Combustion

Combustion is a collection of layers/models/helper functions for deep learning.

Combustion is designed with PyTorch in mind, with some emphasis on the following 3rd party libraries:

• PyTorch-Lightning, a high level API for model training
• Hydra, a library that enables YAML based configuration of hyperparameters

Installation

Install combustion core with pip using

pip install .

Optional dependencies are provided via the following extras:

• hdf5 - HDF5 serialization (combustion.data)
• vision - Vision helpers (combustion.vision)
• points - Point cloud manipulation (combustion.points)
• macs - For counting multiply accumulate operations

Development dependencies can be installed with

pip install combustion[dev] && pip install combustion[points]

Development

Multiple make recipes are provided to aid in development:

• ci-test - Runs tests that will be run by CircleCI
• ci-quality - Runs quality checks that will be run by CircleCI
• docs - Builds documentation
• pre-commit - Installs a pre-commit hook to run code quality tests
• quality - Runs code quality tests
• style - Automatically formats code (using black and autopep8)
• tag-version - Adds a git tag based on version.txt.
• test - Runs all tests, including slow model tests
• test-% - Runs specific tests by pattern match (via pytest -k flag)
• test-pdb-% - Runs specific tests with debugging on failure (via pytest --pdb flag)

References

• PyTorch Lightning
• Hydra