pc-Jiang/NNGraphRepresentation

Course project for DSC205 - Geometry of Data

Neural Network Representation as Graph

Author: Pengcen Jiang, Yingyu Lin

Course: DSC205 Geometry of Data, FA23@UCSD

Instructor: Dr. Gal Mishine

This project is based on the paper: Shnitzer, Tal, et al. "Log-euclidean signatures for intrinsic distances between unaligned datasets." International Conference on Machine Learning. PMLR, 2022. paper [1] github

Instructions

Codes are built with Python version=3.11.5

Run pip install -r requirements.txt to install all the required packages

Optional packages

Optional packages and repositories for comparisons with other algorithms:

• TDA: H0, H1 and H2 bottleneck distances, requires persim , ripser.

  pip install cython
  pip install ripser
  pip install persim
  

• GS [2] - clone and place the gs folder in the current folder.
  Requires GUDHI and Cython.

  pip install cython
  pip install gudhi
  

• GW [3] - requires pot. pip install POT

Run experiments

Run python main.py -a [exp_name] for an experiment. Available experiments are in experiments.py.

For example:

python main.py -a compare_representation_across_models

References

[1] Shnitzer, Tal, et al., "Log-euclidean signatures for intrinsic distances between unaligned datasets", ICML, 2022.
[2] Khrulkov and Oseledets, "Geometry score: A method for comparing generative adversarial networks", ICML, 2018.
[3] Peyré et al., "Gromov-Wasserstein averaging of kernel and distance matrices", ICML, 2016.