Path Imputation Strategies for Signature Models of Irregularly Time Series

Reference

This repository contains code for the following paper:

@article{moor2020path,
  title={Path Imputation Strategies for Signature Models},
  author={Moor, Michael and Horn, Max and Bock, Christian and Borgwardt, Karsten and Rieck, Bastian},
  journal={arXiv preprint arXiv:2005.12359},
  year={2020},
}

Furthermore, this work was subsumed in this short paper which was accepted for presentation at the ICML 2020 workshop on the art of learning with missing values (Artemiss).

Environment

Please install the dependencies as indicated in the requirements.txt or pyproject.toml file:
> poetry install
> poetry shell

Note that if you alternatively use pipenv, the newer GPytorch versions ( >1.0.0 ) tend to overwrite the torch version. If this happens (e.g. with GPytorch 1.0.1), a hacky but working solution was to just >pip intall torch==1.2.0 after the pipenv was installed with >pipenv install --skip-lock

Setting up data:

The physionet 2012 dataset has to be downloaded with the following shell script:
>source data/physionet_2012/download.sh
The other datasets can be downloaded via:
>python3 src/datasets/download_uea_data.py and can then be found in data/Multivariate_ts

Hypersearch Commands

Below commands for training a model assume that a GPU is available, however CPU-only execution is also possible by setting device=cpu (note that GP-based models then also need model.parameters.output_device=cpu)

gpu scheduler

if you want to use a gpu scheduler, simply install this one via: > pip install simple_gpu_scheduler

generate hypersearch commands:

> python scripts/generate_hypersearch_commands.py

This script generates multiple command files, one for GP-based methods and one for the other imputed methods (usually requiring less memory). For instance, this one: scripts/commands/command_LSST_imputed_hypersearches.csv

actually run the hypersearch via gpu scheduler on the first 3 devices of your server:

> simple_gpu_scheduler --gpus 0,1,2 < command_LSST_imputed_hypersearches.csv

If there is a configuration problem with the virtual environment and the gpu scheduler, alternatively those python commands could be started manually or sequentially via
> source scripts/commands/command_LSST_imputed_hypersearches.csv

After having run a hyperparameter search, create repetitions:

> python scripts/generate_repetitions.py This script assumes that the results of the hyperparameter search are stored in experiments/hyperparameter_search Again, as with the hyperparameter search: > simple_gpu_scheduler --gpus 0,1,2 < command_LSST_imputed_repetitions.csv

Quick fitting, testing

Train a end-to-end, posterior moments GP-imputed Signature Model, specifying signature depth (truncation level) to 3

> python exp/train_model.py with model.GPSignatureModel dataset.Physionet2012 model.parameters.sampling_type=moments model.parameters.sig_depth=3
> python exp/train_model.py with model.GPSignatureModel dataset.Physionet2012 model.parameters.sampling_type=monte_carlo model.parameters.sig_depth=2
> python exp/train_model.py with model.GPGRUSignatureModel dataset.Physionet2012 model.parameters.sampling_type=moments model.parameters.sig_depth=2

For training one of the subsampled datasets (PenDigits, CharacterTrajectories, LSST), a subsampler has to be provided. Here, an example command for training a GP-PoM-RNN model on PenDigits using CPU:
>python exp/train_model.py with model.GPRNNModel dataset.PenDigits device=cpu model.parameters.output_device=cpu subsampler_name=MissingAtRandomSubsampler model.parameters.sampling_type=moments

Manually start one hyperparameter search: for the hypersearches, the models and datasets are defined (and extended with hyperparameter spaces) in /exp/hypersearch_configs.py

>python exp/hyperparameter_search.py -F exp_runs/SignatureModel with GP_mom_SignatureModel Physionet2012

To check to current parameter configurations (handled via sacred), use print_config.

For instance, to inspect all set parameters in one of the commands above, use:

> python exp/train_model.py print_config with model.GPSignatureModel dataset.Physionet2012 model.parameters.sampling_type=moments model.parameters.sig_depth=3

Paper configurations

The configurations used in the paper (repetition configs as determined by hyperparameter search), are accessible in the path repe_configs/train_model
Train a model with a stored config.json:
> python exp/train_model.py with path/to/config.json