luz 10.3.0

Lightweight framework for structuring arbitrary reproducible neural network learning procedures using PyTorch.

PyTorch code can easily become complex, unwieldy, and difficult to understand as a project develops. Luz aims to provide a common scaffold for PyTorch code in order to minimize boilerplate, maximize readability, and maintain the flexibility of PyTorch itself.

The basis of Luz is the Runner, an abstraction representing batch-wise processing of data over multiple epochs. Runner has predefined hooks to which code can be attached and a State which can be manipulated to define essentially arbitrary behavior. These hooks can be used to compose multiple Runners into a single algorithm, enabling dataset preprocessing, model testing, and other common tasks.

To further reduce boilerplate, the Learner abstraction is introduced as shorthand for the extremely common Preprocess-Train-Validate-Test algorithm. Simply inherit luz.Learner and define a handful of methods to completely customize your learning algorithm.

Two additional abstractions are provided for convenience: Scorers, which score (i.e. evaluate) a Learner according to some predefined procedure, and Tuners, which tune Learner hyperparameters. These abstractions provide a common interface which makes model selection a two-line process.