Skip to content

Stop training on Inf/NaN loss #2070

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 6 commits into from
Oct 16, 2022
Merged

Stop training on Inf/NaN loss #2070

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
bd8a4dd
stop training on Inf/NaN loss
mcabbott Sep 27, 2022
fe3735d
add a test
mcabbott Sep 27, 2022
77e52ea
improve test
mcabbott Sep 27, 2022
91a93b9
improve test
mcabbott Sep 27, 2022
36b1302
Update train.jl
mcabbott Oct 10, 2022
bacf997
Update optimise.jl
mcabbott Oct 10, 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
8 changes: 6 additions & 2 deletions src/optimise/train.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
using ProgressLogging: @progress, @withprogress, @logprogress
import Zygote: Params, gradient
import Zygote: Params, gradient, withgradient


"""
Expand Down Expand Up @@ -126,9 +126,13 @@ function train!(loss, ps::Params, data, opt::AbstractOptimiser; cb = () -> ())
n = (itrsz == Base.HasLength()) || (itrsz == Base.HasShape{1}()) ? length(data) : 0
@withprogress for (i, d) in enumerate(data)
try
gs = gradient(ps) do
l, gs = withgradient(ps) do
loss(batchmemaybe(d)...)
end
if !isfinite(l)
@warn "Loss is $l on item $i, stopping training"
break
end
update!(opt, ps, gs)
cb()
catch ex
Expand Down
12 changes: 12 additions & 0 deletions test/optimise.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,18 @@ end
Flux.train!(loss, Flux.params(r), (r,), Descent())
end

@testset "Stop on NaN" begin
m = Dense(1 => 1)
m.weight .= 0
CNT = 0
Flux.train!(Flux.params(m), 1:100, Descent(0.1)) do i
CNT += 1
i == 51 ? NaN32 : sum(m([1f0]))
end
@test CNT == 51 # stopped early
@test m.weight[1] ≈ -5 # did not corrupt weights
end

@testset "ExpDecay" begin

@testset "Sanity Check" begin
Expand Down