feat: add broadcast_shapes to the specification
#983
Conversation
kgryte
added
API extension
| - If not provided one or more arguments, the function **must** return an empty tuple. | ||
| .. note:: | ||
| Array libraries which build computation graphs (e.g., ndonnx and Dask) commonly support shapes having dimensions of unknown size. If a shape contains a value other than an integer (e.g., ``None`` for a dimension of unknown size), behavior is unspecified and thus implementation-defined. Array-conforming libraries **may** choose to propagate such values (e.g., if a shape contains a dimension size of ``None``, the returned broadcasted shape also has a corresponding dimension having a size equal to ``None``) or raise an exception. |
There was a problem hiding this comment.
This can be phrased such that it works for computation graphs in the expected manner I believe, since the broadcasting math is abstract anyway and any sentinels can propagate correctly. To be discussed in the next call I think.
There was a problem hiding this comment.
I've updated the copy to now be normative with regard to sentinel propagation.
There was a problem hiding this comment.
There is the question regarding what happens when two shapes use different sentinels, but we can likely resolve this ambiguity in a future PR when someone asks for explicit clarification.
|
the array-api-extra docs just say |
ev-br
left a comment
ev-br
left a comment
There was a problem hiding this comment.
One small suggestion about the unknown dimensions, LGTM otherwise.
| ----- | ||
| - If not provided one or more arguments, the function **must** return an empty tuple. | ||
| - Array libraries which build computation graphs (e.g., ndonnx and Dask) commonly support shapes having dimensions of unknown size. If a shape contains a value other than an integer (e.g., ``None`` for a dimension of unknown size), array-conforming libraries **must** propagate such values (e.g., if a shape contains a dimension size of ``None``, the returned broadcasted shape **must** also have a corresponding dimension having a size equal to ``None``). |
There was a problem hiding this comment.
I've to admit I'm not exactly sure what this clause means for a library that does not support a sentinel value. E.g., what must numpy do for inputs of (1, 2, None) and (1, 2, 1)?
- does it raise a ValueError since the
broadcastingsection does not mentionNone(hence shapes are not broadcast-compatible), per theRaisessection? - or does it propagate the
None, per this clause?
Maybe a way out is to phrase it, roughly, "An array library may accept a value other than integer (e.g. None) for one or multiple dimensions. This is common for libraries that build computational graphs, and is used to indicate a dimension of unknown size. Array-conforming libraries which accept non-integer values for unknown dimensions must propagate such values...."
There was a problem hiding this comment.
numpy will raise ValueError. Your proposed phrasing looks good to me.
4 participants
This PR
broadcast_shapesto the specification #893 by adding support forbroadcast_shapesto the specification.Nonefor a dimension of unknown size, behavior is left unspecified and thus implementation-defined.