https://github.com/jaydu1/ood-ridge

Optimal Ridge Regularization for Out-of-Distribution Prediction
https://github.com/jaydu1/ood-ridge

Last synced: about 1 month ago
JSON representation

Optimal Ridge Regularization for Out-of-Distribution Prediction

• Host: GitHub
• URL: https://github.com/jaydu1/ood-ridge
• Owner: jaydu1
• License: mit
• Created: 2024-03-31T03:30:14.000Z (8 months ago)
• Default Branch: main
• Last Pushed: 2024-03-31T13:15:50.000Z (8 months ago)
• Last Synced: 2024-04-01T04:24:54.038Z (8 months ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 472 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Optimal Ridge Regularization for Out-of-Distribution Prediction

This repository contains code for reproducing results in the paper *Optimal Ridge Regularization for Out-of-Distribution Prediction*.

## Scripts

The following files are included in this repository:

- Util functions:

    - `generate_data.py`: A Python script that generates the data set.

    - `compute_risk.py`: A Python script that computes the empirical and theoretical risks of ridge predictor and ensembles.

    - `fixed_point_sol.py`: A Python script that computes the fixed-point solutions in $(\lambda,\phi)$.

- Ex1

    - `ex1_equiv_lam_min.py`: compute the minimum feasible $\lambda$.

    - `ex1_opt_ridge.py`: compute the in-distribution risk of ridge predictors.

    - `ex1_opt_ridge_ood.py`: compute the OOD risk of ridge predictors.

- Ex2

    - `ex2_MNIST.py`: compute the OOD risk with distribution shifts on MNIST datasets.

- Ex3

    - `ex3_mono.py`: compute the ridge risk at different values of $\lambda$.

    - `ex3_MNIST.py`: compute the OOD risk at different values of $\lambda$ on MNIST datasets.

- Ex4

    - `ex4_equiv_v.py`: compute the fixed-point solutions in $(\lambda,\phi)$.

    - `ex4_equiv_risk.py`: compute the risks of ridge predictors in $(\lambda,\phi)$.

- Ex5: Figure F8

    - `ex5_theory_ridge_opt.py`: theoretical risk of ridge predictors.

    - `ex5_theory_ridgeless.py`: theoretical risk of full-ensemble ridgeless predictors.

- Visualization:

    - `Plot.ipynb`: A Jupyter notebook that visualizes the results.

## Computation details

All the experiments are run on Ubuntu 22.04.2 LTS (GNU/Linux 5.15.0-72-generic x86_64) using 12 cores.

The estimated time to run all experiments is roughly less than 2 hours for each script.