This repository provides the codes for the paper Semi-Implicit Variational Inference via Score Matching by Longlin Yu and Cheng Zhang.
Please use the following commands to install the requirements
pip install -r requirements.txt
The datasets used in our experiments are placed at ./datasets. In our experiments, we use the MNIST and UCI datasets. You can download the UCI datasets from https://archive.ics.uci.edu/.
For the toy examples of 2D synthetic tasks, we show the sampling results of variational distribution in the training dynamics.
python sivism_2d.py --config multimodal.yml
After completing the training process, the KL-divergence between the samples of target distributions and the samples of variational distributions can be estimated via the ITE packages. The samples of target distributions was formed by a long MCMC run, which was simulated through the following command.
python sgld_toyexample.py
We provide the result of approximated posterior distributions of Bayesian logistic regression on the waveform dataset. We compare the posterior estimates of SIVI-SM with the ground truth formed from a long MCMC run.
python sivism_lr.py --config LRwaveform.yml --baseline_sample SGLD_LR/parallel_SGLD_LRwaveform.pt
For the Bayesian multinomial logistic regression problem, we provide the demo results of SIVI-SM on MNIST dataset.
python sivism_mlr.py --config mnist.yml
Lastly, we provide the demo result of SIVI-SM on boston dataset.
python sivism_bnn.py --config boston.yml
If you find this code useful for your research, please consider citing
@inproceedings{
yu2023semiimplicit,
title={Semi-Implicit Variational Inference via Score Matching},
author={Longlin Yu and Cheng Zhang},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023}
}