Skip to content

Clarify usage / implementation of padding for SigLIP model processor #28569

Description

@skysyk

Feature request

Change the implementation of SiglipProcessor to use the global default behavior for padding of False or update the documentation to indicate the usage is different and defaults to 'max_length' if the padding argument is not provided.

Motivation

In the HF documentation for padding (both the docs as well as the function comments for the processor class), the default behavior (argument) is described to be False or 'do_not_pad'. For the SiglipProcessor, max_length is the default behavior implemented in code while the example in the docs omits the padding argument. This is at odds with the overall documentation as well as behavior / usage examples provided in similar models such as CLIP (where explicitly padding=True in the usage example) and could give the wrong impression upon first glance that padding is not used for SigLIP.

Your contribution

Opening this issue to discuss a clarification / improvement. I can help implement the preferred solution.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions