Remove `model.check_bounds` in favor of graph introspection

[ricardoV94]

Description

In the following graph:

import pymc as pm

with pm.Model(check_bounds=True):
  sigma = pm.HalfNormal("sigma")
  y = pm.Normal("y", sigma=sigma)
m.compile_logp().fn.dprint()

There ends up being a switch(exp(i0) > 0, something, -inf) that disappears with check_bounds=False (one other stays regardless).

    ├─ Switch [id E]
    │  ├─ GE [id F]
    │  │  ├─ exp [id G] 't1'
    │  │  │  └─ i0 [id H]
    │  │  └─ 0.0 [id I]
    │  ├─ add [id J]
    │  │  ├─ -0.22579135264472738 [id K]
    │  │  └─ mul [id L]
    │  │     ├─ -0.5 [id M]
    │  │     └─ sqr [id N]
    │  │        └─ exp [id G] 't1'
    │  │           └─ ···
    │  └─ -inf [id O]

We could rewrite exp(x) >= 0) -> True in PyTensor. The one case it's not correct is if it's a nan originally. We could tag such rewrite as nan_unsafe, or we could promise pytensor that we are not going to propose nan for x. I prefer the first approach.

But some of the switches are more complex, such as IntervalTransform(0, 1), for unit variables. We could however check that rvs_to_transforms has the right transform that ensures the constraint, and remove the check like check_bounds does. It may be easier to do that, than to verify the computational graph implies the constraints.

These are a bit myopic points. The general idea is to make check_bounds redundant when a graph is well defined, and remove that flag from the model. Streamlining the UX

[Utkarshkarki]

I am interested in this problem and would give my insights or solution soon.

[ricardoV94]

This is not a beginner friendly or even fully defined issue. I suggest not picking this up

[Utkarshkarki]

I see two paths:
(1) implement smart check removal in PyMC using  rvs_to_transforms
(2) Contribute mathematical rewrites to PyTensor first.
Which would you prefer? Or should we do a hybrid approach—start with PyMC and contribute to PyTensor in parallel?

If you think my way of understanding is correct, we can move ahead with this one issue.

[Utkarshkarki]

but writing rewrites would be little tedious process

[ricardoV94]

Rewrites generalize better, different applications of pytensor will automatically benefit. OTOH, some are non trivial to figure out. I think the best is a hybrid approach.

On the rewrite level exp(x) >= 0 -> True is a good warmup exercise.
On the PyMC Model level just try something and see if complexity stays reasonable