Stochastic Boundary Ordinary Differential Equation (STRODE)

This is the authors' official PyTorch implementation for STRODE. This repo contains code for experiments in the ICML 2021 paper 'STRODE: Stochastic Boundary Ordinary Differential Equation'.

Project Description

The precise timing of each item in real-world data streams can carry important information about the underlying dynamics. However, most algorithms for time-series modeling, e.g., latent ordinary differential equation (ODE), fail to infer timings (timestamps) directly from visual or audio inputs, and still require a large amount of training data with timing annotations. Inspired by neuroscience perspectives on time perception, we generalize neural ODE in handling a special type of boundary value problem with random boundary times and propose a probabilistic ordinary differential equation (ODE), called Stochastic boundaRy ODE (STRODE). Extensive experiments show that STRODE successfully learns both the timings (timestamps) and the dynamics of time series data without requiring any timing annotations during training.

Method Overview

How to run

Environment

• Python 3.7
• PyTorch 1.7.1

Running experiments

Toy Dataset

For training and evaluation on the toy dataset, run the following script:

cd toy
python run.py --dataset exp

where --dataset specifies the dataset used in the experiments. For more specifications of the experiments, see details in config.py.

Rotating MNIST Thumbnail

For training and evaluation on the rotating MNIST thumbnail, run the following script:

cd rotatingMNIST
python run.py --model strode --dataset exp

where --model and --dataset specify the model and the dataset used in the experiments. For more specifications of the experiments, see details in config.py.

Data

You can directly set isLoad=False(default) in config.py and run the script above and new datasets will be automatically generated. ./data/MNIST/ contains the processed MNIST dataset. You can refer to torchvision.datasets.MNIST for details.

Citation

If you use STRODE or this codebase in your own work, please cite our paper:

@inproceedings{huang2021strode,
  title={STRODE: Stochastic Boundary Ordinary Differential Equation},
  author={Huang, Hengguan and Liu, Hongfu and Wang, Hao and Xiao, Chang and Wang, Ye},
  booktitle={International Conference on Machine Learning},
  pages={4435--4445},
  year={2021},
  organization={PMLR}
}

License

MIT