The Adapter Pattern

The Problem

Different ML components may expect different data formats. Consider a scenario where:

• Dataset returns dictionaries of tensors
• Model expects tensors
• Loss function needs specific argument order
• Metrics need different format than loss

Traditionally, you'd implement a pipeline specific to this scenario. This tightly couples components, making reuse and experimentation difficult.

The Solution: Adapters

Data flow through Lighter's System. Adapters bridge components with incompatible interfaces.

In software engineering, the adapter pattern allows incompatible interfaces to work together. Lighter uses adapters to handle variability in data formats.

Lighter's Adapter Types

Adapter	Purpose	When to Use
BatchAdapter	Extract data from batches	Different dataset formats
CriterionAdapter	Format loss inputs	Custom loss functions
MetricsAdapter	Format metric inputs	Third-party metrics
LoggingAdapter	Transform before logging	Visualization needs

Example: Task-Agnostic Configuration

adapters:
  train:
    criterion:
      _target_: lighter.adapters.CriterionAdapter
      pred_transforms:   # Apply sigmoid before loss
        _target_: torch.sigmoid
      pred_argument: 0   # Map pred to first argument
      target_argument: 1 # Map target to second argument

This enables any task—classification, segmentation, self-supervised learning—without framework modifications.

Under the Hood

Adapters are invoked in this order during training:

1. BatchAdapter - Extract input/target from batch
2. Forward pass - Model processes input
3. CriterionAdapter - Format for loss computation
4. MetricsAdapter - Format for metric computation
5. LoggingAdapter - Transform for visualization

Practical Usage

For detailed adapter configuration and examples, see:

• Adapters How-To Guide - Complete usage guide
• Metrics Guide - Using MetricsAdapter
• Writers Guide - Using LoggingAdapter