On Transportation of Mini-batches: A Hierarchical Approach

Authors: Khai Nguyen, Dang Nguyen, Quoc Dinh Nguyen, Tung Pham, Hung Bui, Dinh Phung, Trung Le, 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 experiments on various applications including deep generative models, deep domain adaptation, approximate Bayesian computation, color transfer, and gradient flow to show that the Bo Mb-OT can be widely applied and performs well in various applications.
Researcher Affiliation Collaboration 1Department of Statistics and Data Sciences, The University of Texas at Austin 2Vin AI Research 3Monash University.
Pseudocode Yes Algorithm 1 Mini-batch Deep Generative Model with Bo Mb-OT; Algorithm 2 Mini-batch Deep Domain Adaptation with Bo Mb-OT; Algorithm 3 Color Transfer with Bo Mb-OT; Algorithm 4 Approximate Bayesian Computation; Algorithm 5 Approximate Bayesian Computation with Bo Mb-OT
Open Source Code Yes Python code is published at https://github.com/ UT-Austin-Data-Science-Group/Mini-batch-OT.
Open Datasets Yes We now show the deep generative model result on MNIST (Le Cun et al., 1998), CIFAR10 (32x32) (Krizhevsky et al., 2009), and Celeb A (64x64) (Liu et al., 2015)... For digits datasets, we adapt two digits datasets SVHN (Netzer et al., 2011) and USPS (Hull, 1994) to MNIST (Le Cun et al., 1998).
Dataset Splits No The paper mentions using training and test sets but does not explicitly describe any validation dataset splits (e.g., percentages or counts) or a specific validation strategy for reproducibility.
Hardware Specification Yes All deep learning experiments are done on a GTX 1080 Ti. Other experiments are done on a Mac Book Pro 11inc M1.
Software Dependencies No We use POT (Flamary et al., 2021) for OT solvers and the py ABC (Klinger et al., 2018) for the ABC experiments. The specific version numbers for these software dependencies are not provided.
Experiment Setup Yes Parameter settings: We chose a learning rate equal to 0.0005, batch size equal to 100, number of epochs in MNIST equal to 100, number of epochs in Celeb A equal to 25, number of epochs in CIFAR10 equal to 50. For d = SW2 we use the number of projections L = 1000 on MNIST and L = 100 on CIFAR10 and Celeb A. For d = W ϵ 2, we use ϵ = 1 for MNIST, ϵ = 50 for CIFAR10, ϵ = 40 for Celeb A. For the entropic regularized version of Bo Mb-OT, we choose the best setting for λ {1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 60, 70, 80}