Improving Zygote autodiff perfmance #291
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Trying the example in the issue I get this error and I can't find out why: |
|
At least now trying to call |
|
I finally got there for using DataInterpolations
using Zygote
using Plots
using Random
using ColorSchemes
Random.seed!(3)
pl = plot()
n = 10
t = collect(1.0:10)
u_exact = rand(n)
A_exact = LinearInterpolation(u_exact, t)
t_data = 1 .+ 9rand(10n)
u_data = A_exact.(t_data) .+ (rand(10n) .- 0.5) / 8
scatter!(t_data, u_data; label = "perturbed data")
u_fit = rand(n)
function loss(u_fit)
A = LinearInterpolation(u_fit, t)
values = A.(t_data)
return sum((values - u_data) .^ 2)
end
lr = 1e-2
N = 10
for (i, color) in enumerate(cgrad(:jet, range(0, 1, length = N)))
∇loss = only(gradient(loss, u_fit))
u_fit .-= lr * ∇loss
loss_it = loss(u_fit, u_data, t_data, t)
plot!(LinearInterpolation(u_fit, t); color, label = "Iteration $i, loss = $(round(loss_it, digits = 3))")
end
pl
|
|
@ChrisRackauckas @marcobonici My findings so far: Trying to calculate the gradient w.r.t. Apart from some minor refactors, the main thing I had to do to make the code compatible with |
|
Also got it to work for using DataInterpolations
using Zygote
using ForwardDiff
using Plots
using Random
using ColorSchemes
Random.seed!(3)
pl = plot(legendfontsize = 5)
method = QuadraticSpline
n = 10
t = collect(1.0:10)
u_exact = rand(n)
A_exact = method(u_exact, t)
t_data = 1 .+ 9rand(10n)
u_data = A_exact.(t_data) .+ (rand(10n) .- 0.5) / 10
scatter!(t_data, u_data; label = "perturbed data")
u_fit = rand(n)
function loss(u_fit)
A = method(u_fit, t)
values = A.(t_data)
return sum((values - u_data) .^ 2)
end
lr = 1e-3
N = 200
plot_update = 10
for (i, color) in enumerate(cgrad(:jet, range(0, 1, length = N)))
∇loss = only(Zygote.gradient(loss, u_fit))
∇loss_fd = ForwardDiff.gradient(loss, u_fit)
u_fit .-= lr * ∇loss
loss_it = loss(u_fit)
if i % plot_update == 1
plot!(
method(u_fit, t);
color,
label = "Iteration $i, loss = $(round(loss_it, digits = 3))",
)
end
@assert ∇loss ≈ ∇loss_fd
end
pl
|
|
@marcobonici this is already an order of magnitude better than what you reported for the gradient in your example: However, it leaves much to be desired. Most time is spent on calculating the cached parameters in the constructor, which it seems |
|
Have some more speedup: |
|
Closed in favor of #315 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes #289.
Checklist
contributor guidelines, in particular the SciML Style Guide and
COLPRAC.