First of all, thank you for providing useful code.

I am using LayoutLM and I have 3questions.

1. Image Embedding
   According to your "LayoutLM: Pre-training of Text and Layout for Document Image Understanding" document , it makes image embedding token by Faster R-CNN but I can't find any code related with Faster R-CNN in LayoutLM code.
   Then it means when I using LayoutLM(for Form understanding task) I only use 2-D position Embedding and not using Image embeddings. So, where can I find code related with Image embeddings.

2. Seqeunce-Linking
   Your current code is related with sequence labeling and not related with sequence linking. I think sequence linking is important when extracting useful information from document or papers. So, how can I get sequence linking(key-value pair) results.
   Also you mentioned before issues; just check Funsd dataset paper and add binary classifier layer at the end of the output embeddings. But, I can't understand it and how to make it as a code. So can you let me know any certain code related about it?

3. Input-Output
   LayoutLM - seq_labeling is aiming about making right BIOES tag to each words. But when I check the code input, Input also contains BIOES tag. Then I think there is no meaning because our goal is to make BIOES tag but actually input already contains BIOES tag. So, Can you explain why the input has BIOES-tag?

I'm sorry about that there was too many questions but I hope you reply to all theses questions. Thank you!!