code for the paper: AVSD Huda Alamri, Vincent Cartillier, Abhishek Das, Jue Wang, Stefan Lee, Peter Anderson, Irfan Essa, Devi Parikh, Dhruv Batra, Anoop Cherian, Tim K. Marks, Chiori Hori
website: video-dialog.com
This code has been developed upon batra-mlp-lab/visdial-challenge-starter-pytorch
# create and activate environment
conda env create -n avsd -f=env.yml
conda activate avsd
- download 'split'.json data at: video-dialog.com
-
Build dialogs json file with otions using
makejson_with_options.py
(output: 'split'_options.json) -
Adapt JSON format using
convert_json_to_visdial_style.py
(output: 'split'_options_2.json can be renamed after to 'split'_options.json) -
Build tokenized captions, dialogs and image paths with
prepro.py
(output: dialogs.h5 and params.json) -
Build the image features (if working with images) using
prepro_img_vgg16.lua
orprepro_img_resnet.lua
from the batra-mlp-lab/visdial-challenge-starter-pytorch (output: data_img.h5) -
Build video features I3D (output: data_video.h5) https://github.com/piergiaj/pytorch-i3d.git
-
Build audio features AENET (output: data_audio.h5) https://github.com/znaoya/aenet.git
-
Training: python train.py
-
evaluation: python evaluate.py --use_gt
@article{DBLP:journals/corr/abs-1901-09107,
author = {Huda Alamri and
Vincent Cartillier and
Abhishek Das and
Jue Wang and
Stefan Lee and
Peter Anderson and
Irfan Essa and
Devi Parikh and
Dhruv Batra and
Anoop Cherian and
Tim K. Marks and
Chiori Hori},
title = {Audio-Visual Scene-Aware Dialog},
journal = {CoRR},
volume = {abs/1901.09107},
year = {2019},
url = {http://arxiv.org/abs/1901.09107},
archivePrefix = {arXiv},
eprint = {1901.09107},
timestamp = {Sat, 02 Feb 2019 16:56:00 +0100},
biburl = {https://dblp.org/rec/bib/journals/corr/abs-1901-09107},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
BSD