Amortized Projection Optimization for Sliced Wasserstein Generative Models
Authors: Khai Nguyen, Nhat Ho
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the favorable performance of the proposed sliced losses in deep generative modeling on standard benchmark datasets. We carry out extensive experiments on standard benchmark datasets including CIFAR10, Celeb A, STL10, and Celeb AHQ to demonstrate the favorable performance of A-SW in learning generative models. |
| 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 | Yes | Algorithm 1 Max-sliced Wasserstein distance. Algorithm 2 Training generative models with mini-batch max-sliced Wasserstein loss. Algorithm 3 Training generative models with amortized sliced Wasserstein loss. |
| Open Source Code | Yes | Code for the paper is published at https://github.com/UT-Austin-Data-Science-Group/ Amortized SW. |
| Open Datasets | Yes | For datasets, we choose standard benchmarks such as CIFAR10 (32x32) [26], STL10 (96x96) [7], Celeb A (64x64), and Celeb AHQ (128x128) [32]. |
| Dataset Splits | No | The paper mentions using standard benchmark datasets and discusses training processes, but it does not explicitly provide specific train/validation/test dataset splits (e.g., percentages or sample counts) in the main text or the appendices directly referenced for details. |
| Hardware Specification | No | We would like to recall that reported numbers are under some errors due to the state of the computational device. An implementation technique that utilizes both RAM and GPUs memory for training sliced Wasserstein generative model was introduced in [27]. No specific hardware models or detailed specifications are provided. |
| Software Dependencies | No | The paper does not provide specific software dependency details with version numbers (e.g., Python 3.8, PyTorch 1.9) required to replicate the experiments. |
| Experiment Setup | Yes | The detail of the training processes of all models is given in Appendix C. For amortized models, we fix the slice learning rate 2 = 0.01. Also, we do a grid search on two hyperparameters of Max-SW, namely, the slice maximum number of iterations T2 2 {1, 10, 100} and the slice learning rate 2 2 {0.001, 0.01, 0.1}. The detailed settings about architectures, hyperparameters, and evaluation of FID and IS are given in Appendix E. |