https://github.com/qsh-zh/pis

[ICLR 2022] Path integral sampler
https://github.com/qsh-zh/pis

iclr2022 sampling

Last synced: 4 months ago
JSON representation

[ICLR 2022] Path integral sampler

• Host: GitHub
• URL: https://github.com/qsh-zh/pis
• Owner: qsh-zh
• Created: 2022-02-02T16:31:30.000Z (over 2 years ago)
• Default Branch: master
• Last Pushed: 2023-08-21T10:14:29.000Z (10 months ago)
• Last Synced: 2024-01-16T10:43:27.641Z (5 months ago)
• Topics: iclr2022, sampling
• Language: Python
• Homepage: https://arxiv.org/abs/2111.15141
• Size: 407 KB
• Stars: 38
• Watchers: 2
• Forks: 8
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Lists

• awesome-stars - qsh-zh/pis - [ICLR 2022] Path integral sampler (Python)

README

        
# [Path Integral Sampler: a stochastic control approach for sampling](https://arxiv.org/abs/2111.15141)

[![PIS Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1NOHGt2iHoETgbSh4z7dLep95-zN8FNKL?usp=sharing)

![PIS](asserts/pis.png)

## Setup

The repo heavily depends on [jam](https://github.com/qsh-zh/jam), a versatile toolbox developed by [Qsh.zh](https://github.com/qsh-zh) and [jam-hlt](https://github.com/qsh-zh/jam), a decent deep leanring project template. [⭐️](https://github.com/qsh-zh/jam) if you like them.

*[poetry](https://python-poetry.org/)* (**Recommended**)

```shell

curl -fsS -o /tmp/get-poetry.py https://raw.githubusercontent.com/sdispater/poetry/master/get-poetry.py

python3 /tmp/get-poetry.py -y --no-modify-path

export PATH=$HOME/.poetry/bin:$PATH

poetry shell

poetry install

```

*pip*

```shell

pip install .

```

## Reproduce

```

python run.py experiment=ou.yaml logger=wandb

```

See the [folder](configs/experiment) for more experiment configs.

There are some [results](https://wandb.ai/qinsheng/pub_pis?workspace=user-qinsheng) reproduced by the repo.

## Reference

```tex

@inproceedings{zhang2021path,

  author    = {Qinsheng Zhang and Yongxin Chen},

  title     = {Path Integral Sampler: a stochastic control approach for sampling},

  booktitle = {International Conference on Learning Representations},

  year      = {2022}

}

```

## MICS:

- [sde-sampler](https://gitlab.com/qsh.zh/sde-sampler/-/tree/rings) Uncleaned code used for experiments in paper.

- SDE parameters `dt,g` are modified due to [the issue](https://github.com/google-research/torchsde/issues/109).