BPINN PDE solver #745

AstitvaAggarwal · 2023-10-01T21:37:51Z

for #712 and #205

…PDE.jl into Bpinn_pde

src/discretize.jl

ChrisRackauckas · 2023-10-06T14:27:06Z

The docs successfully built on master, so I put this on top of a successful master.

ChrisRackauckas · 2023-10-07T07:27:28Z

That looks like a legitimate test failure. You may want to split BPINN PDE to a new test group.

…PDE.jl into Bpinn_pde

src/discretize.jl

ChrisRackauckas · 2024-01-05T14:49:12Z

src/advancedHMC_MCMC.jl

+        @info("Sampling Complete.")
+        @info("Current Physics Log-likelihood : ", physloglikelihood(ℓπ, samples[end]))
+        @info("Current Prior Log-likelihood : ", priorweights(ℓπ, samples[end]))
+        @info("Current MSE against dataset Log-likelihood : ",
+            L2LossData(ℓπ, samples[end]))


Is this intended?

yes, helps in choosing better set of parameters and priors

Seems that should be returned information rather than just printed

i could do that, but the info would be returned only after the sampling. The reason ive chosen to display it is so that at the beginning of the sampling the values of different loglikelihoods are visible and if proportion wrt to each other is off (weak priors will need us to have larger L2 loglikelihoods than residual loglikelihoods) we can simply exit sampling and adjust stds,etc instead of waiting out the whole sampling

the final values of loglikelihoods shows which objective has been optimized more.

ChrisRackauckas · 2024-01-05T14:50:00Z

src/advancedHMC_MCMC.jl

+function vector_to_parameters(ps_new::AbstractVector,
+        ps::Union{NamedTuple, ComponentArrays.ComponentVector, AbstractVector})
+    if (ps isa ComponentArrays.ComponentVector) || (ps isa NamedTuple)


Use dispatch...

ChrisRackauckas · 2024-01-05T14:51:50Z

src/PDE_BPINN.jl

+        # multioutput case for Lux chains, for each depvar ps would contain Lux ComponentVectors
+        # which we use for mapping current ahmc sampled vector of parameters onto NNs
+        i = 0
+        Luxparams = Vector{ComponentArrays.ComponentVector}()


This is non-concrete and will slow things down.

ChrisRackauckas · 2024-01-05T14:55:10Z

src/PDE_BPINN.jl

+#         Tar.allstd)
+# end
+
+function setparameters(Tar::PDELogTargetDensity, θ)


Doesn't lux have a utility for the transformation of vector into componentvector for this kind of thing?

i couldnt find it in the docs(atleaast something specific to this case), so i worked it out

Let's make sure an issue get's opened about this after merging.

Typically you store the axes of the ComponentVector via ax = getaxes(ps_ca) and then construct the CA via ComponentArray(theta, ax)

ChrisRackauckas · 2024-01-05T14:55:45Z

src/PDE_BPINN.jl

+    if Kernel == HMCDA
+        δ, λ = MCMCkwargs[:δ], MCMCkwargs[:λ]
+        Kernel(δ, λ)
+    elseif Kernel == NUTS


That won't match if it's a type. Needs to use isa

Kernel here is the constructor name, the constructors are exported via AdvancedHMC.jl

yes but why on the datatype? That's going to ruin inference, and doesn't allow kernel-specific parameters

ChrisRackauckas · 2024-01-05T14:56:51Z

src/PDE_BPINN.jl

+    if nchains > Threads.nthreads()
+        throw(error("number of chains is greater than available threads"))


Why wouldn't that be allowed?

sampling for each chain is done on a single thread, https://turinglang.org/AdvancedHMC.jl/stable/#Parallel-sampling

but you could still have main chains than threads. Nothing stops that, it just batches

src/PDE_BPINN.jl

ChrisRackauckas · 2024-01-05T14:58:17Z

Generally looking good, though there's some performance issues to clean up and some issues with style not using dispatch when it would clean things up

ChrisRackauckas · 2024-01-07T01:45:37Z

src/discretize.jl

+    if bayesian
+        # required as Physics loss also needed on dataset domain points
+        pde_loss_functions1, bc_loss_functions1 = if !(dataset_given isa Nothing)
+            if !(strategy isa  GridTraining)


why would only gridtraining be allowed? Wouldn't you just add these points?

yeah, the points would get included into the training sets, GridTraining strategy is used for loss on dataset domain points anyways. So i was thinking GridTraining throughout domain is better than, having gridtraining for dataset points and Stochastic/etc on other points

ChrisRackauckas · 2024-01-07T01:59:51Z

src/discretize.jl

@@ -401,7 +401,7 @@ to the PDE.
 For more information, see `discretize` and `PINNRepresentation`.
 """
 function SciMLBase.symbolic_discretize(pde_system::PDESystem,
-                                       discretization::PhysicsInformedNN)
+    discretization::PhysicsInformedNN; bayesian::Bool = false,dataset_given=nothing)


This is missing from the docstring.

oh actually, this is an interface break.

ChrisRackauckas · 2024-01-07T02:05:46Z

function SciMLBase.symbolic_discretize(pde_system::PDESystem,
    discretization::PhysicsInformedNN; bayesian::Bool = false,dataset_given=nothing)

that's an interface break, and not one that makes sense. What would bayesian or dataset_given mean for a finite difference method? Those cannot be added to the high level interface as that's not well thought out in a general sense.

It's probably best to make a new discretization type and then have an AbstractPINNDiscretizer that is used in order to make it easy to reuse dispatches. bayesian wouldn't be needed then. I'm not entirely sure about dataset_given though, that should be known by the PDESystem or discretization?

ChrisRackauckas · 2024-01-07T02:09:35Z

Almost there.

So I think make the discretizer object so the interface is kept, and then open issues for the following:

BPINNs on PDEs needs a tutorial and other docs
The vector to parameters stuff is a bit slow for type stability issues, but we can merge without handling that since I don't think the functionality should be here anyways, so we should have an issue about reusing some Lux functionality for this.

ChrisRackauckas · 2024-01-07T17:29:55Z

src/discretize.jl

@@ -401,7 +401,7 @@ to the PDE.
 For more information, see `discretize` and `PINNRepresentation`.
 """
 function SciMLBase.symbolic_discretize(pde_system::PDESystem,
-                                       discretization::PhysicsInformedNN)
+    discretization)


This is type piracy. It needs to specify the top level dispatch.

missed out on this, thanks

src/discretize.jl

AstitvaAggarwal · 2024-01-07T19:11:08Z

awesome, thanks @ChrisRackauckas @Vaibhavdixit02 and @xtalax.

AstitvaAggarwal added 4 commits October 2, 2023 03:01

ready player one

59c14a2

ready player one

8d4d4cc

Merge branch 'Bpinn_pde' of https://github.com/AstitvaAggarwal/Neural…

aba99b4

…PDE.jl into Bpinn_pde

skeleton of it all

eb2c7ff

xtalax reviewed Oct 3, 2023

View reviewed changes

src/discretize.jl Show resolved Hide resolved

AstitvaAggarwal and others added 3 commits October 4, 2023 17:52

minor corrections

eee560a

minor fixes

6960497

Merge branch 'SciML:master' into Bpinn_pde

8c6a46f

ChrisRackauckas force-pushed the Bpinn_pde branch from 8c6a46f to 8684c57 Compare October 6, 2023 14:23

AstitvaAggarwal added 4 commits October 6, 2023 22:50

ready player one

a642c1a

skeleton of it all

d533d5f

minor corrections

8ea0cb3

minor fixes

3e7e784

ChrisRackauckas force-pushed the Bpinn_pde branch from 8684c57 to 3e7e784 Compare October 6, 2023 20:51

AstitvaAggarwal and others added 14 commits October 8, 2023 11:33

Merge branch 'Bpinn_pde' of https://github.com/AstitvaAggarwal/Neural…

9d39deb

…PDE.jl into Bpinn_pde

works but dosent sample correctly

b5cabb8

Merge branch 'SciML:master' into Bpinn_pde

3892b3f

Merge branch 'Bpinn_pde' of https://github.com/AstitvaAggarwal/Neural…

e7dcfbd

…PDE.jl into Bpinn_pde

works for non adaptive Forward solving

5e263eb

parameter estimation works for Flux chains

4b93047

Inverse problem solving now works for Lux chains

805380c

THE PDE SOLVER WORKS COMPLETELY

a2a2292

testing partially complete, minor changes overall

cf2faa4

minor changes

9069028

include dataset points into physicsloglikelhiood

98e0a94

overlapping points losses added

d48b1c8

Added user inference options(stand in for api)

dc3dceb

Merge branch 'SciML:master' into Bpinn_pde

f91adc3

AstitvaAggarwal added 3 commits January 5, 2024 18:11

update runtests.jl

659ef59

changed Mvnormal def in pde Loglikelihoods

4ef549d

TIL: avoid depreciated methods on CI

b02e97c

ChrisRackauckas reviewed Jan 5, 2024

View reviewed changes

src/discretize.jl Outdated Show resolved Hide resolved

ChrisRackauckas reviewed Jan 5, 2024

View reviewed changes

src/discretize.jl Show resolved Hide resolved

ChrisRackauckas reviewed Jan 5, 2024

View reviewed changes

src/PDE_BPINN.jl Outdated Show resolved Hide resolved

Changes from reviews(dispatch etc)

ec9048d

ChrisRackauckas reviewed Jan 7, 2024

View reviewed changes

AstitvaAggarwal added 5 commits January 7, 2024 20:36

Can use diff training_strategies, Implemented dispatch on discretization

7b1ea42

update discretize.jl

4683d74

update BPINN_PDE_tests.jl

185e2d5

update PDE_BPINN.jl

1c64bec

update BPINN_PDEinvsol_tests.jl

cf3dc1f

ChrisRackauckas reviewed Jan 7, 2024

View reviewed changes

src/discretize.jl Outdated Show resolved Hide resolved

Update src/discretize.jl

8008f3a

ChrisRackauckas approved these changes Jan 7, 2024

View reviewed changes

ChrisRackauckas merged commit 0687aaf into SciML:master Jan 7, 2024
10 of 12 checks passed

AstitvaAggarwal deleted the Bpinn_pde branch January 7, 2024 19:25

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

BPINN PDE solver #745

BPINN PDE solver #745

AstitvaAggarwal commented Oct 1, 2023

ChrisRackauckas commented Oct 6, 2023

ChrisRackauckas commented Oct 7, 2023

ChrisRackauckas Jan 5, 2024

AstitvaAggarwal Jan 5, 2024 •

edited

Loading

ChrisRackauckas Jan 5, 2024

AstitvaAggarwal Jan 5, 2024 •

edited

Loading

AstitvaAggarwal Jan 5, 2024

ChrisRackauckas Jan 5, 2024

ChrisRackauckas Jan 5, 2024

ChrisRackauckas Jan 5, 2024

AstitvaAggarwal Jan 5, 2024

ChrisRackauckas Jan 6, 2024

ChrisRackauckas Jan 7, 2024

avik-pal Jan 8, 2024

ChrisRackauckas Jan 5, 2024

AstitvaAggarwal Jan 5, 2024

ChrisRackauckas Jan 5, 2024

ChrisRackauckas Jan 5, 2024

AstitvaAggarwal Jan 5, 2024

ChrisRackauckas Jan 5, 2024

ChrisRackauckas commented Jan 5, 2024

ChrisRackauckas Jan 7, 2024

AstitvaAggarwal Jan 7, 2024 •

edited

Loading

ChrisRackauckas Jan 7, 2024

ChrisRackauckas Jan 7, 2024

ChrisRackauckas commented Jan 7, 2024

ChrisRackauckas commented Jan 7, 2024

ChrisRackauckas Jan 7, 2024

AstitvaAggarwal Jan 7, 2024

AstitvaAggarwal commented Jan 7, 2024

		if nchains > Threads.nthreads()
		throw(error("number of chains is greater than available threads"))

BPINN PDE solver #745

BPINN PDE solver #745

Conversation

AstitvaAggarwal commented Oct 1, 2023

ChrisRackauckas commented Oct 6, 2023

ChrisRackauckas commented Oct 7, 2023

Choose a reason for hiding this comment

AstitvaAggarwal Jan 5, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AstitvaAggarwal Jan 5, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ChrisRackauckas commented Jan 5, 2024

Choose a reason for hiding this comment

AstitvaAggarwal Jan 7, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ChrisRackauckas commented Jan 7, 2024

ChrisRackauckas commented Jan 7, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AstitvaAggarwal commented Jan 7, 2024

AstitvaAggarwal Jan 5, 2024 •

edited

Loading

AstitvaAggarwal Jan 5, 2024 •

edited

Loading

AstitvaAggarwal Jan 7, 2024 •

edited

Loading