Energy-guided Entropic Neural Optimal Transport

Authors: Petr Mokrov, Alexander Korotin, Alexander Kolesov, Nikita Gushchin, Evgeny Burnaev

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

Reproducibility Variable Result LLM Response
Research Type Experimental In practice, we validate its applicability in toy 2D and image domains.
Researcher Affiliation Academia 1Skolkovo Institute of Science and Technology, Moscow, Russia 2Artificial Intelligence Research Institute, Moscow, Russia {petr.mokrov,a.korotin}@skoltech.ru
Pseudocode Yes Algorithm 1: Entropic Optimal Transport via Energy-Based Modelling
Open Source Code Yes Our code is available at: https: //github.com/Petr Mokrov/Energy-guided-Entropic-OT
Open Datasets Yes The AFHQ dataset is taken from the Star GAN v2 (Choi et al., 2020) github: https://github.com/clovaai/stargan-v2. For generating Colored MNIST dataset, we make use of the code, provided by the authors of (Gushchin et al., 2023).
Dataset Splits No The paper mentions drawing samples from source and target distributions (e.g., 'Derive batches {xn}N n=1 = X P, {yn}N n=1 = Y Q of sizes N' in Algorithm 1, and 'For the source and target distributions, we choose P = N(0, ΣX) and Q = N(0, ΣY )' in D.2), but does not explicitly provide specific training, validation, or test dataset split percentages or sample counts.
Hardware Specification Yes The experiment was conducted on a single GTX 1080 Ti (D.1), The training takes approximately a day on 4 A100 GPUs. (D.3), It takes approximately one day on one V100 GPU (D.4).
Software Dependencies No The paper states 'Our code is written in Py Torch', but it does not specify version numbers for PyTorch or any other software dependencies used in the experiments.
Experiment Setup Yes The hyperparameters of Algorithm 1 are as follows: K = 100, σ0 = 1, N = 1024, η = 0.05 for ε = 0.1 and η = 0.005 for ε = 0.001. (D.1) and MLP with Re LU activations and three hidden layers with 1024, 512 and 256 neurons, accordingly. The training hyperparameters are: K = 100, η = 0.008, σ0 = 1.0, N = 128. For both Cat Dog and Wild Dog experiments, we train our model for 11 epochs with a learning rate 10 4 which starts monotonically decreasing to 10 5 after the fifth epoch. (D.3).