Extremal Domain Translation with Neural Optimal Transport

Authors: Milena Gazdieva, Alexander Korotin, Daniil Selikhanovych, Evgeny Burnaev

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our algorithm on toy examples and on the unpaired image-to-image translation task.
Researcher Affiliation Collaboration Milena Gazdieva Skolkovo Institute of Science and Technology Moscow, Russia milena.gazdieva@skoltech.ru Alexander Korotin Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia a.korotin@skoltech.ru Daniil Selikhanovych Skolkovo Institute of Science and Technology Moscow, Russia selikhanovychdaniil@gmail.com Evgeny Burnaev Skolkovo Institute of Science and Technology Artificial Intelligence Research Institute Moscow, Russia e.burnaev@skoltech.ru
Pseudocode Yes Algorithm 1: Procedure to compute the IT map between P and Q for transport cost c(x, y) and weight w.
Open Source Code Yes The code is publicly available at https://github.com/milenagazdieva/Extremal Neural Optimal Transport
Open Datasets Yes Image datasets. We utilize the following publicly available datasets as P, Q: celebrity faces [46], aligned anime faces3, flickr-faces-HQ [36], comic faces4, Amazon handbags from LSUN dataset [68], shoes [67], textures [16] and chairs from Bonn furniture styles dataset [1].
Dataset Splits No The paper specifies a 'Train-test split' with 90% for training and 10% for test, but does not explicitly mention a separate validation split or how hyperparameters were tuned.
Hardware Specification Yes In general, it takes from 2 (for w = 1) up to 6 (for w = 8) days on a single Tesla V100 GPU.
Software Dependencies No The paper mentions 'Py Torch framework' but does not specify a version number for it or other software dependencies.
Experiment Setup Yes Optimization. We employ Adam [38] optimizer with the default betas both for Tθ and fψ. The learning rate is lr = 10 4. We use the Multi Step LR scheduler which decreases lr by 2 after [(5+5 w)K, (20+5 w)K, (40+5 w)K, (70+5 w)K] iterations of fψ where w {1, 2, 4, 8} is a weight parameter. The batch size is |X| = 256 for toy Wi-Fi , |X| = 4096 for toy Accept, and |X| = 64 for image-to-image translation experiments. The number of inner Tθ iterations is k T = 10.