Yet Another Automatic Differentiation package in Julia.

Press ] and use pkg mode in Julia REPL, then type:

pkg> add YAAD

You may want to check my blog post about it: Implement AD with Julia in ONE day

This project aims to provide a similar interface with PyTorch's autograd, while keeping things simple. The core implementation only contains a straight-forward 200 line of Julia. It is highly inspired by AutoGrad.jl and PyTorch

Every operation will directly return a CachedNode, which constructs a computation graph dynamically without using a global tape.

NOTE: This project is for self-use at the moment, it will be a place for me to do AD related experimental coding, I don't guarantee the consistency and stability between versions (different version can be in-compatible). For practical usage, I would suggest you try Flux.Tracker or Zygote. They may have better performance and are aimed to be non-experimental projects.

It is simple. Mark what you want to differentiate with Variable, which contains value and grad. Each time you try to backward evaluate, the gradient will be accumulated to grad.

using LinearAlgebra
x1, x2 = Variable(rand(30, 30)), Variable(rand(30, 30))
y = tr(x1 * x2) # you get a tracked value here
backward(y) # backward propagation
print(x1.grad) # this is where gradient goes

Or you can just register your own

# first define how you want to create a node in computation graph
sin(x::AbstractNode) = register(sin, x)

# then define the gradient
gradient(::typeof(sin), grad, output, x) = grad * cos(x)

Apache License Version 2.0