Revisiting Sliced Wasserstein on Images: From Vectorization to Convolution
Authors: Khai Nguyen, Nhat Ho
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we demonstrate the favorable performance of CSW over the conventional sliced Wasserstein in comparing probability measures over images and in training deep generative modeling on images. In this section, we focus on comparing the sliced Wasserstein (SW) (with the conventional slicing), the convolution-base sliced Wasserstein (CSW-b), the convolution sliced Wasserstein with stride (CSW-s), and the convolution sliced Wassersstein with dilation (CSW-d) in training generative models on standard benchmark image datasets such as CIFAR10 (32x32) [27], STL10 (96x96) [8], Celeb A (64x64), and Celeb A-HQ (128x128) [37]. |
| Researcher Affiliation | Academia | Khai Nguyen Department of Statistics and Data Sciences The University of Texas at Austin Austin, TX 78712 khainb@utexas.edu Nhat Ho Department of Statistics and Data Sciences The University of Texas at Austin Austin, TX 78712 minhnhat@utexas.edu |
| Pseudocode | No | The paper includes mathematical definitions and derivations but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code for the paper is published at https://github.com/UT-Austin-Data-Science-Group/CSW. |
| Open Datasets | Yes | training deep generative modeling on standard benchmark image datasets such as CIFAR10 (32x32) [27], STL10 (96x96) [8], Celeb A (64x64), and Celeb A-HQ (128x128) [37]. |
| Dataset Splits | No | The paper mentions using standard benchmark datasets and training set sizes (e.g., '50k images from CIFAR10 training set'), but does not explicitly provide details about training, validation, and test splits (e.g., percentages, sample counts for each, or citations to specific predefined splits). |
| Hardware Specification | Yes | We train all models on 1 NVIDIA Tesla V100 GPU. |
| Software Dependencies | Yes | The code is implemented in PyTorch 1.10.1. |
| Experiment Setup | Yes | The details of the training are given in Appendix D.2. The detailed settings about architectures, hyperparameters, and evaluation of FID and IS are given in Appendix E. We use the Adam optimizer [23] with β1 = 0.5, β2 = 0.9, and the learning rate 0.0001. |