Explanation

In this section, specific aspects of Lightning-Boost are explained in depth.

Models vs. Systems

In Lightning-Boost, there is a differentiation between a model and a system:

• A model is a parametrized function. It has a well-defined signature, taking one or more tensors as input and producing one or more tensors as output.
• A system is an instance that embeds one or more models, and manages the entire training process, given data through a datamodule.

This clear logical separation has several advantages in terms of modularity and code structure:

• A model can be exchanged by another model, if their signature is the same.
• Multiple models can easily be combined in a multi-model system, e.g., for a multi-task pretraining-finetuning setup.
• In general, code becomes clearer and easier to maintain.

Task-oriented Data Processing

In more complex scenarios, a model might have to solve a variety of tasks at once.

Lightning-Boost allows tackling several tasks in one system through task-oriented data processing. This is enabled by the data structures used for input, target and prediction data, which are dictionaries. Task-specific data is then logically assigned to a task by setting its key accordingly, making the code easy to understand. Furthermore, it allows the automatic processing of target and prediction data through loss functions and metrics within a specific task.