trOCR in onnx format does not read full text

[feff2]

System Info

platform: Windows 10
optimum version 1.19.2
transformers version 4.40.2
onnx version 1.16.0
onnxruntime version 1.17.3

Who can help?

@amyeroberts
@pacman100

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

I have trained the small-printed trocr on my custom dataset having multiline images. The trained model can read full text. But while converting the model to onnx, the model detects only first line or part of it in first line. I have used this [https://github.com/huggingface/transformers/issues/19811#issuecomment-1303072202](https://gist.github.com/mht-sharma/f38c670930ac7df413c07327e692ee39)
for inference and this command "optimum-cli export onnx -m {model_checkpoints} --task vision2seq-lm onnx/ --atol 1e-3" for convert to onnx

Expected behavior

It is unclear why the model recognizes only the first line of text (with almost no loss of quality)

[amyeroberts]

Hi @feff2, thanks for raising an issue!

I'm transferring this issue to the optimum repo, as it seems this is more related to that library.

[feff2]

@amyeroberts , thanks!