Zygote

Welcome! Zygote extends the Julia language to support differentiable programming. With Zygote you can write down any Julia code you feel like – including using existing Julia packages – then get gradients and optimise your program. Deep learning, ML and probabilistic programming are all different kinds of differentiable programming that you can do with Zygote.

At least, that's the idea. We're still in beta so expect some adventures.

Setup

Zygote can be installed from the package manager in Julia's REPL:

] add Zygote

Taking Gradients

Zygote is easy to understand since, at its core, it has a one-function API (pullback), along with a few simple conveniences. Before explaining pullback, we'll look at the higher-level function gradient.

gradient calculates derivatives. For example, the derivative of $3x^2 + 2x + 1$ is $6x + 2$, so when x = 5, dx = 32.

julia> using Zygote

julia> gradient(x -> 3x^2 + 2x + 1, 5)
(32.0,)

gradient returns a tuple, with a gradient for each argument to the function.

julia> gradient((a, b) -> a*b, 2, 3)
(3.0, 2.0)

This will work equally well if the arguments are arrays, structs, or any other Julia type, but the function should return a scalar (like a loss or objective $l$, if you're doing optimisation / ML).

julia> W = rand(2, 3); x = rand(3);

julia> gradient(W -> sum(W*x), W)[1]
2×3 Array{Float64,2}:
 0.0462002  0.817608  0.979036
 0.0462002  0.817608  0.979036

julia> gradient(x -> 3x^2 + 2x + 1, 1//4)
(7//2,)

Control flow is fully supported, including recursion.

julia> function pow(x, n)
         r = 1
         for i = 1:n
           r *= x
         end
         return r
       end
pow (generic function with 1 method)

julia> gradient(x -> pow(x, 3), 5)
(75.0,)

julia> pow2(x, n) = n <= 0 ? 1 : x*pow2(x, n-1)
pow2 (generic function with 1 method)

julia> gradient(x -> pow2(x, 3), 5)
(75.0,)

Data structures are also supported, including mutable ones like dictionaries. Arrays are currently immutable, though this may change in future.

julia> d = Dict()
Dict{Any, Any}()

julia> gradient(5) do x
         d[:x] = x
         d[:x] * d[:x]
       end
(10.0,)

julia> d[:x]
5

Structs and Types

Julia makes it easy to work with custom types, and Zygote makes it easy to differentiate them. For example, given a simple Point type:

import Base: +, -

struct Point
  x::Float64
  y::Float64
end

a::Point + b::Point = Point(a.x + b.x, a.y + b.y)
a::Point - b::Point = Point(a.x - b.x, a.y - b.y)
dist(p::Point) = sqrt(p.x^2 + p.y^2)
julia> a = Point(1, 2)
Point(1.0, 2.0)

julia> b = Point(3, 4)
Point(3.0, 4.0)

julia> dist(a + b)
7.211102550927978

julia> gradient(a -> dist(a + b), a)[1]
(x = 0.5547001962252291, y = 0.8320502943378437)

Zygote's default representation of the "point adjoint" is a named tuple with gradients for both fields, but this can of course be customised too.

This means we can do something very powerful: differentiating through Julia libraries, even if they weren't designed for this. For example, colordiff might be a smarter loss function on colours than simple mean-squared-error:

julia> using Colors

julia> colordiff(RGB(1, 0, 0), RGB(0, 1, 0))
86.60823557376344

julia> gradient(colordiff, RGB(1, 0, 0), RGB(0, 1, 0))
((r = 0.4590887719632896, g = -9.598786801605689, b = 14.181383399012862), (r = -1.7697549557037275, g = 28.88472330558805, b = -0.044793892637761346))

Explicit and Implicit Parameters

It's easy to work with even very large and complex models, and there are few ways to do this. Autograd-style models pass around a collection of weights. Depending on how you write your model, there are multiple ways to explicity take gradients with respect to parameters. For example, the function linear accepts the parameters as an argument to the model. So, we directly pass in the parameters, θ, as an argument to the function being differentiated.

Zygote.gradientMethod
gradient(f, args...)

Returns a tuple containing ∂f/∂x for each argument x, the derivative (for scalar x) or the gradient.

f(args...) must be a real number, see jacobian for array output.

See also withgradient to keep the value f(args...), and pullback for value and back-propagator.

julia> gradient(*, 2.0, 3.0, 5.0)
(15.0, 10.0, 6.0)

julia> gradient(x -> sum(abs2,x), [7.0, 11.0, 13.0])
([14.0, 22.0, 26.0],)

julia> gradient([7, 11], 0, 1) do x, y, d
         p = size(x, d)
         sum(x.^p .+ y)
       end
([14.0, 22.0], 2.0, nothing)
source
julia> linear(θ, x) = θ[:W] * x .+ θ[:b]
linear (generic function with 1 method)

julia> x = rand(5);

julia> θ = Dict(:W => rand(2, 5), :b => rand(2))
Dict{Any,Any} with 2 entries:
  :b => [0.0430585, 0.530201]
  :W => [0.923268 … 0.589691]

# Alternatively, use a named tuple or struct rather than a dict.
# θ = (W = rand(2, 5), b = rand(2))

julia> θ̄ = gradient(θ -> sum(linear(θ, x)), θ)[1]
Dict{Any,Any} with 2 entries:
  :b => [1.0, 1.0]
  :W => [0.628998 … 0.433006]

We can combine the role of the dictionary and the function here by making a callable struct which contains the parameters, equivalent to a closure. Passed explicitly to gradient, we get a named tuple with the same field names:

julia> struct Linear
         W
         b
       end

julia> (l::Linear)(x) = l.W * x .+ l.b

julia> model = Linear(rand(2, 5), rand(2))
Linear([0.267663 … 0.334385], [0.0386873, 0.0203294])

julia> dmodel = gradient(model -> sum(model(x)), model)[1]
(W = [0.652543 … 0.683588], b = [1.0, 1.0])

Zygote also supports another way to take gradients, via implicit parameters. Here the loss function takes zero arguments, but the variables of interest are indicated by a special Params object. The function linear which depends on W and b is executed when the loss function () -> sum(linear(x)) is called, and hence this dependence is visible to Zygote:

Zygote.gradientFunction
gradient(() -> loss(), ps::Params) -> Grads

Gradient with implicit parameters. Takes a zero-argument function, and returns a dictionary-like container, whose keys are arrays x in ps.

julia> x = [1 2 3; 4 5 6]; y = [7, 8]; z = [1, 10, 100];

julia> g = gradient(Params([x, y])) do
         sum(x .* y .* z')
       end
Grads(...)

julia> g[x]
2×3 Matrix{Float64}:
 7.0  70.0  700.0
 8.0  80.0  800.0

julia> haskey(g, z)  # only x and y are parameters
false
source
julia> W = rand(2, 5); b = rand(2);

julia> linear(x) = W * x .+ b
linear (generic function with 2 methods)

julia> grads = gradient(() -> sum(linear(x)), Params([W, b]))
Grads(...)

julia> grads[W], grads[b] # access gradients using arrays as keys
([0.652543 … 0.683588], [1.0, 1.0])

Here grads is a dictionary-like object, whose keys are the same parameters we indicated in Params. (In fact it wraps a dictionary using objectid(W) as keys, which does not change if the values in W are mutated).

This implicit style is the one presently used by Flux.jl, a closely related machine learning library. It uses structs like Linear above to define layers, and the function Flux.params(model) returns a Params object containing all the parameters of all layers. See its documentation for more details. When using Zygote for most other purposes, however, the explicit style is usually preferred.