Kernel Methods for Machine Learning

MoSIG DSAI, MSIAM 2021/22 Grenoble INP - Ensimag

Introduction

This repository contains the code for the data challenge of the Kernel Methods for Machine Learning course. The task is image classification with kernel methods.

Authors

• Gabriele Degola

Public score: 0.616 Private score: 0.602 Leaderboard position: 3

Data

Images are drawn from the CIFAR-10 dataset and organized as csv files: each row contains an image of 32 x 32 pixels.

Run

The following command reproduces the best submission. Predictions are stored in Yte.csv.

python start.py --xtr data/Xtr.csv --ytr data/Ytr.csv --xte data/Xte.csv

Structure

Code is organized in the following files:

• start.py is the main script;
• models.py contains our implementation of kernel ridge regression;
• kernels.py contains implemented kernels;
• utils.py contains functions and classes for data processing.

Dependencies

Code relies on the following Python libraries:

• numpy, scipy for basic operations;
• pandas for data loading and storing;
• tqdm for showing progresses;
• scikit-image for computation of histograms of oriented gradients;
• scikit-learn for label binarization only, required for multi-class classification.