new solvers structure by ndem0 · Pull Request #777 · mathLab/PINA

ndem0 · 2026-03-19T15:38:51Z

Description

This PR fixes #782

Checklist

Code follows the project’s Code Style Guidelines
Tests have been added or updated
Documentation has been updated if necessary
Pull request is linked to an open issue

ndem0 · 2026-03-19T15:49:26Z

+            condition_losses[condition_name] = self._apply_reduction(
+                condition_loss_tensor
+            )


add weighting

Do not forget weightings!

dario-coscia · 2026-03-27T20:08:16Z

Just changed the title, cuz generation was making me think about generative models

GiovanniCanali

Thanks @ndem0 for implementing this substantial refactor of the solver logic. The overall direction looks good to me, and the change is definitely significant.

Before proceeding with the merge, though, I think there are still a few points that should be addressed. Some are minor, while others seem blocking to ensure consistency and compatibility with the current API. I am reporting here the ones I consider the most important; the remaining review comments are still useful for completeness.

Several solvers are still missing, especially PINN-related ones (CausalPINN, CompetitivePINN, GradientPINN, RBAPINN, SelfAdaptivePINN). We can reintroduce them in a later phase, but they should still be restored before the 0.3 release.
The same applies to the .rst documentation files, which currently do not yet reflect the new structure.
There are also a couple of files that do not appear to be used anywhere. I think we should clarify whether they are meant to be integrated into the new flow or removed to avoid dead code.
Regarding the new solver logic, I still see a few critical issues:
- loss aggregation weights currently seem to be missing;
- More importantly, condition.evaluate is not supported by all Condition types. At the moment, it is defined for TimeSeriesCondition and InputTargetCondition, but not for the other condition classes. In its current state, the logic inside optimization_cycle will break as soon as a different condition type is encountered. I noticed that BaseEquationCondition exists, but currently no condition inherits from it. This could perhaps be used to centralize the evaluate logic for equation-based conditions, although DataCondition would still remain uncovered.
- Finally, I thought we had agreed to remove the optimization_cycle / _optimization_cycle duality in favor of a unified logic shared across train/validation/test. If we instead want to keep both, I would still suggest revisiting the naming, because in the current form it may easily become confusing.

GiovanniCanali · 2026-05-11T07:28:40Z

To delete: this file has been split into test_data/test_graph_data_manager.py and test_data/test_tensor_data_manager.py.

GiovanniCanali · 2026-05-11T07:31:16Z

    Abstract interface for all losses.
    """

+    def __init__(self, reduction="mean"):


Remove __init__: interfaces should not implement methods. This is already declared and defined in the respective base class.

GiovanniCanali · 2026-05-11T07:33:06Z



-class LossInterface(_Loss, metaclass=ABCMeta):
+class DualLossInterface(_Loss, metaclass=ABCMeta):


Please, rename the file and update the import statements in the __init__.py file according to PINA conventions. The file should be renamed to dual_loss_interface.py.

GiovanniCanali · 2026-05-11T07:33:59Z

 from pina._src.core.label_tensor import LabelTensor
 from pina._src.core.utils import check_consistency
-from pina._src.loss.loss_interface import LossInterface
+from pina._src.loss.loss_interface import DualLossInterface as LossInterface


Please, import DualLossInterface without aliasing.

GiovanniCanali · 2026-05-11T07:34:41Z

-from pina._src.solver.physics_informed_solver.pinn_interface import (
-    PINNInterface,
-)
+from pina._src.solver.pinn import PINN as PINNInterface


Please, import PINN without aliasing.

GiovanniCanali · 2026-05-11T10:16:28Z

+                self.forward = (  # noqa: E731
+                    lambda x, _idx=idx: self.models[_idx].forward(x)
+                )
+                from pina._src.core.utils import labelize_forward


Move to import section

GiovanniCanali · 2026-05-11T10:18:12Z

Delete commented lines if they are unnecessary.

GiovanniCanali · 2026-05-11T10:21:28Z

Restore the following: AbstractProblem -> BaseProblem, Scheduler -> SchedulerInterface, Optimizer -> OptimizerInterface in the docs

GiovanniCanali · 2026-05-11T10:22:35Z

+        print('dio setup')
+        return super(PINN, self).setup(stage)
+
+    def validation_step(self, batch, **kwargs):


See PINNs for both validation and test steps

GiovanniCanali · 2026-05-11T10:32:58Z

+            condition = self.problem.conditions[condition_name]
+            condition_data = dict(data)
+
+            condition_loss_tensor = condition.evaluate(


The instruction condition.evaluate is not supported by all Condition types: it is defined for InputTargetCondition and TimeSeriesCondition, but it does not appear to be available for all other condition types.

ndem0 changed the title ~~0.3 solver~~ new generation solvers Mar 19, 2026

ndem0 commented Mar 19, 2026

View reviewed changes

dario-coscia changed the title ~~new generation solvers~~ new solvers structure Mar 27, 2026

ndem0 added the 0.3 Related to 0.3 release label Apr 8, 2026

GiovanniCanali mentioned this pull request Apr 14, 2026

New optimizers #618

Open

4 tasks

ndem0 force-pushed the 0.3-solver branch from 9b94d42 to 8a86eb5 Compare May 8, 2026 07:50

ndem0 added 9 commits May 8, 2026 09:52

evaluate for physics condition

6719eb2

input target evaluate

a873e28

first new generation solver

1450330

pinn and supervised

9dac226

autoregressive

b9a094c

multi model

a6c1b0e

ensemble solver

24fec4f

labelize_forward on ensemble

9b549f0

Clean inheritance

37f0755

ndem0 force-pushed the 0.3-solver branch from 8a86eb5 to 37f0755 Compare May 8, 2026 07:57

Fix conflicts

05f3059

ndem0 marked this pull request as ready for review May 8, 2026 08:26

ndem0 requested a review from a team as a code owner May 8, 2026 08:26

GiovanniCanali requested changes May 11, 2026

View reviewed changes

fix import

3d4c6e5



		class LossInterface(_Loss, metaclass=ABCMeta):
		class DualLossInterface(_Loss, metaclass=ABCMeta):

Conversation

ndem0 commented Mar 19, 2026 • edited by GiovanniCanali Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dario-coscia commented Mar 27, 2026

Uh oh!

GiovanniCanali left a comment

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!

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!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ndem0 commented Mar 19, 2026 •

edited by GiovanniCanali

Loading