Skip to content

Latest commit

 

History

History
59 lines (33 loc) · 2.3 KB

README.md

File metadata and controls

59 lines (33 loc) · 2.3 KB
Raw

PI-Net: A Deep Learning Approach to Extract Topological Persistence Images

Paper

This repository contains:

  • Sample code to compute persistence images (PIs) using the proposed Image PI-Net and Signal PI-Net models.

  • The provided pretrained Image PI-Net model was trained using the CIFAR10 dataset, and the provided pretrained Signal PI-Net model was trained using the USC-HAD dataset. Scikit-TDA was used to generate ground-truth PIs.

Note, performance of the provided Signal PI-Net model differs from that described in the paper, since the original model was pretrained using the GENEactiv dataset which is not publicly released.

Image PI-Net

For sample test-set images in CIFAR10, both "main_image_pinet.py" and "main_image_pinet.ipynb" do the following:

  • Load weights from a pretrained Image PI-Net model;

  • Compute PIs using the Image PI-Net model;

  • Compare the generated PIs to ground-truth PIs obtained using conventional topological data analysis (TDA) tools.

Additionally, the "main.py" file saves the PI comparisons for each sample image in the "Examples folder".

Signal PI-Net

Here, we provide pretrained model weights and sample code to train the Signal PI-Net model using the USC-HAD dataset.

To extract ground-truth persistence images for your time-series data, please refer to the following repository: Sublevel-Set-TDA.

Required Packages

Python 3.8.5 was used to create an environment with the following packages.

  • tensorflow == 2.10.0
  • keras == 2.10.0
  • matplotlib
  • numpy
  • scikit-learn
  • scikit-tda

Citation

@inproceedings{som2020pi,
  title={PI-Net: A Deep Learning Approach to Extract Topological Persistence Images},
  author={Som, Anirudh and Choi, Hongjun and Ramamurthy, Karthikeyan Natesan and Buman, Matthew P and Turaga, Pavan},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops},
  pages={834--835},
  year={2020}
}