PSWE: Pooling by Sliced-Wasserstein Embedding (NeurIPS 2021)

PSWE is a permutation-invariant feature aggregation/pooling method based on sliced-Wasserstein distances that generates an output embedding from a set of input features, whose dimensionality does not depend on the input set size.

Run ModelNet40 experiments

Use the following command to run the complete experiments (for PSWE and other pooling methods) on the ModelNet40 point cloud dataset:

python3 ModelNet40_train_test.py

Dependencies

• PyTorch
• NumPy
• torchinterp1d
• h5py
• tqdm

Citation

Please use the following BibTeX citation if you use this repository in your work:

@inproceedings{naderializadeh2021PSWE,
  title={Pooling by Sliced-{Wasserstein} Embedding},
  author={Navid Naderializadeh and Joseph F. Comer and Reed W Andrews and Heiko Hoffmann and Soheil Kolouri},
  booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021},
  url={https://openreview.net/forum?id=1z2T01DKEaE}
}