Adds new broadcast_to atom for nd broadcasting #2724

Transurgeon · 2025-02-21T03:19:55Z

Description

Please include a short summary of the change.
This PR introduces the broadcast_to atom in hopes of fixing issues with broadcasting in higher dimensions.
The broadcast_to atom is used internally inside of the expression class in its broadcast method.
This method is called in many places, including binary ops, basic ops and others.

The implementation of the atom follows other similar shape manipulation atoms such as vstack and concatenate. The idea is to form a set of indices which are reshaped to the expression's shape, then using numpy's API we are able to mimic the behavior of np.broadcast_to on cvxpy's tensor views in the backend.

CPP backend compatibility

In the #2728 discussion, I brought up some issues with the difference in behavior for the backends. Those issues were resolved as follows:

only call cp.broadcast_to when one of the operands has dim > 2, which ensures that the SCIPY backend is used.
full broadcasting is supported in those cases, even the edge cases mentioned in the discussion above.
in hopes of keeping the CPP backend behavior the same, both the SCIPY and CPP backends don't do full broadcasting when both operands are <= 2. However, this only omits some edge cases, as most practical scenarios are covered by existing code.

Improved error messages

Recent changes to the codebase (notably adding the cp.concatenate atom) showed that we are able to instantiate empty numpy arrays (using dtype=np.empty([])) and then perform analogous operations on those arrays to get the error handling from numpy for free. This technique has been reused here to improve on the existing multiply atom error handling.

Issue link (if applicable): #2739

Type of change

New feature (backwards compatible)
New feature (breaking API changes)
Bug fix
Other (Documentation, CI, ...)

Contribution checklist

Add our license to new files.
Check that your code adheres to our coding style.
Write unittests.
Run the unittests and check that they’re passing.
Run the benchmarks to make sure your change doesn’t introduce a regression.

CLAassistant · 2025-02-21T03:20:02Z

All committers have signed the CLA.

Transurgeon

I think this is ready for a first round of reviews. I have still yet to add backend specific tests which could help understand the operator (although it is mostly about duplicating existing tensors along the rows).
Things that are open for discussion, do we want to support broadcasting for both operands? (i.e. cp.multiply((3,1), (5)) is allowed and will give an (3,5) expression)
this means we will need to broadcast both the constant data and the variable/expr which differs from the current behavior (although matches the NumPy API for broadcasting)

cvxpy/tests/test_expressions.py

cvxpy/expressions/expression.py

cvxpy/atoms/affine/binary_operators.py

SteveDiamond · 2025-03-07T18:00:17Z

cvxpy/tests/test_expressions.py

+    @pytest.mark.parametrize("shapes", [((5),(3, 1)),
+                                        ((15, 1),(8)),
+                                        ((3), (2, 1))])
+    def test_segfault_multiply(self, shapes) -> None:


Can you say more about this test? Would these shapes be compatible with numpy broadcasting rules?

yes they are compatible with numpy broadcasting. And these tests would pass for the SCIPY backend. However, I was not able to enforce the SCIPY backend to be used in these cases. (CPP would raise errors because it doesn't have the broadcast_to atom implemented).

SteveDiamond · 2025-03-07T18:09:53Z

This is a really elegant solution for the backend! To be clear, in this MR we are not handling cases such as (5) + (5,1) -> (5,5)? We should still handle these cases but it can be in a separate MR.

Transurgeon · 2025-03-07T19:08:12Z

cvxpy/atoms/affine/broadcast_to.py

+    def _supports_cpp(self) -> bool:
+        return False


I've added this method, but I was still getting tests going through the CPP backend.
I think it's because the broadcast_to atom is being called from another atom (i.e. multiply), but not sure.
Ideally we should fix this, but maybe for another PR.

SteveDiamond · 2025-03-07T21:15:09Z

There is a bug in _aggregate_metrics. It's not recursive. It should be

    @pu.compute_once
    def _aggregate_metrics(self) -> dict:
        """
        Aggregates and caches metrics for expression trees in self.args. So far
        metrics include the maximum dimensionality ('max_ndim') and whether
        all sub-expressions support C++ ('all_support_cpp').

        """
        max_ndim = self.ndim
        cpp_support = self._supports_cpp()

        for arg in self.args:
            max_ndim = max(max_ndim, arg._max_ndim())
            cpp_support = cpp_support and arg._all_support_cpp()

        metrics = {
            "max_ndim": max_ndim,
            "all_support_cpp": cpp_support
        }
        return metrics

cvxpy/expressions/expression.py

* adding some initial ideas to fix broadcasting * reverting changes and adding new tests/broadcast atoms * add more boilerplate for broadcasting operator * updating methods for the broadcast_to atom * some improvements * add more tests for broadcasting, raise exception * update shape from args and validate args * fixing some tests, but still issues with CPP backend not implemented * adding test from eric's code snippet * fixing all tests and adding new ones for multiply broadcast * revert these changes and remove failing case as separate test * reverting changes to typing * some updates and progress towards a truly working backend * truly reverting np.typing issue now * fixing implementation for broadcast_to atom, and cleaning up tests * final changes for broadcasting to be fully working in nd * adding hypothesis tests and renaming some other stuff * fix bug * maybe fixed all the tests? * remove breakpoint * final touches, changing test name and case in expr.broadcast * revert changes to expr.broadcast --------- Co-authored-by: William Zijie Zhang <[email protected]> Co-authored-by: Steven Diamond <[email protected]>

* patch empty breaking empty rownames for xpress>=9.5 (#2745) Co-authored-by: Marc Bataillou Almagro <[email protected]> * Adds new broadcast_to atom for nd broadcasting (#2724) * adding some initial ideas to fix broadcasting * reverting changes and adding new tests/broadcast atoms * add more boilerplate for broadcasting operator * updating methods for the broadcast_to atom * some improvements * add more tests for broadcasting, raise exception * update shape from args and validate args * fixing some tests, but still issues with CPP backend not implemented * adding test from eric's code snippet * fixing all tests and adding new ones for multiply broadcast * revert these changes and remove failing case as separate test * reverting changes to typing * some updates and progress towards a truly working backend * truly reverting np.typing issue now * fixing implementation for broadcast_to atom, and cleaning up tests * final changes for broadcasting to be fully working in nd * adding hypothesis tests and renaming some other stuff * fix bug * maybe fixed all the tests? * remove breakpoint * final touches, changing test name and case in expr.broadcast * revert changes to expr.broadcast --------- Co-authored-by: William Zijie Zhang <[email protected]> Co-authored-by: Steven Diamond <[email protected]> * change is_released to true --------- Co-authored-by: Marc Bataillou Almagro <[email protected]> Co-authored-by: Marc Bataillou Almagro <[email protected]> Co-authored-by: William Zijie Zhang <[email protected]> Co-authored-by: Steven Diamond <[email protected]>

Transurgeon and others added 3 commits February 5, 2025 04:39

adding some initial ideas to fix broadcasting

86b4c7f

reverting changes and adding new tests/broadcast atoms

2218795

add more boilerplate for broadcasting operator

fe79f2b

Transurgeon added 8 commits February 20, 2025 23:13

updating methods for the broadcast_to atom

36b24ca

some improvements

2c748ff

add more tests for broadcasting, raise exception

f8523b8

update shape from args and validate args

0d101b4

fixing some tests, but still issues with CPP backend not implemented

d6e7aa9

adding test from eric's code snippet

d3a559b

fixing all tests and adding new ones for multiply broadcast

097424d

revert these changes and remove failing case as separate test

86ba518

Transurgeon marked this pull request as ready for review February 22, 2025 07:20

Transurgeon changed the title ~~BUGFIX: Adds new broadcast_to atom for nd broadcasting~~ Adds new broadcast_to atom for nd broadcasting Feb 22, 2025

Transurgeon commented Feb 22, 2025

View reviewed changes

cvxpy/tests/test_expressions.py Outdated Show resolved Hide resolved

cvxpy/expressions/expression.py Show resolved Hide resolved

cvxpy/atoms/affine/binary_operators.py Show resolved Hide resolved

Transurgeon added 6 commits March 6, 2025 00:55

reverting changes to typing

5ce490a

some updates and progress towards a truly working backend

a73b0b9

truly reverting np.typing issue now

9805737

fixing implementation for broadcast_to atom, and cleaning up tests

5a1740b

final changes for broadcasting to be fully working in nd

08013a1

adding hypothesis tests and renaming some other stuff

fa85446

SteveDiamond reviewed Mar 7, 2025

View reviewed changes

SteveDiamond approved these changes Mar 7, 2025

View reviewed changes

Transurgeon added the PR backport needed label Mar 7, 2025

Transurgeon commented Mar 7, 2025

View reviewed changes

SteveDiamond added 3 commits March 8, 2025 17:10

fix bug

96b4b27

maybe fixed all the tests?

32e9e37

remove breakpoint

37c4b4a

final touches, changing test name and case in expr.broadcast

e15fe33

SteveDiamond reviewed Mar 9, 2025

View reviewed changes

cvxpy/expressions/expression.py Outdated Show resolved Hide resolved

revert changes to expr.broadcast

4e56150

Transurgeon merged commit ba14f1a into cvxpy:master Mar 9, 2025
51 of 52 checks passed

langestefan mentioned this pull request Mar 20, 2025

BROADCAST_TO is not supported #2764

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Adds new broadcast_to atom for nd broadcasting #2724

Adds new broadcast_to atom for nd broadcasting #2724

Transurgeon commented Feb 21, 2025 •

edited

Loading

Uh oh!

CLAassistant commented Feb 21, 2025 •

edited

Loading

Uh oh!

Transurgeon left a comment •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

SteveDiamond Mar 7, 2025

Uh oh!

Transurgeon Mar 7, 2025

Uh oh!

SteveDiamond commented Mar 7, 2025

Uh oh!

Transurgeon Mar 7, 2025

Uh oh!

SteveDiamond commented Mar 7, 2025

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Adds new broadcast_to atom for nd broadcasting #2724

Adds new broadcast_to atom for nd broadcasting #2724

Conversation

Transurgeon commented Feb 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

CPP backend compatibility

Improved error messages

Type of change

Contribution checklist

Uh oh!

CLAassistant commented Feb 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Transurgeon left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

SteveDiamond Mar 7, 2025

Choose a reason for hiding this comment

Uh oh!

Transurgeon Mar 7, 2025

Choose a reason for hiding this comment

Uh oh!

SteveDiamond commented Mar 7, 2025

Uh oh!

Transurgeon Mar 7, 2025

Choose a reason for hiding this comment

Uh oh!

SteveDiamond commented Mar 7, 2025

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Transurgeon commented Feb 21, 2025 •

edited

Loading

CLAassistant commented Feb 21, 2025 •

edited

Loading

Transurgeon left a comment •

edited

Loading