Trainers

FENN provides a modular trainer architecture for training deep learning models. The base Trainer class is abstract and provides the core functionality for managing the training cycle, including checkpointing, early stopping, validation, and prediction.

Available Trainers

• Base Trainer: Abstract base class with core training functionality
• ClassificationTrainer: For binary, multi-class, and multi-label classification tasks
• RegressionTrainer: For regression tasks
• LoRATrainer: For parameter-efficient fine-tuning with LoRA (Low-Rank Adaptation)

Common Features

All trainers share common functionality:

Training

• fit() method for training loops
• Optional validation every N epochs
• Automatic loss calculation per epoch

Prediction

• predict() method for inference
• Support for DataLoaders and batch tensors

Checkpointing

• Save and resume training from checkpoints
• Automatic best model saving
• Checkpoint loading at specific epochs

Early Stopping

• Automatic training termination when model stops improving
• Configurable patience counter
• Tracks training or validation loss depending on configuration

State Management

• Training state tracking through epochs
• Model and optimizer state persistence
• Metrics logging (accuracy, loss, F1, etc.)

Shared Functionality

See the following pages for shared functionality across all trainers:

• Checkpoint Management: Saving, loading, and resuming from checkpoints
• Advanced Features: Early stopping, validation metrics, training state management