Adds interoperability with PyTensor (and PyMC)

[cako]

Description

Adds the ability to call use PyLops Linear Operators as PyTensor Ops. This allows PyMC to use PyLops LinearOperators inside of their models.

The PR also includes an example of Bayesian linear regression (examples/plot_bayeslinearregr.py) demonstrating this capability.

Changes

In addition to the main functionality, this PR issues several fixes:

• Ignore __pycache__
• Disables html_theme_options not available for current theme
• Various formatting issues in RST, math, etc.

Notes

Although I added everything to intersphinx, the documentation links is not appearing in my local documentation. Don't know if this is by design or if I need to change anything.

[cako]

@mrava87 I've been having a very hard time satisfying the version requirements of all libraries. PyTensor does not support numpy 2 (pymc-devs/pytensor#689). Devito only supports numpy 2 for the latest versions, so dropping to numpy 1.x forces one to use a very old version of devito... which in turns makes it not support newer versions of sympy (above 0.13). On the other hand, PyTorch requires versions of sympy>=0.13.1.

Errata: looks like devito does support numpy 1.x for the newest versions. I'll give this another try!

[cako]

@mrava87 I think I managed to build it, the trick was to prevent the pip command which install pytorch to overwrite the sympy installation. I did this setting pytorch version to be below 2.5.

Note that these version changes are only required for the requirements-dev.txt and requirements-torch.txt files because they are used separately when installing. They are not needed in the requirements-doc.txt or the yml files because the dependency managers have access to all packages at once and therefore are able to figure out that you need numpy<2 and pytorch<2.5 to be compatible with pytensor.

[cako]

@mrava87 managed to get everything working after a lot of back and forth with versions :)

https://pylops--621.org.readthedocs.build/en/621/gallery/plot_bayeslinearregr.html#sphx-glr-gallery-plot-bayeslinearregr-py

[mrava87]

Great! I’ll take a look in the next couple of days 😀

[mrava87]

This looks great 🚀

I just left a few minor comments... I am happy for this to be merged as soon as you feel it's ready 😄

[cako]

@mrava87 I actually tested the yml file as it current stands, it didnt install the correct numpy. The only way I could get it to work was to version pin:

- numpy>=1.21.0,<2
- scipy>=1.11.0,<1.13
- pytorch>=1.2.0,<2.5
 - pip:
  - devito==4.8.6

I'm reluctant to add this to the "standard" environment-dev.yml since that would force everyone to use old versions just so that pytensor still works. I think instead I will remove pytensor from the yml files, and put a check to not execute the pytensor tests if numpy>=2. What do you think?

[mrava87]

@mrava87 I actually tested the yml file as it current stands, it didnt install the correct numpy. The only way I could get it to work was to version pin:

- numpy>=1.21.0,<2
- scipy>=1.11.0,<1.13
- pytorch>=1.2.0,<2.5
 - pip:
  - devito==4.8.6

I'm reluctant to add this to the "standard" environment-dev.yml since that would force everyone to use old versions just so that pytensor still works. I think instead I will remove pytensor from the yml files, and put a check to not execute the pytensor tests if numpy>=2. What do you think?

I agree. In general I would strive for avoiding forcing upper limits on versions and doing that only as a temporary measure... especially when it comes to numpy, which is arguably the most used python package so every serious library will have to ultimately use numpy>2.

For the requirements.txt files, I am happy with the changes to allow tests/doc to be build in the CI... if we face issues in future we will decide what to do with respect to pytensor, but hopefully the PR you pointed to will be ready soon and the problem fixed 🚀

Feel free to merge after you add the check to the pytensor tests