Parametrize offset models

[CeliaBenquet]

• Parametrize offset models into a single OffsetNModel class so that we don't have to recreate a specific model for each new receptive field wanted. Instead, it can just be added to the list of parameters in the parametrize decorator.
• This handles both odd and even receptive field sizes.
• Also restructured the tests/ folder so that deprecation test code is grouped in a subfolder.

[stes]

Good initiative, thanks --- but add tests that confirm that the previously hardcoded models are 1:1 the same. You can do this by:

• copy pasting the old model code to the test folder as reference implementation
• instantiate with the new code, and check that all kinds of properties match 1:1

If this is hard to achieve, the alternative is to just let old and new models co-exist :)

e.g. use "varoffset-10-model" or so for the new ones?

[MMathisLab]

@CeliaBenquet lets wrap this one up?

[Copilot]

Pull request overview

This PR refactors multiple offset model classes into a single parametrized OffsetNModel class that dynamically generates models with different receptive field sizes. This reduces code duplication by consolidating 7 separate model class definitions (Offset5Model, Offset10Model, Offset15Model, Offset20Model, Offset36Model, Offset40Model, Offset50Model) into one parameterized class that handles both odd and even receptive field sizes.

Changes:

• Introduced OffsetNModel class with parametrize decorator to generate models for receptive fields of 5, 10, 15, 18, 20, 31, 36, 40, and 50 samples
• Removed 7 redundant offset model class definitions
• Updated documentation to reflect the new available model names

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
cebra/models/model.py	Replaced multiple offset model classes with a single parametrized OffsetNModel class and removed deprecated implementations
docs/source/usage.rst	Updated the example output to show the new parametrized model names available

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[CeliaBenquet]

I got this error, related to my changes, but results from an existing latent bug in registry.py:

FAILED tests/test_solver.py::test_single_session[demo-discrete-offset10-model-DiscreteDataLoader-SingleSessionSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
FAILED tests/test_solver.py::test_single_session[demo-continuous-offset10-model-ContinuousDataLoader-SingleSessionSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
FAILED tests/test_solver.py::test_single_session[demo-mixed-offset10-model-MixedDataLoader-SingleSessionSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
FAILED tests/test_solver.py::test_single_session_hybrid[demo-continuous-offset10-model-HybridDataLoader-SingleSessionHybridSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
FAILED tests/test_solver.py::test_multi_session[demo-discrete-multisession-offset10-model-DiscreteMultiSessionDataLoader-MultiSessionSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
FAILED tests/test_solver.py::test_multi_session[demo-continuous-multisession-offset10-model-ContinuousMultiSessionDataLoader-MultiSessionSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
FAILED tests/test_solver.py::test_unified_session[demo-continuous-unified-offset10-model-UnifiedLoader-UnifiedSolver] - TypeError: OffsetNModel.__init__() got multiple values for argument 'num_neurons'
= 7 failed, 1525 passed, 1399 skipped, 31 deselected, 108 warnings in 72.44s (0:01:12) =
make: *** [Makefile:45: test_reduced] Error 1
Error: Process completed with exit code 2.

This was due to the @parametrize wrapper. The default_kwargs dictionary is updated in-place rather than constructing one per call. Appears here because I instantiate the OffsetNModel class multiple times across tests.

Solved by removing the in-place update() and constructing a "fresh" kwargs per instance instead.