Code for the NeurIPS 2021 paper Independent Prototype Propagation for Zero-Shot Compositionality.
To train and evaluate the model change the parameters in run_args/args.json and run train/train_protoprop.py, or change the parameters in train/grid_search.py and run for a grid search or results with error bars.
The data can be obtained from the repositories in the attribution section below.
NOTE: ImageNet transforms (as used in dataloaders/ut_zappos.py) crop out significant portions of the objects depicted in UT-Zappos and C-GQA images.
We use them for fair comparison with previous works, but we would recommend using different val and test crops otherwise.
Parts of the code has been taken or adapted from the following sources.
Evaluation & loading UT-Zappos (these repositories also contain dowload links for the data):
https://github.com/Tushar-N/attributes-as-operators
https://github.com/ExplainableML/czsl
HSIC calculation:
https://github.com/choasma/HSIC-bottleneck
@inproceedings{ruis2021protoprop,
title={Independent Prototype Propagation for Zero-Shot Compositionality},
author={Ruis, Frank and Burghouts, Gertjan J and Bucur, Doina},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
volume={34},
year={2021}
}