Implement pad

[jessegrabowski]

Description

Implement pt.pad, following the np.pad API with feature parity.

Very preliminary draft, uploading it in this state so I can ask @ricardoV94 to look at the _linear_ramp_pad function and tell me if I'm missing something obvious related to shapes. It should follow numpy.lib.arraypad._get_linear_ramps. Also the reflection pad uses a scan, curious if we can avoid that somehow or if we think it will be no big deal (probably the 2nd).

Also I'm not sure where to put this. I put it in tensor/basic but it might be better in tensor/extra_ops?

Related Issue

• Closes ENH: Implement pt.pad #743
• Related to Restore fast convolution Ops, rewrites, and docs #548

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

[ricardoV94]

What about a padding.py file?

[jessegrabowski]

Sure, I'll make a new file. It's just not my default. I agree it doesn't belong in basic.

[jessegrabowski]

Not quite 1:1 on numpy features but close. The more exotic padding schemes I would need more time to understand.

Still needs jax/numba overloads, but these should be very trivial.

[codecov]

Codecov Report

Attention: Patch coverage is 94.30380% with 18 lines in your changes missing coverage. Please review.

Project coverage is 81.48%. Comparing base (c6d85d1) to head (bbeb300).
Report is 97 commits behind head on main.

Files with missing lines	Patch %	Lines
pytensor/tensor/extra_ops.py	80.95%	8 Missing and 4 partials ⚠️
pytensor/tensor/pad.py	97.14%	3 Missing and 3 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #748      +/-   ##
==========================================
+ Coverage   81.38%   81.48%   +0.09%     
==========================================
  Files         172      174       +2     
  Lines       46868    47166     +298     
  Branches    11423    11471      +48     
==========================================
+ Hits        38145    38434     +289     
- Misses       6542     6548       +6     
- Partials     2181     2184       +3     

Files with missing lines	Coverage Δ
pytensor/link/jax/dispatch/__init__.py	100.00% <100.00%> (ø)
pytensor/link/jax/dispatch/pad.py	100.00% <100.00%> (ø)
pytensor/tensor/subtensor.py	89.31% <100.00%> (+0.11%)	⬆️
pytensor/tensor/pad.py	97.14% <97.14%> (ø)
pytensor/tensor/extra_ops.py	87.64% <80.95%> (-0.98%)	⬇️

... and 4 files with indirect coverage changes

[ricardoV94]

Looks great so far, left some small suggestions.

[jessegrabowski]

Draft of the JAX overload. Need your input on the Pad OpFromGraph. I needed some way to hang on to the keyword arguments.

It seems like there might be a difference between how JAX and numpy handle mode=mean padding, because tests pass against numpy but not against JAX. I'll investigate more carefully, but it might be a JAX bug (I doubt this padding mode is used ever)

I also think my loopy pads (symmetric, wrap) need to be redone, because they are failing a new test that arbitrarily pads every dimension of an nd input differently. So all that probably needs a re-design from the ground up.

[ricardoV94]

I also think my loopy pads (symmetric, wrap) need to be redone, because they are failing a new test that arbitrarily pads every dimension of an nd input differently. So all that probably needs a re-design from the ground up.

You may need an operation per dimension

[ricardoV94]

Regarding JAX do you need to implement a specific dispatch? For instance for the einsum I don't think we'll need because the OFG expression will be as good as what they do internally (since we copied it from them)

[jessegrabowski]

No idea on the JAX dispatch. I just assumed I should.

[jessegrabowski]

Don't understand why the doctest for pad is failing

[ricardoV94]

Don't understand why the doctest for pad is failing

It says there is an output that was not expected. If you have a print somewhere, you need to always test it afterward

[ricardoV94]

You can also run doctest locally btw

[ricardoV94]

Something like pytest --doctest-modules pytensor/tensor/pad.py --verbose

[ricardoV94]

We should open a follow up issue for performance. With the reshape and concatenation, we're doing a lot of copies. We should see how much better it would be to have scans with set_subtensors like you tried halfway.

[jessegrabowski]

I kind of just want to skip the segfault test and come back to it later. I am trying to debug, but not really sure what's going on. It runs fine when NUMBA_DISABLE_NJIT flag is set. My suspicion is an out-of-range index, but I can't reproduce the error in an NB. I'd be fine just removing the numba tests all-together and considering it unsupported, but every other mode passes. So idk. I just want to get this over the finish line.

[ricardoV94]

This is huge!