Skip to content
Code for reproducing results in the sliced score matching paper
Python
Branch: master
Clone or download
Latest commit a9bb852 Aug 20, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
configs release all Jun 25, 2019
evaluations release all Jun 25, 2019
losses release all Jun 25, 2019
models release all Jun 25, 2019
runners release all Jun 25, 2019
.gitignore release all Jun 25, 2019
204_supplement.pdf add supplementary material Jun 27, 2019
LICENSE Create LICENSE Jun 28, 2019
README.md Update README.md Aug 20, 2019
main.py release all Jun 25, 2019

README.md

Sliced Score Matching: A Scalable Approach to Density and Score Estimation

This repo contains a PyTorch implementation for the paper Sliced Score Matching: A Scalable Approach to Density and Score Estimation, UAI 2019. Sliced score matching is a scalable variant of score matching that can be used to train unnormalized statistical models or estimating the score (derivatives of the log-density function) of data.

Dependencies

The following are packages needed for running this repo.

  • PyTorch==1.0.1
  • TensorFlow==1.12.0
  • tqdm
  • tensorboardX
  • Scipy
  • PyYAML

Running the experiments

python main.py --runner [runner name] --config [config file]

Here runner name is one of the following:

  • DKEFRunner. This corresponds to experiments on deep kernel exponential families.
  • NICERunner. This corresponds to the sanity check experiment of training a NICE model.
  • VAERunner. Experiments on VAEs.
  • WAERunner. Experiments on Wasserstein Auto-Encoders (WAEs).

and config file is the directory of some YAML file in configs/.

For example, if you want to train an implicit VAE of latent size 8 on MNIST with Sliced Score Matching, just run

python main.py --runner VAERunner --config vae/mnist_ssm_8.yml

References

If you find the idea or code useful for your research, please consider citing our paper:

@article{song2019sliced,
  title={Sliced Score Matching: A Scalable Approach to Density and Score Estimation},
  author={Song, Yang and Garg, Sahaj and Shi, Jiaxin and Ermon, Stefano},
  journal={arXiv preprint arXiv:1905.07088},
  year={2019}
}
You can’t perform that action at this time.