Re-export OCR to use single weights

[chmjkb]

Problem description

Currently our OCR pipeline needs using 3 different input shapes, so there is 3 different models exported, each with a different input shape, but with the same weights. This creates unnecessary overhead, since the users need to download the same weights 3 times.

Proposed solution

see https://docs.pytorch.org/executorch/stable/using-executorch-export.html#multi-method-models

Alternative solutions

No response

Benefits to React Native ExecuTorch

Less overhead for the users

Additional context

No response

[msluszniak]

Don't dynamic shapes work here still?

[chmjkb]

Don't dynamic shapes work here still?

@msluszniak If you're suggesting to export the model with dynamic shapes then unfortunately the export fails when we do so, only static shape export passes AFAIK

[msluszniak]

Yeap, that's what I meant

[msluszniak]

Maybe we should share the export errors with dynamic shapes with executorch team then?