Improving Mini-batch Optimal Transport via Partial Transportation

Authors: Khai Nguyen, Dang Nguyen, The-Anh Vu-Le, Tung Pham, Nhat Ho

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we carry out extensive experiments on various applications such as deep domain adaptation, partial domain adaptation, deep generative model, color transfer, and gradient flow to demonstrate the favorable performance of m-POT compared to current mini-batch methods.
Researcher Affiliation Collaboration Khai Nguyen 1 * Dang Nguyen 2 * The-Anh Vu-Le 2 Tung Pham 2 Nhat Ho 1 1Department of Statistics and Data Sciences, The University of Texas at Austin 2Vin AI Research.
Pseudocode Yes Algorithm 1 Mini-batch Deep Domain Adaptation, Algorithm 2 Two-stage mini-batch Deep Domain Adaptation, Algorithm 3 Mini-batch Deep Generative Model
Open Source Code Yes Python code is published at https://github.com/UT-Austin-Data-Science-Group/Mini-batch-OT.
Open Datasets Yes The USPS dataset (Hull, 1994) consists of 7, 291 training and 2, 007 testing images... The MNIST dataset (Le Cun et al., 1998) contains 60, 000 training and 10, 000 testing grayscale images... The SVHN dataset (Netzer et al., 2011) contains house numbers extracted from Google Street View images... Next, we consider the Office-Home dataset (Venkateswara et al., 2017)... Finally, Vis DA-2017 (Peng et al., 2017) is a large-scale dataset... We train generators on CIFAR10 (Krizhevsky et al., 2009) and Celeb A (Liu et al., 2015) datasets.
Dataset Splits Yes The USPS dataset (Hull, 1994) consists of 7, 291 training and 2, 007 testing images... The MNIST dataset (Le Cun et al., 1998) contains 60, 000 training and 10, 000 testing grayscale images... The SVHN dataset (Netzer et al., 2011) has 73, 212 training images, and 26, 032 testing RGB images... For Vis DA: Vis DA consists of 152, 397 source (synthetic) images and 55, 388 target (real) images... Following JUMBOT (Fatras et al., 2021a), we evaluate all methods on the validation set. For CIFAR10: CIFAR10 dataset contains 10 classes, with 50, 000 training and 10, 000 testing color images of size 32 32.
Hardware Specification Yes Color transfer and gradient flow applications are conducted on a Mac Book Pro 11 inch M1. While other deep learning experiments are done on NVIDIA Tesla V100 GPUs.
Software Dependencies No The paper mentions using the 'POT package (Flamary et al., 2021)' and that 'Our source code is based on https://github.com/kilian Fatras/JUMBOT.' It does not specify version numbers for these software dependencies or other programming languages/libraries used.
Experiment Setup Yes Parameter settings for Digits datasets: We optimize using Adam with the initial learning rate η0 = 0.0004}. The number of mini-batches k for each method is set to either 1 or 2... We train all algorithms with a batch size of 500 during 100 epochs. The hyperparameters in equation (14) follow the setting in m-UOT: α = 0.1, λt = 0.1. For computing UOT, we also set τ to 1 and ϵ to 0.1.