions. Typically, for gradients, reverse mode AD might be a better fit, so let's try the state-of-the-art Enzyme.jl!

import Enzyme

backend2 = AutoEnzyme()
AutoEnzyme{Nothing}(nothing)

Once the backend is created, things run smoothly with exactly the same syntax as before:

gradient(f, backend2, x)
5-element Vector{Float64}:
  2.0
  4.0
  6.0
  8.0
 10.0

And you can run the same benchmarks to see what you gained (although such a small input may not be realistic):

@benchmark gradient!($f, _grad, $backend2, $x, _extras) evals=1 setup=(
    _grad=similar($x);
    _extras=prepare_gradient($f, $backend2, $x)
)
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
 Range (minmax):  112.000 ns 10.467 μs   GC (min … max): 0.00% … 0.00%
 Time  (median):     119.000 ns                GC (median):    0.00%
 Time  (mean ± σ):   120.776 ns ± 103.809 ns   GC (mean ± σ):  0.00% ± 0.00%

                      ▄   ▇                             
  ▂▁▁▁▂▁▁▁▃▁▁▁▄▁▁▁█▁▁▁█▁▁▁█▁▁▁▁▁▁█▁▁▁▁▁▁█▁▁▁▆▁▁▁▄▁▁▁▃▁▁▁▃▁▁▁▂ ▃
  112 ns           Histogram: frequency by time          127 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

In short, DifferentiationInterface.jl allows for easy testing and comparison of AD backends. If you want to go further, check out the documentation of DifferentiationInterfaceTest.jl. This related package provides benchmarking utilities to compare backends and help you select the one that is best suited for your problem.