Simplify Model new and metaclass #7473

get_context returns an instance of a Model, not a ContextMeta object We don't need the typevar, since we don't use it for anything special

All of these are supported on python>=3.9.

We create a global instance of it within this module, which is similar to how it worked before, where a `context_class` attribute was attached to the Model class. We inherit from threading.local to ensure thread safety when working with models on multiple threads. See pymc-devs#1552 for the reasoning. This is already tested in `test_thread_safety`.

UNSET is the instance of the _UnsetType type. We should be typing the latter here.

We use the new ModelManager.parent_context property to reliably set any parent context, or else set it to None.

We set this directly on the class as a classmethod, which is clearer than going via the metaclass.

The original function does not behave as I expected. In the following example I expected that it would return only the final model, not root. This method is not used anywhere in the pymc codebase, so I have dropped it from the codebase. I originally included the following code to replace it, but since it is not used anyway, it is better to remove it. ```python` @classmethod def get_contexts(cls) -> list[Model]: """Return a list of the currently active model contexts.""" return MODEL_MANAGER.active_contexts ``` Example for testing behaviour in current main branch: ```python import pymc as pm with pm.Model(name="root") as root: print([c.name for c in pm.Model.get_contexts()]) with pm.Model(name="first") as first: print([c.name for c in pm.Model.get_contexts()]) with pm.Model(name="m_with_model_None", model=None) as m_with_model_None: # This one doesn't make much sense: print([c.name for c in pm.Model.get_contexts()]) ```

We only keep the __call__ method, which is necessary to keep the model context itself active during that model's __init__.

@ricardoV94

In pymc/distributions/distribution.py, this change allows the type checker to infer that `rv_out` can only be a TensorVariable. Thanks to @ricardoV94 for type hint on rv_var.

I originally tried numpy's ArrayLike, replacing Sequence entirely, but then I realized that ArrayLike also allows non-sequences like integers and floats. I am not certain if `values="a string"` should be legal. With the type hint sequence, it is. Might be more accurate, but verbose to use `list | tuple | set | np.ndarray | None`.

…unction We don't want to allow the user to pass a `dims=[None, None]` to our function, but current behaviour set `dims=[None] * N` at the end of `determine_coords`. To handle this, I created a `new_dims` with a larger type scope which matches the return type of `dims` in `determine_coords`. Then I did the same within def Data to support this new type hint.

The only case where dims=[None, ...] is when the user has passed dims=None. Since the user passed dims=None, they shouldn't be expecting any coords to match that dimension. Thus we don't need to try to add any more coords to the model.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Simplify Model new and metaclass #7473

Simplify Model new and metaclass #7473

Commits on Oct 10, 2024

Simplify Model __new__ and metaclass #7473

Simplify Model __new__ and metaclass #7473

Commits on Oct 10, 2024

Simplify Model new and metaclass #7473

Simplify Model new and metaclass #7473