https://github.com/pnnl/ddks

A high-dimensional Kolmogorov-Smirnov distance for comparing high dimensional distributions
https://github.com/pnnl/ddks

machine-learning physics statistics

Last synced: 29 days ago
JSON representation

A high-dimensional Kolmogorov-Smirnov distance for comparing high dimensional distributions

• Host: GitHub
• URL: https://github.com/pnnl/ddks
• Owner: pnnl
• License: other
• Created: 2021-06-24T14:50:30.000Z (over 3 years ago)
• Default Branch: master
• Last Pushed: 2022-09-20T14:27:38.000Z (over 2 years ago)
• Last Synced: 2024-10-29T10:55:34.604Z (about 2 months ago)
• Topics: machine-learning, physics, statistics
• Language: Jupyter Notebook
• Homepage:
• Size: 2.15 MB
• Stars: 23
• Watchers: 5
• Forks: 6
• Open Issues: 3
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.md

Awesome Lists containing this project

README

        
# ddKS - a d-dimensional Kolmogorov-Smirnov Test

*Alex Hagen1, Shane Jackson1, James Kahn2, Jan Strube1, Isabel Haide2, Karl Pazdernik1, and Connor Hainje1*

1: Pacific Northwest National Laboratory, 

2: Karlsruhe Institute of Technology

This code accompanies our paper submitted to IEEE Transactions on

Pattern Analysis and Machine Intelligence titled "Accelerated Computation of a

High Dimensional Kolmogorov-Smirnov Distance" ([arXiv](https://arxiv.org/abs/2106.13706)).  

As of 6/25/2021 there are 3 methods implemented:

* ddKS - d-dimensional KS test caclulated per

    * Variable splitting of space (all points, subsample, grid spacing)

* rdKS - ddKS approximation using distance from (d+1) corners

* vdKS - ddKS approximation calculating ddks distance between voxels instead of points

# Quickstart

Installation of `ddks` should be pretty easy, simple run

```bash

pip install git+https://github.com/pnnl/DDKS

```

or, if you want to develop on DDKS, simply clone this repository into a safe

spot on your computer and run

```bash

pip install -e .

```

from the top level of the repository.

Then, you can get started used the

repository by starting a `ddks` object and performing the distance calculation

on any pair of torch tensors that are `sample_size` x `dimension`.

```python

import torch

import ddks

p = torch.rand((100, 3))

t = torch.rand((50, 3))

calculation = ddks.methods.ddKS()

distance = calculation(p, t)

print(f"The ddKS distance is {distance}")

```

To operate on GPU, all you need to do is move the tensors to the device before

calculation:

```python

p = torch.rand((100, 3)).to('cuda:0')

t = torch.rand((50, 3)).to('cuda:0')

calculation = ddks.methods.ddKS()

distance = calculation(p, t)

```

If you want to use a different accelerated method, simply use

`ddks.methods.rdKS` or `ddks.methods.vdKS`. Note that rdKS and vdKS cannot use

GPU.

# Package Structure:

1. methods - Callable classes for xdks methods [x=d,r,v]

1. data - Contains several data generators to play around with

1. run_scripts - Contains an example run script 

1. Unit_tests - Contains unit tests for repo