Light Unbalanced Optimal Transport

Authors: Milena Gazdieva, Arip Asadulaev, Evgeny Burnaev, Aleksandr Korotin

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide the generalization bounds for our solver (M4.4) and experimentally test it on several tasks (M5.1, M5.2). ... In this section, we test our U-Light OT solver on several setups from the related works.
Researcher Affiliation Academia Milena Gazdieva Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia milena.gazdieva@skoltech.ru Arip Asadulaev ITMO University Artificial Intelligence Research Institute Moscow, Russia asadulaev@airi.net Evgeny Burnaev Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia e.burnaev@skoltech.ru Alexander Korotin Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia a.korotin@skoltech.ru
Pseudocode No The paper does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The code is publicly available at https://github.com/milenagazdieva/Light Unbalanced Optimal Transport
Open Datasets Yes We follow the setup of [41, Section 5.4] and use pre-trained ALAE autoencoder [55] for 1024 1024 FFHQ dataset [34] of human faces.
Dataset Splits No The paper mentions 'Number of train FFHQ images for each subset' in Table 2, but does not explicitly state specific train/validation/test dataset splits (e.g., percentages or sample counts for each split) for its experiments.
Hardware Specification Yes Each experiment requires several minutes of training on CPU with 4 cores.
Software Dependencies No The code is written using Py Torch framework and is publicly available at... We use the Adam optimizer... To obtain an optimal transport plan π (x, y) discrete OT solvers from the POT [21] package were used.
Experiment Setup Yes We use K = L = 5, ε = 0.05, lr = 3e 4 and batchsize 128. We do 2 104 gradient steps... for our solver, we use weighted DKL divergence with parameters τ specified in Appendix C, and set K = L = 10, ε = 0.05, lr = 1, and batch size to 128. We do 5 103 gradient steps using Adam optimizer [35] and Multi Step LR scheduler with parameter γ = 0.1 and milestones= [500, 1000].