Stochastic Optimization for Large-scale Optimal Transport

Authors: Aude Genevay, Marco Cuturi, Gabriel Peyré, Francis Bach

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We backup these claims on a set of discrete, semi-discrete and continuous benchmark problems.Numerical Illustrations on Bags of Word-Embeddings.Numerical Illustrations. Simulations are performed in X = Y = R3.
Researcher Affiliation Academia Aude Genevay CEREMADE, Université Paris-Dauphine INRIA Mokaplan project-team genevay@ceremade.dauphine.fr; Marco Cuturi CREST, ENSAE Université Paris-Saclay marco.cuturi@ensae.fr; Gabriel Peyré CNRS and DMA, École Normale Supérieure INRIA Mokaplan project-team gabriel.peyre@ens.fr; Francis Bach INRIA Sierra project-team DI, ENS francis.bach@inria.fr
Pseudocode Yes Algorithm 1 SAG for Discrete OT; Algorithm 2 Averaged SGD for Semi-Discrete OT; Algorithm 3 Kernel SGD for continuous OT
Open Source Code No The paper does not provide any statement or link for concrete access to source code for the described methodology.
Open Datasets Yes We use Glove word embeddings [14] to represent words, namely X = Y = R300.
Dataset Splits No The paper describes sampling of data (e.g., 'sample N = 20, 000 words'), but does not provide specific dataset split information (train/validation/test percentages or counts) needed to reproduce data partitioning for model training or evaluation.
Hardware Specification Yes We used 4 Tesla K80 cards to compute both SAG and Sinkhorn results.
Software Dependencies No The paper mentions algorithms like Sinkhorn and SAG, but does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes We set ε to 0.01 (other values are considered in the supplementary material).Following the recommendations in [19], SAG s stepsize is tested for 3 different settings, 1/L, 3/L and 5/L.We used minibatches of size 200 for SAG.SGD for 107 iterations.Each mixture is composed of three gaussians whose means are drawn randomly in [0, 1]3, and their correlation matrices are constructed as Σ = 0.01(RT + R) + 3I3 where R is 3 3 with random entries in [0, 1].