A User's Guide to Sampling Strategies for Sliced Optimal Transport

Authors: Keanu Sisouk, Julie Delon, Julien Tierny

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on both simulated and real-world data offer a representative comparison of the strategies, culminating in practical recommendations for their best usage.
Researcher Affiliation Academia Keanu SISOUK EMAIL CNRS, LIP6 Sorbonne University Julie Delon EMAIL MAP5 Paris Cite University Julien Tierny EMAIL CNRS, LIP6 Sorbonne University
Pseudocode No The paper describes algorithms and methods in detail using mathematical notation and textual explanations, but it does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes The implementations used are grouped and are available here https://github.com/Keanu-Sisouk/SW-Sampling-Guide.
Open Datasets Yes Extensive experiments on both simulated and real-world data offer a representative comparison of the strategies... We compare different estimates of SW 2 2 (µd, νd) for d {2, 3, 5, 10, 20, 50}... We took three 3D point clouds issued from the Shapenet Core dataset introduced by [Chang et al., 2015]... we select the classical MNIST dataset [Le Cun, 1998].
Dataset Splits No The paper describes how specific datasets (Gaussian mixtures, Shapenet Core, MNIST) were constructed or sampled for the experiments (e.g., 'N = 1000' for Gaussian, 'select randomly 600 digit images and divide them into groups of 200' for MNIST). However, it does not specify explicit training/validation/test splits as it primarily focuses on evaluating sampling methods for distance computation rather than training a predictive model.
Hardware Specification No The paper does not explicitly mention any specific hardware used for running the experiments, such as CPU or GPU models, or specific computational resources.
Software Dependencies No The paper mentions several software components and libraries (e.g., python, scipy.stats.ortho_group, scipy.stats.qmc, POT library), but it does not provide specific version numbers for these dependencies, nor for the Python interpreter itself.
Experiment Setup Yes Riesz point configuration: We use a code provided by François Clement... where we choose the number of iterations as T = 10, the gradient step as 1 and s = 0.1. Spherical Sliced Wasserstein: For the hyper-parameters we set the number of iteration T = 250, the learning rate ϵ = 150 and the number of great circles L = 500. Spherical Harmonics Control Variates: They provide two possible functions SHCV and SW_CV... we use both functions and always keep only the minimal error among the two.