Fix a numpy deprecation related to dtype

[Armavica]

Deprecated in numpy 1.21.0.
This fixes 369 deprecation warnings in PyMC's test suite.

See https://numpy.org/doc/stable/release/1.21.0-notes.html#the-dtype-attribute-must-return-a-dtype

Thank you for opening a PR!

Here are a few important guidelines and requirements to check before your PR can be merged:

• There is an informative high-level description of the changes.
• The description and/or commit message(s) references the relevant GitHub issue(s).
• pre-commit is installed and set up.
• The commit messages follow these guidelines.
• The commits correspond to relevant logical changes, and there are no commits that fix changes introduced by other commits in the same branch/BR.
• There are tests covering the changes introduced in the PR.

Don't worry, your PR doesn't need to be in perfect order to submit it. As development progresses and/or reviewers request changes, you can always rewrite the history of your feature/PR branches.

If your PR is an ongoing effort and you would like to involve us in the process, simply make it a draft PR.

[brandonwillard]

Thanks for the PR! I just fixed a related issue in 8867a72. Can you try the same numpy.obj2sctype approach from there? It appears to be a concise way to determine whether or not a string or type is a valid dtype.

[Armavica]

I am sorry, but I am not sure how this would work. The function obj2sctype doesn't seem to return either a dtype or an object whose .dtype is a dtype, as shown by

for a in ["float64", np.int8, np.int8(4).dtype]:
    assert not isinstance(np.obj2sctype(a), np.dtype)
    assert not isinstance(np.obj2sctype(a).dtype, np.dtype)

It is possible that I am missing something though, as this dtype manipulation is extremely confusing to me.

[brandonwillard]

I am sorry, but I am not sure how this would work. The function obj2sctype doesn't seem to return either a dtype or an object whose .dtype is a dtype, as shown by

for a in ["float64", np.int8, np.int8(4).dtype]:
    assert not isinstance(np.obj2sctype(a), np.dtype)
    assert not isinstance(np.obj2sctype(a).dtype, np.dtype)

It is possible that I am missing something though, as this dtype manipulation is extremely confusing to me.

Likewise, I'm probably not thinking of the same issue that you are.

Are the deprecation warnings caused by something like the following?

import numpy as np
import aesara.tensor as at


np.dtype(at.vector("x"))
# <ipython-input-11-4480ef785bc5>:1: DeprecationWarning: in the future the `.dtype` attribute of a given datatype object must be a valid dtype instance. `data_type.dtype` may need to be coerced using `np.dtype(data_type.dtype)`. (Deprecated NumPy 1.20)
#   np.dtype(at.vector("x"))

[Armavica]

Yes, I think that is the idea, at least for some of the warnings.

[brandonwillard]

Yes, I think that is the idea, at least for some of the warnings.

I think I found a test that provides an example of the warnings you're talking about: pymc/tests/test_logprob.py::test_joint_logp_basic.

In that case, pymc.distributions.logprob.joint_logp is passing a TensorType—via var[0].type—to the TensorVariable.astype method here, when it should be using var[0].dtype.

[Armavica]

Amazing, it looks like this fixes everything! Thank you very much!