Trainer

The trainer is the concrete engine that fits a model to a dataset. It manages the training, validation, and testing lifecycle: it iterates the training data in epochs, periodically evaluates on the validation split, computes metrics, and drives checkpointing and reporting through callbacks.

Usage

A trainer is built from a configuration and run with execute; see the usage guide for a complete script.

from icegraph.trainer import Trainer

with Trainer.from_yaml(config_path) as trainer:
    trainer.execute()

Configuration

The trainer configuration is an engine configuration (services, policy, and components) plus a few run-level keys:

Option	Description	Type	Default
outdir	Directory where checkpoints, logs, and plots are written.	path	required
max_epochs	Maximum number of training epochs.	int	required
val_interval	Number of epochs between validation passes.	int	required

It uses the state, record, data, metrics, and decode services, a policy, and the full set of components (model, transformer, normalizer, optimizer, loss).

How it works

Each epoch runs the training split, updating the model through the optimizer against the loss. Every val_interval epochs the trainer runs the validation split, where it computes metrics and serves predictions to callbacks but performs no weight updates. A final test pass evaluates the trained model. The run is reproducible through the state service’s seed, and can be scaled across ranks with the Distributed wrapper.

• Trainer Callbacks
  • Registering
  • Built-in callbacks
  • Writing a callback