Skip to content

Latest commit

 

History

History
80 lines (69 loc) · 2.96 KB

README.md

File metadata and controls

80 lines (69 loc) · 2.96 KB
Raw

README

This directory contains the files for declaration generation. T5-small is exploited as the encoder-decoder model for training and evaluation.

INSTALL

Data

Assume the root data dir is [ROOT_DATA_DIR], then the declaration dataset for training and validation is placed in [ROOT_DATA_DIR]/declaration/*:

|- [ROOT_DATA_DIR]
    |- declaration/
        |- question_to_declarative_train.json
        |- question_to_declarative_val.json

Model Training

Run the script for training:

bash run_train.sh

or the fine-tuned T5-small model (model/declaration/checkpoint-480000) can be downloaded from Baidu Yun (PSW:8888). The fine-tuned or downloaded checkpoint is placed in data/model/declaration/checkpoint-480000.

Declaration Generation

Once T5 is trained on the declaration dataset, the model can be used to generate declarative sentences for GQA and VQA datasets. Just follow the steps as bellow:

  1. Transform the questions (from GQA and VQA v2.0 datasets) into the translation format (one sample per line), where en_q denotes the source question string and en_a denotes the target declarative sentence we want to generate. The file is named source_file.txt: 
    {"translation": {"en_q": "Is the sky dark?", "en_a": ""}}
{"translation": {"en_q": "What is on the white wall?", "en_a": ""}}
{"translation": {"en_q": "Is that pipe red?", "en_a": ""}}
...
  2. Assume the path of source_file.txt is [SOURCE_FILE_DIR]/source_file.txt, then run the script: 
    bash run_predict.sh
    Finally, there will be a .txt file in output dir, i.e., generated_predictions.txt. This file contains one sentence per line, representing the declarative sentence of the corresponding question in source_file.txt. The format of generated_predictions.txt is shown as follows: 
    [MASK], the sky [BE] dark.
the [MASK] is on the wall.
[MASK], that pipe [BE] red.
  3. We provide the pre-generated declaration files (from the questions of GQA and VQA v2.0 datasets) for easy-to-use. The files can be downloaded from Baidu Yun (PSW:8888), the files are arranged as follows: 
    |- [ROOT_DATA_DIR]
    |- declaration/
        |- gqa/
            |- gqa_all_submission_declaration.json
            |- gqa_all_train_declaration.json
            |- gqa_all_val_declaration.json
            |- gqa_bal_train_declaration.json
            |- gqa_bal_val_declaration.json
        |- vqa/
            |- test2015_declarative.json
            |- test-dev2015_declarative.json
            |- train2014_declarative.json
            |- val2014_declarative.json
            |- vqa_vg_declarative.json