Implement faster-trocr code for trocr-small-handwritten/printed/stage1

[vu0607]

Dear @IlyasMoutawwakil, I have experienced optimize trocr-small-handwritten. I have run your code with my checkpoint trocr-small-handwritten in Vietnamese language.
It didn't work well, It can't optimize time inference and predicted label always begin with "to, the, and". Ex: the chào Việt Nam
Maybe your code compatible with trocr-base or may I wrong somewhere.
Hope you can help me with this issue.
Thanks for your attention and your help <3

[IlyasMoutawwakil]

sorry but this experiment is not maintained. the trick I used is already implemented in the transformers library.
if you want to optimize your TrOCR, I would suggest using optimum.

[vu0607]

@IlyasMoutawwakil Please, can you give me link reference optimize TrOCR in optimum, thanks for you help <3