Finetuning on DONUT-proto about donut

[Veason-silverbullet]

Hi, @gwkrsrch @logan-markewich . Seems that when the input size is set as [1024, 768], donut-proto works on finetuning, observed from the DONUT-base config that pretrain[2560x1920] -> finetuning[1280x960]. Does it make sense?

[Veason-silverbullet]

@gwkrsrch , yes the above example notebook works well!

BTW, do you plan to release the donut-proto pretraining YAML config file? I would like to train donut-proto from scratch on other datasets, so the example pretraining config file would help (e.g., optimizer settings). Thanks a lot!

[logan-markewich]

+1!

Also, if you ever have results for docvqa with donut-proto, I'd love to seem them 🥇

[gwkrsrch]

Hi @Veason-silverbullet, yes it make sense :) The window size of donut-proto is 8, hence, there should be no problem if the size of each axis is set to a multiple of 256 (e.g., 768, 1024, 2048, etc).

[gwkrsrch]

@Veason-silverbullet I made a toy example notebook about training and testing donut-proto on CORD. You can see the colab notebook at this link. Hope this helps :)