Add utility function to generate neural ode problem #38

BSnelling · 2025-10-20T09:16:14Z

(also adds a .gitignore)

sebapersson · 2025-10-20T13:36:14Z

src/python/petab_sciml/problem_utils/neural_ode.py

+from typing import Iterable
+from ruamel.yaml import YAML
+
+def generate_neural_ode_problem(


A few things:

This function should, as optional argument, take experimental conditions.

It should not have default names for measurements and neural network file right, as these need to exist?

How about allowing users to, optionally, provide initial values for the species following valid PEtab math syntax?

sebapersson · 2025-10-20T13:38:18Z

@dilpath could you also take a look at this (you are a bit more familiar with Python than I am)

dilpath

Overall I like it, but could you specify the target audience/intended use case? Perhaps prepare a short docs page that describes it's use in the docs Examples? That would help me understand the purpose better and make it easier to review.

src/python/petab_sciml/problem_utils/neural_ode.py

dilpath · 2025-10-21T18:21:14Z

src/python/petab_sciml/problem_utils/neural_ode.py

+from typing import Iterable
+from ruamel.yaml import YAML
+
+def generate_neural_ode_problem(


Is the target audience for this function, people who have never used PEtab before? If so then looks OK overall. If not, then I would prefer a different design

Can you elaborate on different design?

I would argue target is both experienced PEtab users, but especially those newer to the format.

Ah, this is a duplicate comment of what we discussed (i.e., the basic function will just accept a pytorch model + list of species IDs).

For experienced PEtab users, this might be cumbersome -- they might prefer the modular design we discussed.

dilpath · 2025-10-21T18:22:33Z

src/python/petab_sciml/problem_utils/neural_ode.py

+        s = model.createSpecies()
+        s.setId(species)
+        s.setConstant(False)
+        s.setInitialAmount(0.5)


Why 0.5? 0 or 1? Or something on a similar scale to the first measurement?

src/python/petab_sciml/problem_utils/neural_ode.py

dilpath · 2025-10-21T18:39:16Z

src/python/petab_sciml/problem_utils/neural_ode.py

+    observables = {
+        "observableId": observable_ids,
+        "observableFormula": species,
+        "noiseFormula": [0.05] * len(species),


src/python/petab_sciml/problem_utils/neural_ode.py

dilpath · 2025-10-21T18:47:42Z

src/python/petab_sciml/problem_utils/neural_ode.py

+    parameters = {
+        "parameterId": [f"{network_name}_ps"],
+        "parameterScale": ["lin"],
+        "lowerBound": ["-inf"],
+        "upperBound": ["inf"],
+        "nominalValue": [None],
+        "estimate": [1],
+    }


Add priors to initialize these to be small? I think the bounds should also be finite.

I do not really see how bounds on neural network parameters make sense, as usually they are estimated unconstrained, and algorithms for estimating them (e.g., Adam) does not handle constraints in its native form?

Yes you're right. I rather meant for initialization, there should at least be bounds or something like a prior that can be sensibly sampled from. Unbounded uniform distribution for initialization would be ... difficult.

Or is initialization out-of-scope?

As initialization priors will be out of the scope of PEtab v2, I think it it outside the scope here as well.

dilpath · 2025-10-21T18:48:36Z

src/python/petab_sciml/problem_utils/neural_ode.py

+            {
+                "model_files": {
+                    "model": {
+                        "location": model_filename,
+                        "language": "sbml",
+                    },    
+                },
+                "measurement_files": [measurements_filename],
+                "observable_files": ["observables.tsv"],
+                "condition_files": ["conditions.tsv"],
+                "mapping_files": ["mapping.tsv"],
+            }


It looks like you are using a mix of both PEtab v1 and v2 here. We should target v2 but fine for now if you're working with the current state of PEtab v2 and the libpetab-python library.

I think it is fine for now. Currently the test-suite is something of a mix as v2 is not done, but PEtab SciML will target v2 in the end. Also updating the test-suite, and these examples will also be straightforward later on as we are not really using features that will be hard to port.

As a side-note, including events later in this example could be great to show the benefit with PEtab SciML.

dilpath · 2025-10-21T18:52:43Z

src/python/tests/problem_utils/test_neural_ode.py

+        with open("hybridization.tsv", "r") as f:
+            # number of lines is header plus inputs and outputs for each species
+            assert len(f.readlines()) == 1 + len(species) * 2


Replace with expected content of the files, since they are not too long? Would make the tests more understandable.

sebapersson · 2025-10-21T19:28:45Z

Overall I like it, but could you specify the target audience/intended use case? Perhaps prepare a short docs page that describes it's use in the docs Examples? That would help me understand the purpose better and make it easier to review.

@BSnelling knows more about the doc, but in #37 there should be a how-to-guide on neural ODE. But briefly, I see it as a convenient way to reproducible setup neural ODEs models, and to also benefit from the PEtab features such as experimental conditions, events (with v2). We should probably highlight these things in the designated doc page more clearly (that is, why it is nice to use PEtab SciML for neural ODEs). As for target audience, would be nice if we could more generally capture neural ODE practioners.

dilpath · 2025-10-21T19:49:54Z

Overall I like it, but could you specify the target audience/intended use case? Perhaps prepare a short docs page that describes it's use in the docs Examples? That would help me understand the purpose better and make it easier to review.

@BSnelling knows more about the doc, but in #37 there should be a how-to-guide on neural ODE. But briefly, I see it as a convenient way to reproducible setup neural ODEs models, and to also benefit from the PEtab features such as experimental conditions, events (with v2). We should probably highlight these things in the designated doc page more clearly (that is, why it is nice to use PEtab SciML for neural ODEs). As for target audience, would be nice if we could more generally capture neural ODE practioners.

Thanks, those docs look good, and makes sense to advertise PEtab a bit there too to a potential non-PEtab audience.

convenient way to reproducible setup neural ODEs models

As for target audience, would be nice if we could more generally capture neural ODE practioners.

For me it would make sense to have a function that takes the PEtab SciML NNModel and a list of species IDs, then generates an appropriate SBML file and mapping/hybridization tables.

I don't see much benefit from including measurements in this PR, so would prefer to remove that code, unless the current PR directly addresses a very popular use case for neural ODEs.

sebapersson · 2025-10-21T20:00:44Z

For me it would make sense to have a function that takes the PEtab SciML NNModel and a list of species IDs, then generates an appropriate SBML file and mapping/hybridization tables.

Good point. The argument for measurements here is for creating a complete PEtab Problem (including PEtab problem yaml), overall reducing the work of users. Might be worthwhile to go with both, one function for building the SBML, mapping and hybridization table, then a function based on it that given measurements (and other tables) can build the full problem?

dilpath · 2025-10-21T20:54:45Z

For me it would make sense to have a function that takes the PEtab SciML NNModel and a list of species IDs, then generates an appropriate SBML file and mapping/hybridization tables.

Good point. The argument for measurements here is for creating a complete PEtab Problem (including PEtab problem yaml), overall reducing the work of users. Might be worthwhile to go with both, one function for building the SBML, mapping and hybridization table, then a function based on it that given measurements (and other tables) can build the full problem?

Sounds good! I think it would be useful because this function would generate all of the PEtab SciML-specific files based on just the pytorch model and a list of species. Then if a user already knows PEtab, they can do the rest, or else use the measurements function.

BSnelling · 2025-10-23T09:08:30Z

Thank you both for the reviews! The two function approach makes sense to me: one for the model, network, hybridization and mapping and a second for measurements and other tables. I will implement that and come back for re-review.

BSnelling added 8 commits October 13, 2025 09:38

write sbml for neural ode

23e6848

write petab tsv and yaml files

91e6a87

unit tests and ruff

bf7abee

gitignore

9404535

rm tracked __pycache__

80d46a7

more optional kwwargs and docstrings

da9278f

fix yaml indentation issue

36ca07a

rm units

f094946

sebapersson requested changes Oct 20, 2025

View reviewed changes

sebapersson requested a review from dilpath October 20, 2025 13:37

dilpath requested changes Oct 21, 2025

View reviewed changes

Add utility function to generate neural ode problem #38

Are you sure you want to change the base?

Add utility function to generate neural ode problem #38

Uh oh!

Conversation

BSnelling commented Oct 20, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sebapersson commented Oct 20, 2025

Uh oh!

dilpath left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dilpath Oct 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dilpath Oct 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sebapersson commented Oct 21, 2025

Uh oh!

dilpath commented Oct 21, 2025

Uh oh!

sebapersson commented Oct 21, 2025

Uh oh!

dilpath commented Oct 21, 2025

Uh oh!

BSnelling commented Oct 23, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

dilpath Oct 21, 2025 •

edited

Loading

dilpath Oct 21, 2025 •

edited

Loading