yixuan / temperflow

Efficient Multimodal Sampling via Tempered Distribution Flow

TemperFlow

This repository stores the code files for the article Efficient Multimodal Sampling via Tempered Distribution Flow by Yixuan Qiu and Xiao Wang.

Workflow

We provide two implementations of the TemperFlow algorithm, one using the PyTorch framework (in the torch folder), and the other using the TensorFlow framework (in the tf folder).

The workflow to reproduce the results and plots in the article is as follows:

1. Download the following two model files into the tf/pretrained folder.

   • face-gmmvae-generator.npz: https://1drv.ms/u/s!ArsORq8a24WmoHtrued5dY6APVtH?e=noYKZr
   • face-gmmvae-flow.npz: https://1drv.ms/u/s!ArsORq8a24WmoHwOx8nlOoXpXYEW?e=7aQ6ks

2. Install the CUDA and cuDNN environments if you use GPU for computing. An installation guide can be found at https://www.tensorflow.org/install/pip.

3. Install the PyTorch and TensorFlow frameworks, following the installation guides at https://pytorch.org/get-started/locally/ and https://www.tensorflow.org/install/pip.

4. Run the main script as follows:

   sh run_experiments.sh

   This will call individual scripts under the torch and tf directories. When the script finishes, two new folders, both named model, will be created under the torch and tf directories, respectively. They will contain model data that are further used to create tables and plots.

5. When Step 4 finishes, the two model folders would already contain image files for Figures 7, 9, 10, S2, S3, S4, and S11.

6. The individual R scripts in the visualization directory provide the code to generate other figures and tables, based on the model data created in previous steps. For example, the script fig1-figs1-kl-sampler.R produces Figure 1 and S1, and tabs1-benchmark.R outputs the numbers in Table S1.

Software environment

GPU computing:

• CUDA 11.7
• cuDNN 8.6.0

Python:

• Python 3.10.8
• Numpy 1.23.5
• SciPy 1.9.3
• Pandas 1.5.2
• POT 0.8.2
• Matplotlib 3.6.2
• Seaborn 0.12.1
• PyTorch 1.13.1
• Pyro 1.8.3
• TensorFlow 2.10.0
• TensorFlow Probability 0.18

R:

• R 4.2.0
• reticulate 1.25
• readr 2.1.2
• jsonlite 1.8.0
• tibble 3.1.8
• reshape2 1.4.4
• dplyr 1.0.9
• ggplot2 3.3.6
• GGally 2.1.2
• gridExtra 2.3
• rgl 0.108.3.2
• showtext 0.9.5
• transport 0.12.2
• kernlab 0.9.31
• nor1mix 1.3.0
• copula 1.1.0

Citation

Please consider to cite our work if you find our algorithm or software useful in your research.

@article{qiu2023efficient,
  title={Efficient Multimodal Sampling via Tempered Distribution Flow},
  author={Qiu, Yixuan and Wang, Xiao},
  journal={Journal of the American Statistical Association},
  number={just-accepted},
  year={2023},
  publisher={Taylor \& Francis},
  doi = {10.1080/01621459.2023.2198059}
}