SSNS

Code for Safe and Sparse Newton Method for Entropic-Regularized Optimal Transport.

Environment

Use the following commands to create a virtual environment and install necessary packages:

conda create -n ssns
conda activate ssns
conda install python=3.10 numpy scipy matplotlib seaborn pandas scikit-learn
pip install regot python-mnist 

Experiments

We evaluate the SSNS algorithm's performance through various numerical experiments, focusing on entropic-regularized optimal transport (OT). We compare SSNS with other optimization methods, including the Sinkhorn algorithm, APDAGD, L-BFGS, and the globalized Newton method.

Datasets and Experiment Setup

We use three benchmark datasets to define the OT problem, with experiments conducted on both (Fashion-)MNIST and ImageNet datasets:

• (Fashion-)MNIST: fashion and data_mnist
• ImageNet: train_feature, train_feature_60 and train_feature_90

Code for Experiments

Convert ImageNet data to feature vectors

feature.py

Save runtime data for Different Algorithms:

• Comparative Experiments: get-pkl-for-all.py
• Impact of Feature Dimension: get-pkl-for-feature.py
• Scalability: get-pkl-for-large-scale.py

Visualization

• Comparative Experiments: plot-pdf-for-all.py
• Impact of Feature Dimension: plot-pdf-for-feature.py
• Scalability: plot-pdf-for-large-scale.py