slingpy

The slingpy python package provides starter code for structured, reproducible and maintainable machine learning projects. slingpy aims to be maximally extensible while maintaining sensible defaults. It is agnostic in terms of modelling backend (e.g., supporting scikit-learn, torch and tensorflow) and suitable for both production and research-grade machine learning projects.

slingpy contains utilities for standard model evaluation workflows, such as nested cross validation, model serialisation, dataset handling, managing high performance computing (HPC) interfaces such as slurm, and by default writes all experiment artefacts to disk.

Install

pip install slingpy

Use

A minimal slingpy project consists of a base application that defines your basic modelling workflow.

import slingpy as sp
from typing import AnyStr, Dict, List
from sklearn.linear_model import LogisticRegression


class MyApplication(sp.AbstractBaseApplication):
    def __init__(self, output_directory: AnyStr = "",
                 schedule_on_slurm: bool = False,
                 split_index_outer: int = 0,
                 split_index_inner: int = 0,
                 num_splits_outer: int = 5,
                 num_splits_inner: int = 5):
        super().__init__(
            output_directory=output_directory,
            schedule_on_slurm=schedule_on_slurm,
            split_index_outer=split_index_outer,
            split_index_inner=split_index_inner,
            num_splits_outer=num_splits_outer,
            num_splits_inner=num_splits_inner
        )

    def get_metrics(self, set_name: AnyStr) -> List[sp.AbstractMetric]:
        return [
            sp.metrics.AreaUnderTheCurve()
        ]

    def load_data(self) -> Dict[AnyStr, sp.AbstractDataSource]:
        data_source_x, data_source_y = sp.datasets.Iris.load_data(self.output_directory)

        stratifier = sp.StratifiedSplit()
        rest_indices, training_indices = stratifier.split(data_source_y, test_set_fraction=0.6,
                                                          split_index=self.split_index_inner)
        validation_indices, test_indices = stratifier.split(data_source_y.subset(rest_indices), test_set_fraction=0.5,
                                                            split_index=self.split_index_outer)

        return {
            "training_set_x": data_source_x.subset(training_indices),
            "training_set_y": data_source_y.subset(training_indices),
            "validation_set_x": data_source_x.subset(validation_indices),
            "validation_set_y": data_source_y.subset(validation_indices),
            "test_set_x": data_source_x.subset(test_indices),
            "test_set_y": data_source_y.subset(test_indices)
        }

    def train_model(self) -> sp.AbstractBaseModel:
        model = sp.SklearnModel(LogisticRegression())
        model.fit(self.datasets.training_set_x, self.datasets.training_set_y)
        return model


if __name__ == "__main__":
    app = sp.instantiate_from_command_line(MyApplication)
    app._run()

Your new app can then be executed locally from the command line using:

python /project_path/my_application.py

slingpy also enables execution of your project on a remote HPC cluster, e.g. via slurm, by using:

python /project_path/my_application.py --do_schedule_on_slurm

Application parameters are automatically parsed from the command line, e.g.:

python /project_path/my_application.py --output_directory=/path/to/output/dir

Development

The slingpy codebase is formatted with Black and Import Sort to ensure consistant code formatting. These are run through pre-commit. If making code changes to slingpy, install the development dependencies and pre-commit git hook with:

pip install -r requirements-dev.txt
pre-commit install

License

License

Authors

Patrick Schwab, GlaxoSmithKline plc
Arash Mehrjou, GlaxoSmithKline plc
Andrew Jesson, University of Oxford
Ashkan Soleymani, MIT

Acknowledgements

PS and AM are employees and shareholders of GlaxoSmithKline plc.