This repo contains the modified source code of the original Visual Pivoting for (Unsupervised) Entity Alignment and various Jupyter Notebooks we used for the experimental setup.

• eva_gridsearch.ipynb: A grid search to discover the trade-offs in directly reducing the size of the relation, attribute, image, and graph embeddings. (Ben)
• eva_graphencoder.ipynb: Our experiments in using an autoencoder to reduce the size of the embeddings (Shreesh, Ben)
• eva_effnet.ipynb: Extracting the second-to-last layer activations of EfficientNet B0 and B7 to be used as image features in EVA. (Ben)
• eva_altimgfeatures.ipynb: An evaluation of the performance of alternative image features generated by EfficientNet B0 and B7 pre-trained on ImageNet. (Ben)
• splitpkl.py: A script to load the original image data (~20GiB pkl file) and split it into smaller pkl files that fit on low memory machines. (Ben)
• rclone_sync.sh: A utility script to sync this repository with google drive. (Ben)
• gridsearch.csv: The Hits@1 results of the grid search on embedding sizes.
• gridsearch.pkl: A dictionary containing the results of the grid search on embedding sizes. The dictionary is of the form:

{
    (rel, att, img, graph): (hits@1_FRtoEN, hits@1_ENtoFR, MR_FRtoEN, MR_ENtoFR, MRR_FRtoEN, MRR_ENtoFR)
}

where rel, att, img, graph are the embedding sizes for the relation, attribute, image, and graph embeddings respectively

This code is modified from EVA. We thank the authors for providing their source code with an open license.