SLOSH

Set Locality Sensitive Hashing via Sliced Wasserstein Embedding

Datasets:

• Point Cloud MNIST 2d: put downloaded data in /dataset/pointcloud_mnist_2d
• ModelNet40: put downloaded data in /dataset/modelnet
• Oxford 5K: put the extracted 8-dimensional features train_test_AE8.pkl in /dataset/oxford/

Baselines:

• WE: Wasserstein Embedding
• FSPool: Featurewise Sort Pooling.
• Cov: Covariance Pooling.
• GeM-1: Generalized-Mean Pooling for power=1(average pooling).
• GeM-2: Generalized-Mean Pooling for power=2.
• GeM-4: Generalized-Mean Pooling for power=4.

Code:

• experiments.ipynb: notebook to reproduce results in table 1 and wall-clock analysis in our paper.
• sensitivity_to_code_length.py: scripts to perform sensitivity analysis on hash code length.
• sensitivity_to_num_slices.py: scripts to perform sensitivity analysis on the number of slices used in SLOSH.