Skip to content
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

Jump to bottom

bump compat of AdvancedHMC #2050

Merged
merged 20 commits into from
Jul 29, 2023
Merged

bump compat of AdvancedHMC #2050

Show file tree
Hide file tree
Changes from 12 commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
7175c1d
bump compat of AdvancedHMC
JaimeRZP Jul 24, 2023
5e2ed5a
Update Project.toml
JaimeRZP Jul 27, 2023
01f835e
Merge branch 'master' into bump_AdvancedHMC
yebai Jul 27, 2023
9e130d2
upd make_ahmc_kernel
JaimeRZP Jul 27, 2023
ebeade8
bugs
JaimeRZP Jul 28, 2023
3045bb2
bugs
JaimeRZP Jul 28, 2023
a6838c1
bug
JaimeRZP Jul 28, 2023
d8ccd8d
Update Project.toml
yebai Jul 28, 2023
8f51898
Update Inference.jl
yebai Jul 29, 2023
0581174
Update Inference.jl
yebai Jul 29, 2023
bbd890c
Apply suggestions from code review
yebai Jul 29, 2023
476f90c
Update src/inference/Inference.jl
yebai Jul 29, 2023
1ade2fd
Update Inference.jl
yebai Jul 29, 2023
b3b1d33
Update Inference.jl
yebai Jul 29, 2023
05a2dba
Update Inference.jl
yebai Jul 29, 2023
1125f72
Update Project.toml
yebai Jul 29, 2023
a1fcb21
Update Inference.jl
yebai Jul 29, 2023
6838b4e
Typofix in test.
yebai Jul 29, 2023
307040e
Implement sample for external spl + ensemble
yebai Jul 29, 2023
2bf8beb
Revert default n_adapts value.
yebai Jul 29, 2023
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
4 changes: 2 additions & 2 deletions Project.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
name = "Turing"
uuid = "fce5fe82-541a-59a6-adf8-730c64b5f9a0"
version = "0.27"
version = "0.27.1"

[deps]
AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
Expand Down Expand Up @@ -38,7 +38,7 @@ Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"

[compat]
AbstractMCMC = "4"
AdvancedHMC = "0.3.0, 0.4"
AdvancedHMC = "0.5.2"
devmotion marked this conversation as resolved.
Show resolved Hide resolved
AdvancedMH = "0.6.8, 0.7"
AdvancedPS = "0.4"
AdvancedVI = "0.2"
Expand Down
6 changes: 3 additions & 3 deletions src/inference/Inference.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,10 +189,10 @@ function AbstractMCMC.sample(
kwargs...
)
if resume_from === nothing
return AbstractMCMC.mcmcsample(rng, model, sampler, N;
chain_type=chain_type, progress=progress, kwargs...)
return AbstractMCMC.mcmcsample(rng, model, sampler, N + min(div(N, 10), 1_000);
Copy link
Member

Choose a reason for hiding this comment

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

This means users will get more than the requested N samples - I think this should be reverted?

Copy link
Member

Choose a reason for hiding this comment

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

There is a change in how the number of adaptions is handled in the external sampler interface. Instead of storing n_adapts in the sampling algorithm (see here), we now only pass them in the AbstractMCMC.sample call. This means there is no adaption by default if the change is reverted.

Copy link
Member

Choose a reason for hiding this comment

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

Maybe we can set disgard_initial = n_adapts here so we have a default adaption but will return the same number of MCMC samples. In addition, we can allow the user to pass a n_adapt argument to override the default adaption settings.

Copy link
Member

Choose a reason for hiding this comment

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

Sure, we have to take into account the adaption steps - but changing this line here means that a call such as sample(model, NUTS(), 10) will return not 10 but 1010 samples? These discarded steps only have to be specified as keyword argument but not added to the positional argument (number of samples). AbstractMCMC will added discard_initial to these internally to the requested number of samples automatically.

Copy link
Member

@yebai yebai Jul 29, 2023 •
edited
Loading

Choose a reason for hiding this comment

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

It turns out we have a typo in tests: we already explicitly pass the n_adapts argument to the sample function. Due to the typo, AHMC complains about missing n_adapts. After correcting this typo, these default values are no longer necessary. More generally, we should introduce default options for n_adapts in the AbstractMCMC package; see this PR.

n_adapts = min(div(N, 10), 1_000), chain_type=chain_type, progress=progress, kwargs...)
else
return resume(resume_from, N; chain_type=chain_type, progress=progress, kwargs...)
return resume(resume_from, N; n_adapts = 0, chain_type=chain_type, progress=progress, kwargs...)
end
end

Expand Down
3 changes: 2 additions & 1 deletion src/inference/hmc.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -444,7 +444,8 @@ end
function make_ahmc_kernel(alg::HMCDA, ϵ)
return AHMC.HMCKernel(AHMC.Trajectory{AHMC.EndPointTS}(AHMC.Leapfrog(ϵ), AHMC.FixedIntegrationTime(alg.λ)))
end
make_ahmc_kernel(alg::NUTS, ϵ) = AHMC.NUTS(AHMC.Leapfrog(ϵ), alg.max_depth, alg.Δ_max)
make_ahmc_kernel(alg::NUTS, ϵ) =
AHMC.HMCKernel(AHMC.Trajectory{AHMC.MultinomialTS}(AHMC.Leapfrog(ϵ), AHMC.GeneralisedNoUTurn(alg.max_depth, alg.Δ_max)))

####
#### Compiler interface, i.e. tilde operators.
Expand Down
2 changes: 1 addition & 1 deletion test/contrib/inference/abstractmcmc.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@ function initialize_nuts(model::Turing.Model)
# - multinomial sampling scheme,
# - generalised No-U-Turn criteria, and
# - windowed adaption for step-size and diagonal mass matrix
proposal = AdvancedHMC.NUTS{AdvancedHMC.MultinomialTS,AdvancedHMC.GeneralisedNoUTurn}(integrator)
proposal = AdvancedHMC.HMCKernel(AdvancedHMC.Trajectory{AdvancedHMC.MultinomialTS}(integrator, AdvancedHMC.GeneralisedNoUTurn()))
adaptor = AdvancedHMC.StanHMCAdaptor(
AdvancedHMC.MassMatrixAdaptor(metric),
AdvancedHMC.StepSizeAdaptor(0.65, integrator)
Expand Down
Loading