Automated Spectral Kernel Learning

Intro

This repository provides the code used to run the experiments of the paper "Automated Spectral Kernel Learning" (https://arxiv.org/abs/1909.04894).

Environments

• Python 3.7.4
• Pytorch 1.2.0
• CUDA 10.1.168
• cuDnn 7.6.0
• GPU: Tesla P100 16GB

Core functions

• auto_kernel_learning.py implements the algorithm to construct an one-layer neural network, including initialization of trainable weights and untrainable biases as well as feature mapping (cosine as activation).
• utils.py implements useful tools including load svmlight style dataset and classic datasets used in Pytorch but also various loss functions are introduced.
• optimal_parameters.py records optimal parameters for the proposed algorithm.

Experiments

1. Download datasets for multi-class classification (https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/).
2. Run the script to tune parameters and record them in optimal_parameters.py.

python run_parameter_tune.py

3. Run the script to obtain results in Experiment section

python run_exp1.py