RFC: add isin for elementwise set inclusion test #854

Activity

1. asmeurer commented on Nov 22, 2024

   

   asmeurer

   on Nov 22, 2024

   Member

   More actions

   Another potential reason for adding it is that it uses a nontrivial implementation which depends on some heuristics based on the input size.
2. rgommers commented on Nov 22, 2024

   

   rgommers

   on Nov 22, 2024

   Member

   More actions

   Thanks @lucascolley. I've added ndonnx (which has it) and MLX (which doesn't) to the issue description.

   This seems like a very reasonable proposal to me. Implementing isin in terms of other primitives in the standard is a little complex indeed.

   The return type should always be a boolean array. The NumPy docs say it can be a bool for a single input element, but that's actually a bug in the docs. I checked NumPy, JAX, and PyTorch and all return a 0-D array.
3. rgommers

   added

   API extensionAdds new functions or objects to the API.Adds new functions or objects to the API.

   on Nov 22, 2024
4. kgryte

   added

   RFCRequest for comments. Feature requests and proposed changes.Request for comments. Feature requests and proposed changes.

   on Nov 22, 2024
5. kgryte

   added this to Proposalson Nov 22, 2024
6. kgryte

   moved this to Stage 0 in Proposalson Nov 22, 2024
7. rgommers commented on Nov 28, 2024

   

   rgommers

   on Nov 28, 2024

   Member

   More actions

   The thing to discuss here is what keywords are desired I think. NumPy and Dask use:

   def isin(element, test_elements, assume_unique=False, invert=False, *, kind=None)

   The private scikit-learn implementation here is:

   def isin(element, test_elements, xp, assume_unique=False, invert=False)

   JAX:

   def isin(element, test_elements, assume_unique=False, invert=False, *, method='auto')

   PyTorch:

   def isin(elements, test_elements, *, assume_unique=False, invert=False)

   ndonnx:

   def isin(x: Array, items: Sequence[Scalar]) -> Array

   The assume_unique and invert keywords seem easy to support and useful. So this should probably work:

   def isin(x: Array, test_elements: Array, /, *, assume_unique : bool = False, invert : bool =False) -> Array[bool]

   The type of test_elements is a bit TBD, could be a union between arrays and sequences perhaps?
8. rgommers commented on Nov 28, 2024

   

   rgommers

   on Nov 28, 2024

   Member

   More actions

   We discussed this in the community meeting today. A summary with a couple of points to follow up on:

   • In general, folks were 👍🏼 on adding isin
   • For the second argument, accepting arrays and scalars but not sequences seemed preferred.
   • Some discussion about promotion behavior, this needs to be specified. Since the semantics of isin are element-wise comparison like, it is probably a good idea to match what == does.
   • For the second argument: limit to 1-D, or accept any shape and then reshape to 1-D? NumPy does the latter.
9. kgryte

   added this to the v2025 milestone on May 29, 2025
10. kgryte

    added a commit that references this issue on Jun 12, 2025

    feat: add `isin` to the specification

    ...

    7c09df3
11. kgryte

    moved this from Stage 0 to Stage 2 in Proposalson Jun 12, 2025
12. kgryte commented on Jun 12, 2025

    

    kgryte

    on Jun 12, 2025

    Contributor

    More actions

    A PR adding isin to the specification is now up for review: #959
13. kgryte

    added

    AcceptedRFC feature request which has been accepted.RFC feature request which has been accepted.

    on Feb 5, 2026
14. kgryte

    closed this as completedin #959on Feb 19, 2026
15. kgryte

    added a commit that references this issue on Feb 19, 2026

    feat: add `isin` to the specification

    ...

    ffaf7fa