What is Skeltorch?
Skeltorch is a light-weight framework that helps researchers to prototype faster using PyTorch. To do so, Skeltorch provides developers with a set of predefined pipelines to organize projects and train/test their models.
Skeltorch is an experiment-based framework. What that means is that every possible variation of your model will be represented by a different experiment. Every experiment is uniquely identified by its name and contains:
A set of immutable configuration parameters, specified during its creation.
A copy of the data object, also created during the creation of the experiment.
The checkpoints of the model associated with the experiment.
A set of TensorBoard files with a graphical evolution of the losses and other data that may be logged.
A textual log of the actions performed on the experiment.
Easy creation and loading of experiments.
Automatic restoration of interrupted training.
Readable JSON configuration files with the option to validate them using a schema.
Visual logging using TensorBoard.
Automatic logging using the native Python logging package.
Automatic handling of random seeds, specified during the creation of an experiment.
Easy implementation of custom pipelines.