daixiongming/mirror-stein-samplers

Sampling with Mirrored Stein Operators

Sampling with Mirrored Stein Operators

Code for reproducing the experimental results in https://arxiv.org/abs/2106.12506

---

Update (2023/2/13):

We added a new example of sampling from a uniform distribution in a $[-1, 1]^d$ rectangle in uniform.ipynb.

---

Requirements

• R >= 4.0.4 (only needed for the non-toy post-selection inference experiments)
• python >= 3.8

rpy2 >= 3.4.4
tensorflow >= 2.4.0
tensorflow_probability >= 0.12
numpy >= 1.19.5
scipy >= 1.6.3
matplotlib >= 3.4.1
pandas >= 1.2.4
seaborn >= 0.11.1
scikit-learn >= 0.24.2
tqdm
absl-py

Experiments

Approximation quality on the simplex

• Sparse Dirichlet: dirichlet.ipynb
• Quadratic: quad.ipynb

Confidence intervals for post-selection inference

Note: The R_HOME variable must be set correctly before running the scripts.

• Simulation
  • 2D example: selective_inf.ipynb
  • Nominal coverage vs. actual coverage: coverage.py
  • Coverage vs. number of samples: coverage_wrt_k.py
  • Plotting: plot_coverage.ipynb
• HIV drug resistance: hiv.ipynb

Large-scale posterior inference with non-Euclidean geometry

• Run script: lr.py
• Plotting: plot_lr.ipynb

Citation

If you find this repository useful, please cite:

@article{shi2021sampling,
  title={Sampling with Mirrored {S}tein Operators}, 
  author={Jiaxin Shi and Chang Liu and Lester Mackey},
  journal={International Conference on Learning Representations},
  year={2022}
}