Skip to content

Mark OptimiserChain as @functor and improve inference for apply! #115

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Nov 27, 2022
Merged

Mark OptimiserChain as @functor and improve inference for apply! #115

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
41a0dbb
mark `OptimiserChain` with `@functor`, and improve type inference for…
jondeuce Oct 25, 2022
1735625
fix doc tests
jondeuce Oct 25, 2022
9ac27ff
fix more doctests
jondeuce Oct 26, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 5 additions & 5 deletions src/adjust.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ through training.
To change just the learning rate, provide a number `η::Real`.

# Example
```jldoctest
```jldoctest adjust
julia> m = (vec = rand(Float32, 2), fun = sin);

julia> st = Optimisers.setup(Nesterov(), m) # stored momentum is initialised to zero
Expand All @@ -27,18 +27,18 @@ julia> st = Optimisers.adjust(st, 0.123) # change learning rate, stored momentu
To change other parameters, `adjust` also accepts keyword arguments matching the field
names of the optimisation rule's type.

```
```jldoctest adjust
julia> fieldnames(Adam)
(:eta, :beta, :epsilon)

julia> st2 = Optimisers.setup(OptimiserChain(ClipGrad(), Adam()), m)
(vec = Leaf(OptimiserChain(ClipGrad{Float32}(10.0), Adam{Float32}(0.001, (0.9, 0.999), 1.19209f-7)), [nothing, (Float32[0.0, 0.0], Float32[0.0, 0.0], (0.9, 0.999))]), fun = nothing)
(vec = Leaf(OptimiserChain(ClipGrad{Float32}(10.0), Adam{Float32}(0.001, (0.9, 0.999), 1.19209f-7)), (nothing, (Float32[0.0, 0.0], Float32[0.0, 0.0], (0.9, 0.999)))), fun = ())
Copy link
Member

@ToucheSir ToucheSir Oct 26, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wouldn't have expected fun = nothing to change to fun = () with this PR. Is this the output you see locally? If so, can you change this codeblock to a jldoctest so that we catch changes like this in the future?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Changed by #106 I think.

I guess we just need to remember the syntax for naming a doctest block, so that this one uses m, st from the one before.

Copy link
Contributor Author

@jondeuce jondeuce Oct 26, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes that change does not arise from this PR, sorry for the confusion. I changed nothing to () for consistency with the doctest above it; I can lookup the syntax to make that block into a proper doctest, as well @mcabbott.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed now.


julia> Optimisers.adjust(st2; beta = (0.777, 0.909), delta = 11.1) # delta acts on ClipGrad
(vec = Leaf(OptimiserChain(ClipGrad{Float32}(11.1), Adam{Float32}(0.001, (0.777, 0.909), 1.19209f-7)), [nothing, (Float32[0.0, 0.0], Float32[0.0, 0.0], (0.9, 0.999))]), fun = nothing)
(vec = Leaf(OptimiserChain(ClipGrad{Float32}(11.1), Adam{Float32}(0.001, (0.777, 0.909), 1.19209f-7)), (nothing, (Float32[0.0, 0.0], Float32[0.0, 0.0], (0.9, 0.999)))), fun = ())

julia> Optimisers.adjust(st; beta = "no such field") # silently ignored!
(vec = Leaf(Nesterov{Float32}(0.001, 0.9), Float32[-0.016, -0.088]), fun = nothing)
(vec = Leaf(Nesterov{Float32}(0.123, 0.9), Float32[-0.016, -0.088]), fun = ())
```
"""
adjust(tree, eta::Real) = map(st -> adjust(st, eta), tree)
Expand Down
14 changes: 7 additions & 7 deletions src/rules.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -607,7 +607,7 @@ julia> o = OptimiserChain(ClipGrad(1.0), Descent(0.1));
julia> m = (zeros(3),);

julia> s = Optimisers.setup(o, m)
(Leaf(OptimiserChain(ClipGrad{Float64}(1.0), Descent{Float64}(0.1)), [nothing, nothing]),)
(Leaf(OptimiserChain(ClipGrad{Float64}(1.0), Descent{Float64}(0.1)), (nothing, nothing)),)

julia> Optimisers.update(s, m, ([0.3, 1, 7],))[2] # clips before discounting
([-0.03, -0.1, -0.1],)
Expand All @@ -618,15 +618,15 @@ struct OptimiserChain{O<:Tuple} <: AbstractRule
end
OptimiserChain(opts...) = OptimiserChain(opts)

init(o::OptimiserChain, x::AbstractArray) = [init(opt, x) for opt in o.opts]
@functor OptimiserChain

init(o::OptimiserChain, x::AbstractArray) = map(opt -> init(opt, x), o.opts)

function apply!(o::OptimiserChain, states, x, dx, dxs...)
new_states = similar(states)
for (i, (opt, state)) in enumerate(zip(o.opts, states))
new_states[i], dx = apply!(opt, state, x, dx, dxs...)
foldl(tuple.(o.opts, states); init = ((), dx)) do (states′, dx′), (opt, state)
state′, dx′ = apply!(opt, state, x, dx′, dxs...)
return (states′..., state′), dx′
end

return new_states, dx
end

function Base.show(io::IO, c::OptimiserChain)
Expand Down