Sliced-Wasserstein Flows: Nonparametric Generative Modeling via Optimal Transport and Diffusions

Authors: Antoine Liutkus, Umut Simsekli, Szymon Majewski, Alain Durmus, Fabian-Robert Stöter

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results support our theory and show that our algorithm is able to successfully capture the structure of different types of data distributions. In this section, we evaluate the SWF algorithm on a synthetic and a real data setting.
Researcher Affiliation Academia 1Inria and LIRMM, Univ. of Montpellier, France 2LTCI, T el ecom Paristech, Universit e Paris-Saclay, Paris, France 3Institute of Mathematics, Polish Academy of Sciences, Warsaw, Poland 4CNRS, ENS Paris-Saclay,Universit e Paris-Saclay, Cachan, France.
Pseudocode Yes Algorithm 1: Sliced-Wasserstein Flow (SWF)
Open Source Code No The paper does not include any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We test the SWF algorithm on two real datasets. (i) The traditional MNIST dataset that contains 70K binary images corresponding to different digits. (ii) The popular Celeb A dataset (Liu et al., 2015), that contains 202K color-scale images.
Dataset Splits No The paper mentions using a "training set" but does not provide specific details on how the data was split into training, validation, or test sets (e.g., percentages, sample counts, or citations to predefined splits).
Hardware Specification No The paper states that the algorithm "can be easily run on an everyday laptop CPU" but does not provide specific details about the hardware (e.g., CPU model, GPU model, memory) used for the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes In all our experiments, the initial distribution µ0 is selected as the standard Gaussian distribution on Rd, we take Q = 100 quantiles and N = 5000 particles, which proved sufficient to approximate the quantile functions accurately. Here, we set Nθ = 30, h = 1 and λ = 10 4. In the following, we set λ = 0, Nθ = 40000, d = 32 for MNIST and d = 64 for Celeb A.