Neural Optimal Transport
Authors: Alexander Korotin, Daniil Selikhanovych, Evgeny Burnaev
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the performance of our optimal transport algorithm on toy examples and on the unpaired image-to-image translation. (a) Celeba (female) anime, outdoor church, deterministic (one-to-one, W2). (b) Handbags shoes, stochastic (one-to-many, W2,1). Figure 1: Unpaired translation with our Neural Optimal Transport (NOT) Algorithm 1. |
| Researcher Affiliation | Academia | 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: Neural optimal transport (NOT) |
| Open Source Code | Yes | The code is written in Py Torch framework and is publicly available at https://github.com/iamalexkorotin/Neural Optimal Transport |
| Open Datasets | Yes | We use the following publicly available datasets as P, Q: aligned anime faces3, celebrity faces (Liu et al., 2015), shoes (Yu & Grauman, 2014), Amazon handbags, churches from LSUN dataset (Yu et al., 2015), outdoor images from the MIT places database (Zhou et al., 2014). |
| Dataset Splits | No | Train-test split. We pick 90% of each dataset for unpaired training. The rest 10% are considered as the test set. No explicit validation split is mentioned for the main experiments. |
| Hardware Specification | Yes | Our networks converge in 1-3 days on a Tesla V100 GPU (16 GB); wall-clock times depend on the datasets and the image sizes. Training stochastic T(x, z) is harder since we sample multiple random z per x (we use |Z| = 4). Thus, we learn stochastic maps on 4 Tesla V100 GPUs. |
| Software Dependencies | No | The paper mentions 'Py Torch framework' and 'Adam optimizer (Kingma & Ba, 2014)' but does not specify exact version numbers for these software dependencies. |
| Experiment Setup | Yes | The learning rate is lr = 1 × 10−4. The batch size is |X| = 64. The number of inner iterations is kT = 10. When training with the weak cost (4), we sample |Zx| = 4 noise samples per each image x in batch. |