This repository was used for the experiments described in SacreEOS.
- python >= 3.7
- numpy
- torch
- h5py
- sacreeos
The MS-COCO 2014 visual inputs are generously provided by
here, where features are extracted by a Faster-RCNN
using the ResNet-101 backbone. The data can be converted into the required format of
this repository using our features_generation.py
script by specifying the source file location:
python features_generation.py --tsv_features_path trainval_file_path
Download dataset_coco.json
here
and place it in the directory github_ignore_material/raw_data/
.
XE Training:
python train.py $(cat confs/xe.conf) &> xe_output.txt &
SCST with eos training:
python train.py $(cat confs/rf_standard.conf) &> rl_std_output.txt &
SCST without eos training:
python train.py $(cat confs/rf_no_eos.conf) &> rl_noeos_output.txt &
If you find this repository useful, please consider citing (no obligation):
@article{hu2023request,
title={A request for clarity over the End of Sequence token in the Self-Critical Sequence Training},
author={Hu, Jia Cheng and Cavicchioli, Roberto and Capotondi, Alessandro},
journal={arXiv preprint arXiv:2305.12254},
year={2023}
}