Skip to content

Fix ADADelta calculations and broken tests not catching the problems #1299

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
bors merged 3 commits into from
Aug 6, 2020
Merged

Fix ADADelta calculations and broken tests not catching the problems #1299

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
22812e9
Fix ADADelta doing nothing
scimas Aug 2, 2020
2fbd2ce
Fix tests not testing
scimas Aug 4, 2020
50092c8
Increase learning rate of OADAM
scimas Aug 5, 2020
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
8 changes: 5 additions & 3 deletions src/optimise/optimisers.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -283,7 +283,7 @@ mutable struct OADAM
state::IdDict
end

OADAM(η = 0.0001, β = (0.5, 0.9)) = OADAM(η, β, IdDict())
OADAM(η = 0.001, β = (0.5, 0.9)) = OADAM(η, β, IdDict())

function apply!(o::OADAM, x, Δ)
η, β = o.eta, o.beta
Expand Down Expand Up @@ -357,7 +357,9 @@ function apply!(o::ADADelta, x, Δ)
ρ = o.rho
acc, Δacc = get!(o.state, x, (zero(x), zero(x)))
@. acc = ρ * acc + (1 - ρ) * Δ^2
@. Δ *= √Δacc/ (√acc + ϵ)
# DON'T remove epsilon from numerator
# or even out of the square roots
@. Δ *= √(Δacc + ϵ) / √(acc + ϵ)
@. Δacc = ρ * Δacc + (1 - ρ) * Δ^2
return Δ
end
Expand Down Expand Up @@ -599,4 +601,4 @@ function apply!(o::ClipNorm, x, Δ)
rmul!(Δ, o.thresh / Δnrm)
end
return Δ
end
end
7 changes: 7 additions & 0 deletions test/optimise.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,12 +2,17 @@ using Flux.Optimise
using Flux.Optimise: runall
using Flux: Params, gradient
using Test
using Random

@testset "Optimise" begin
# Ensure rng has different state inside and outside the inner @testset
# so that w and w' are different
Random.seed!(84)
w = randn(10, 10)
@testset for opt in [ADAMW(), ADAGrad(0.1), AdaMax(), ADADelta(0.9), AMSGrad(),
NADAM(), RADAM(), Descent(0.1), ADAM(), OADAM(), Nesterov(), RMSProp(),
Momentum()]
Random.seed!(42)
w′ = randn(10, 10)
loss(x) = Flux.Losses.mse(w*x, w′*x)
for t = 1: 10^5
Expand All @@ -21,8 +26,10 @@ using Test
end

@testset "Optimiser" begin
Random.seed!(84)
w = randn(10, 10)
@testset for Opt in [InvDecay, WeightDecay, ExpDecay]
Random.seed!(42)
w′ = randn(10, 10)
loss(x) = Flux.Losses.mse(w*x, w′*x)
opt = Optimiser(Opt(), ADAM(0.001))
Expand Down