https://github.com/pomonam/jax-influence
A simple Jax implementation of influence functions.
https://github.com/pomonam/jax-influence
influence-functions
Last synced: 9 months ago
JSON representation
A simple Jax implementation of influence functions.
- Host: GitHub
- URL: https://github.com/pomonam/jax-influence
- Owner: pomonam
- License: apache-2.0
- Created: 2022-11-16T21:21:46.000Z (over 3 years ago)
- Default Branch: main
- Last Pushed: 2024-04-09T18:21:30.000Z (about 2 years ago)
- Last Synced: 2024-12-17T18:39:26.062Z (over 1 year ago)
- Topics: influence-functions
- Language: Python
- Homepage:
- Size: 39.1 KB
- Stars: 16
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
## JAX-Influence
[](LICENSE.txt)
JAX-Influence is a JAX implementation of influence functions, a classical technique from robust statistics that
estimates the impact of removing a single training data point on a model's learned parameters. This repository
complements the paper ["If Influence Functions are the Answer, Then What is the Question?"](https://arxiv.org/abs/2209.05364).
The repository aims to provide a simple and minimal implementation of influence functions in JAX. For those interested in
implementations in other frameworks, a PyTorch version is available [here](https://github.com/alstonlo/torch-influence), and
a PyTorch EK-FAC implementation can be found [here](https://github.com/pomonam/kronfluence).
## Installation
To install JAX-Influence, you can use pip to install from the source:
```bash
git clone https://github.com/pomonam/jax-influence
cd jax-influence
pip install -e .
pip install -e '.[jax_gpu]' -f 'https://storage.googleapis.com/jax-releases/jax_cuda_releases.html' # Replace `jax_gpu` with `jax_cpu` if you wish to install the CPU version.
```
## Contributors
- [Juhan Bae](https://www.juhanbae.com/)
- [Nathan Ng](https://nng555.github.io/)
- [Alston Lo](https://alstonlo.github.io/)