Unify +simplify collators API among tasks

huggingface / pixparse

Pixel Parsing. A reproduction of OCR-free end-to-end document understanding models with open data

11 stars 3 forks source link

Unify +simplify collators API among tasks #27

Closed molbap closed 11 months ago

molbap commented 11 months ago

This WIP PR is a stab at #26 to move collators to a unified API followinf RVLCDIP

[x] Add BaseCollate
[x] Add CORD collate
[x] Add DocVQA collate
[x] move pack_inputs to base class
[x] run tests

molbap commented 11 months ago

Added a few checks and moved to flake + autopep8 with line length 120, feels more natural to read as well

rwightman commented 11 months ago

@molbap re the merging, I merged the smaller ones but left this since it still had tasks remaining

molbap commented 11 months ago

Added tests, solving conflicts with mega_refactor before merge

molbap commented 11 months ago

@rwightman should be good to go within scope of collators/basic tests, no guarantee on the forward method for this but will base necessary modifs for #28 accordingly