Table-Recognition-base-on-Transformer-Decoder

An end to end model for two sub-tasks of Table Recognition: table structure recognition, cell detection

Dataset: Pubtabnet

Architecture: base on this paper

• Consists one of Shared Encoder, one Shared Decoder and three separate Decoder for three sub-tasks
  • Shared Encoder using a CNN backbone network as the feature extractor
  • Four Decoders are inspired by original Transformer decoder

Proposal Architecture:

• Add FPN and Encoder Transformer

• Add Multi-Aspect Global Context Attention

Description:

• config.py contains hyperparameters
• parsing_data.py match raw data from Pubtabnet to anotation
• tokenizer.py encode characters, html tags
• sub_module.py build necessary sub-modules like Cross Attention, Self Attention, Positional Encoding, ...
• main_model build last model from sub-modules
• train_infer.py train loop