S2TD

code for "S2TD: A Tree-Structured Decoder for Image Paragraph Captioning" accepted by MMAsia 2021

Overview

Notes

The code is based on soloist97/region-hierarchical-pytorch. For convenient, we recommend you to read its README.md first and follow the instructions to setup the environment.

• The code uses the same image features & preprocessed paragraph input files.
• The code shares almost the same training & evaluating commends. (If you want to generate sentence trees during evaluating, please add --tree.)
• Extra package requirements:
  • caesar0301/treelib
  • UKPLab/sentence-transformers
• Download pre-parsed tree files here.