Nonparametric Generative Modeling with Conditional Sliced-Wasserstein Flows

Authors: Chao Du, Tianbo Li, Tianyu Pang, Shuicheng Yan, Min Lin

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we first examine the efficacy of the proposed techniques of locally-connected projections and pyramidal schedules. We then demonstrate that with these techniques our ℓ-CSWF further enables superior performances on conditional modeling tasks, including class-conditional generation and image inpainting. We report the FID scores (Heusel et al., 2017) on CIFAR10 and Celeb A in Table 1 for quantitative evaluation.
Researcher Affiliation Industry 1Sea AI Lab, Singapore. Correspondence to: Chao Du <duchao@sea.com>, Min Lin <linmin@sea.com>.
Pseudocode Yes Algorithm 1: Conditional Sliced-Wasserstein Flow. Algorithm 2: Sliced-Wasserstein Flow (SWF) (Liutkus et al., 2019)
Open Source Code Yes Code is available at https://github.com/duchao0726/Conditionial-SWF.
Open Datasets Yes We use MNIST, Fashion-MNIST (Xiao et al., 2017), CIFAR10 (Krizhevsky et al., 2009) and Celeb A (Liu et al., 2015) datasets in our experiments.
Dataset Splits No The paper mentions using training and test data, but it does not explicitly specify train/validation/test dataset splits with percentages or sample counts for reproducibility. It states 'We augment the CIFAR-10 dataset with horizontally flipped images, resulting in a total of 100000 training images' and later 'test split of each dataset'.
Hardware Specification No The paper does not provide specific details about the hardware used to run its experiments. It mentions 'limited computing resources' but no model numbers or specifications.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with versions) needed to replicate the experiment.
Experiment Setup Yes For all experiments, we set H = 10000 for the number of projections in each step and set the step size η = d. The number of simulation steps K varies from 10000 to 20000 for different datasets, due to different resolutions and pyramidal schedules. For MNIST and Fashion-MNIST, we set M = 2.5 105. For CIFAR10 and Celeb A, we set M = 7 105 and M = 4.5 105, respectively. We set the amplifier ξ = 10 for all datasets.