This repository contains data and code for the EMNLP 2017 paper Zero-Shot Activity Recognition with Verb Attribute Induction. (For more information, see the paper). If you use our Verbs with Attributes corpus or if the paper significantly inspires you, we request that you cite our work:
@inproceedings{emnlp17_zellers,
author = {Rowan Zellers and Yejin Choi},
title = {Zero-Shot Activity Recognition with Verb Attribute Induction},
url = {https://arxiv.org/abs/1707.09468},
booktitle = "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
year = "2017",
}
Our annotations are available in this folder. See the readme there for more information.
To download the imSitu images, you'll need to follow the instructions from the install script here.
I originally wrote this code with PyTorch 1.12, but I've updated it to hopefully work for PyTorch 3.0. See requirements.txt for dependencies. For ease of use, I recommend installing everything in a virtualenv. Ping me if there's a dependency missing (I tried to prune dependencies not needed for this project in particular).
- You'll need to download the imsitu dataset from here and update
IMSITU_PATH
in config.py accordingly. - If you want to pretrain on the dictionary challenge dataset, then download it from here. I renamed the file
training_data.pkl
(in the archivetraining_data.tgz
) todictionary_challenge.pkl
and moved it to my data folder.
Here's the rough outline to reproduce my results:
# Pretrain dictionary challenge
python models/def_to_atts_pretrain.py
# pretrain IMSITU
python models/imsitu_pretrain.py
# Train the ensembling text model
python models/def_to_atts_train.py
# Train imsitu
python models/imsitu_train.py
For evaluation, use the scripts def_to_atts_eval.py
and imsitu_eval.py
This documentation is a work in progress, so flag an issue or contact me if you have any questions.