Skip to content

opSupportLimits: Level of detail for output tensor(s)? #835

@inexorabletash

Description

@inexorabletash

Per discussion in #828 (comment) there are a range variety of opinions about how much detail opSupportLimits() should include for the output tensors. Currently as specified and per the Chromium implementation, output tensor(s) are included and list just the data types, not ranks (#828).

  • On one extreme, the output data type and rank can (almost) always be inferred from the inputs. So providing the bare minimum of data in the dict would mean that output tensor(s) are not described at all - neither data type(s) nor rank(s).
  • On the other, it is informative to list what possible outputs are, and this can help with debugging. So providing both possible output data types and ranks would be useful.

Note that for ops where the output data type can be explicitly passed - argMin/argMax and cast - opSupportLimits() must include the supported types.

Metadata

Metadata

Assignees

Labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions