Variational Wasserstein gradient flow
Authors: Jiaojiao Fan, Qinsheng Zhang, Amirhossein Taghvaei, Yongxin Chen
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The performance and scalability of the proposed method are illustrated with the aid of several numerical experiments involving high-dimensional synthetic and real datasets.In this section, we present several numerical examples to illustrate our algorithm. |
| Researcher Affiliation | Academia | Jiaojiao Fan 1 Qinsheng Zhang 1 Amirhossein Taghvaei 2 Yongxin Chen 1 1Georgia Institute of Technology 2University of Washington, Seattle. |
| Pseudocode | Yes | Algorithm 1 Primal-dual gradient flow |
| Open Source Code | Yes | Our code is written in Py Torch-lightning and is publicly available at https://github.com/sbyebss/ variational_wgf. |
| Open Datasets | Yes | We numerically demonstrate the performance of our algorithm on several representative problems including sampling from high-dimensional Gaussian mixtures, porous medium equation, and learning generative models on MNIST and CIFAR10 datasets.We test on 8 relatively small datasets (S <= 7400) from Mika et al. (1999) and one large Covertype dataset (S = 0.58M). The dataset is randomly split into training dataset and test dataset according to the ratio 4:1. (https://www.csie.ntu.edu.tw/ cjlin/ libsvmtools/datasets/binary.html) |
| Dataset Splits | No | The dataset is randomly split into training dataset and test dataset according to the ratio 4:1. The paper does not explicitly mention a separate validation split or its proportion. |
| Hardware Specification | Yes | Our experiments are conducted on Ge Force RTX 3090 or RTX A6000. We always make sure the comparison is conducted on the same GPU card when comparing training time with other methods. |
| Software Dependencies | No | Our code is written in Pytorch-Lightning (Falcon & Cho, 2020). We use other wonderful python libraries including W&B (Biewald, 2020), hydra (Yadan, 2019), seaborn (Waskom, 2021), etc. Specific version numbers for these software dependencies are not provided. |
| Experiment Setup | Yes | Without further specification, we use the following parameters: The number of iterations: J1 = 600. J2 = 3. J3 = 1. The batch size is fixed to be M = 100. The learning rate is fixed to be 0.001. All the activation functions are set to be PRe Lu. h has 3 layers and 16 neurons in each layer. T has 4 layers and 16 neurons in each layer. |