ENH: Join looses static shape information

[michaelosthege]

Describe the issue:

A simple pt.concatenate of tensors with known static shapes could easily result in a static shape.

Currently it does not.

Reproducable code example:

def test_concatenate_static_shapes():
    a = pt.ones((2, 3))
    b = pt.ones((2, 5))
    c = pt.concatenate([a, b], axis=1)
    assert c.eval().shape == (2, 8)
    assert c.type.shape == (2, 8)
    pass

Error message:

>       assert c.type.shape == (2, 8)
E       assert (None, None) == (2, 8)
E         At index 0 diff: None != 2
E         Full diff:
E         - (2, 8)
E         + (None, None)

tests\test_setsubtensor.py:103: AssertionError

PyTensor version information:

main

Context for the issue:

No response

[ricardoV94]

That's not a bug. It's fine for an Op to lose shape information (as long as it's still correct). We can improve it of course.

Other parts of the library do assume shape is not lost (e.g., some asserts in rewrites). Those are bugs.