diff --git a/docs/src/training/optimisers.md b/docs/src/training/optimisers.md index 9222a0db2d..600cae4d88 100644 --- a/docs/src/training/optimisers.md +++ b/docs/src/training/optimisers.md @@ -4,53 +4,24 @@ CurrentModule = Flux # [Optimisers](@id man-optimisers) -Consider a [simple linear regression](../models/basics.md). We create some dummy data, calculate a loss, and backpropagate to calculate gradients for the parameters `W` and `b`. +Flux builds in many optimisation rules for use with [`train!`](@ref Flux.Optimise.train!) and +other training functions. -```julia -using Flux - -W = rand(2, 5) -b = rand(2) - -predict(x) = (W * x) .+ b -loss(x, y) = sum((predict(x) .- y).^2) +The mechanism by which these work is gradually being replaced as part of the change +from "implicit" dictionary-based to "explicit" tree-like structures. +At present, the same struct (such as `Adam`) can be used with either form, +and will be automatically translated. -x, y = rand(5), rand(2) # Dummy data -l = loss(x, y) # ~ 3 - -θ = Flux.params(W, b) -grads = gradient(() -> loss(x, y), θ) -``` +For full details of how the new "explicit" interface works, see the [Optimisers.jl documentation](https://fluxml.ai/Optimisers.jl/dev/). -We want to update each parameter, using the gradient, in order to improve (reduce) the loss. Here's one way to do that: - -```julia -η = 0.1 # Learning Rate -for p in (W, b) - p .-= η * grads[p] -end -``` +For full details on how the "implicit" interface worked, see the [Flux 0.13.6 manual](https://fluxml.ai/Flux.jl/v0.13.6/training/optimisers/#Optimiser-Interface). -Running this will alter the parameters `W` and `b` and our loss should go down. Flux provides a more general way to do optimiser updates like this. - -```julia -using Flux: update! - -opt = Descent(0.1) # Gradient descent with learning rate 0.1 - -for p in (W, b) - update!(opt, p, grads[p]) -end -``` - -An optimiser `update!` accepts a parameter and a gradient, and updates the parameter according to the chosen rule. We can also pass `opt` to our [training loop](training.md), which will update all parameters of the model in a loop. However, we can now easily replace `Descent` with a more advanced optimiser such as `Adam`. ## Optimiser Reference All optimisers return an object that, when passed to `train!`, will update the parameters passed to it. ```@docs -Flux.Optimise.update! Descent Momentum Nesterov @@ -67,44 +38,6 @@ OAdam AdaBelief ``` -## Optimiser Interface - -Flux's optimisers are built around a `struct` that holds all the optimiser parameters along with a definition of how to apply the update rule associated with it. We do this via the `apply!` function which takes the optimiser as the first argument followed by the parameter and its corresponding gradient. - -In this manner Flux also allows one to create custom optimisers to be used seamlessly. Let's work on this with a simple example. - -```julia -mutable struct Momentum - eta - rho - velocity -end - -Momentum(eta::Real, rho::Real) = Momentum(eta, rho, IdDict()) -``` - -The `Momentum` type will act as our optimiser in this case. Notice that we have added all the parameters as fields, along with the velocity which we will use as our state dictionary. Each parameter in our models will get an entry in there. We can now define the rule applied when this optimiser is invoked. - -```julia -function Flux.Optimise.apply!(o::Momentum, x, Δ) - η, ρ = o.eta, o.rho - v = get!(o.velocity, x, zero(x))::typeof(x) - @. v = ρ * v - η * Δ - @. Δ = -v -end -``` - -This is the basic definition of a Momentum update rule given by: - -```math -v = ρ * v - η * Δ -w = w - v -``` - -The `apply!` defines the update rules for an optimiser `opt`, given the parameters and gradients. It returns the updated gradients. Here, every parameter `x` is retrieved from the running state `v` and subsequently updates the state of the optimiser. - -Flux internally calls on this function via the `update!` function. It shares the API with `apply!` but ensures that multiple parameters are handled gracefully. - ## Composing Optimisers Flux defines a special kind of optimiser simply called `Optimiser` which takes in arbitrary optimisers as input. Its behaviour is similar to the usual optimisers, but differs in that it acts by calling the optimisers listed in it sequentially. Each optimiser produces a modified gradient diff --git a/docs/src/training/train_api.md b/docs/src/training/train_api.md index 5ef6ba12a1..0a6b41c934 100644 --- a/docs/src/training/train_api.md +++ b/docs/src/training/train_api.md @@ -1,10 +1,16 @@ # Training API - ```@docs Flux.Train.setup -Flux.Train.update! -Flux.Train.train! +Flux.Optimise.train!(loss, model, data, opt; cb) +``` + +The new version of Flux's training code was written as an independent package, called Optimisers.jl. +However, at present all Flux models contain parameter arrays (such as `Array`s and `CuArray`s) +which can be updated in-place. Thus objects returned by `update!` can be ignored. + +```@docs +Optimisers.update! ``` ## Implicit style @@ -15,14 +21,12 @@ Flux 0.13 is the transitional version which supports both. For full details on how to use the implicit style, see [Flux 0.13.6 manual](https://fluxml.ai/Flux.jl/v0.13.6/training/training/). - ```@docs Flux.params -Flux.Optimise.update! -Flux.Optimise.train! +Optimisers.update!(opt::Flux.Optimise.AbstractOptimiser, xs::Params, gs) +Flux.Optimise.train!(loss, ps::Flux.Params, data, opt::Flux.Optimise.AbstractOptimiser; cb) ``` - Note that, by default, `train!` only loops over the data once (a single "epoch"). A convenient way to run multiple epochs from the REPL is provided by `@epochs`. @@ -69,3 +73,4 @@ cb = function () accuracy() > 0.9 && Flux.stop() end ``` + diff --git a/src/Flux.jl b/src/Flux.jl index 80eb87df1e..dfb2645d19 100644 --- a/src/Flux.jl +++ b/src/Flux.jl @@ -34,9 +34,11 @@ export Descent, Adam, Momentum, Nesterov, RMSProp, AdamW, RAdam, AdaBelief, InvDecay, ExpDecay, WeightDecay, ClipValue, ClipNorm +export ClipGrad, OptimiserChain # these are const defined in deprecations, for ClipValue, Optimiser + include("train.jl") using .Train -# using .Train: setup, @train_autodiff +using .Train: setup using CUDA const use_cuda = Ref{Union{Nothing,Bool}}(nothing) diff --git a/src/optimise/optimisers.jl b/src/optimise/optimisers.jl index ce72a4b0ce..8a60ba97a5 100644 --- a/src/optimise/optimisers.jl +++ b/src/optimise/optimisers.jl @@ -564,6 +564,9 @@ end Combine several optimisers into one; each optimiser produces a modified gradient that will be fed into the next, and this is finally applied to the parameter as usual. + +!!! note + This will be replaced by `Optimisers.OptimiserChain` in Flux 0.14. """ mutable struct Optimiser <: AbstractOptimiser os::Vector{Any} @@ -699,6 +702,9 @@ end ClipValue(thresh) Clip gradients when their absolute value exceeds `thresh`. + +!!! note + This will be replaced by `Optimisers.ClipGrad` in Flux 0.14. """ mutable struct ClipValue{T} <: AbstractOptimiser thresh::T