`convert_size` wrongly assumes symbolic sizes have to be scalars

[ricardoV94]

import pymc as pm
import aesara.tensor as at

s = at.scalar('s')
size = at.stack([s, s])
x = at.random.normal(size=size)
x.eval({s: 2})
# array([[-0.66285345,  0.97341598],
#        [ 0.05158132,  2.03399059]])

pm.Normal.dist(size=size)
# ValueError: The `size` parameter must be a tuple, TensorVariable, int or list. Actual: <class 'aesara.tensor.var.TensorVariable'>

The error comes from this check here:

pymc/pymc/distributions/shape_utils.py

Line 475 in 9155922

if isinstance(size, int) or (isinstance(size, TensorVariable) and size.ndim == 0):

CC @michaelosthege

[michaelosthege]

The functions signature is correct, but the implementation doesn't account for the TensorVariable case it advertised in the type hint.

Probably needs TensorVariable to be added to the allow-list in the first else branch.
Plus a raise on size.ndim > 1.

[5hv5hvnk]

I did some changes and the code executes without any error
in line 475 size.ndim>=1
in line 476 size=tiple(size,)

how should I proceed further (should I generate the pull request?)

[ricardoV94]

@5hv5hvnk I am not sure what you are proposing is correct but feel free to open a PR and we can have a look over there, with the actual code changes :)

[5hv5hvnk]

Oh okay is there a way I can run tests on it to check if its been fixed ?

[ricardoV94]

You can try pytest pymc/tests/test_distributions_random.py to check nothing obvious is broken. You should also add a mew test for this new logic, wherever the relevant function is being tested.

When you open a PR with us it will automatically run the entire test suite plus whatever new tests you added.

[5hv5hvnk]

I have opened PR #5417 please have a look at it and share if any changes I have to make in order to fix the issue.