Metric Learning in an RKHS

This repository provides code for the simulations, experiments, and datasets in our paper: Metric Learning in an RKHS

Requirements

• python3.12
• You need to have a comet.ml account, as we upload the data to the comet.ml server the pull them to analyze the results.

How to run the code

Run the simulation

python run_tatli_new.py -m r=2,3,4,5,6,7,8,9,10 T=10000 seed=1,2,3,4,5 solver=MOSEK project_name=[some project name] d=10 eigen_cutoff=10 noise_param=100 noise=true

Run the spiral simulation

python run_spiral.py -m T=100,500,1000,2500,5000,7500,10000 seed=0,1,2,3,4 project_name=spiral kernel_type=linear,gaussian,sigmoid,poly,laplacian

Run the experiment

python run_tatli_experiment_v1.py -m project_name=food100_img2vec_v1_gaussian_hypsearch kernel_params.sigma=0.1,0.2,0.5,1,2,5 loss_type=hinge

Acknowledgements

This work was partially supported by NSF grants NCS-FO
2219903 and NSF CAREER Award CCF 2238876