Package io.github.kirstenali.deepj.training

Class Trainer

java.lang.Object

io.github.kirstenali.deepj.training.Trainer

public final class Trainer
extends Object

A small, reusable training loop wrapper.

This library supports different model/data shapes (e.g. supervised Tensor->Tensor models, and causal language models that operate on token ids). Rather than duplicating full trainers, Trainer delegates a single training step to a pluggable Trainer.StepFunction.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static interface	Trainer.StepFunction
static interface	Trainer.StepHook

Constructor Summary

Constructors

Constructor	Description
Trainer(Trainer.StepFunction stepFn)

Method Summary

Modifier and Type	Method	Description
TrainingResult	train(int maxSteps, int batchSize, int logEvery, double emaBeta, Double targetEmaLoss)	Train until maxSteps or until EMA loss goes below targetEmaLoss (if provided).
TrainingResult	train(int maxSteps, int batchSize, int logEvery, double emaBeta, Double targetEmaLoss, int releaseEverySteps)
TrainingResult	train(int maxSteps, int batchSize, int logEvery, double emaBeta, Double targetEmaLoss, int releaseEverySteps, Trainer.StepHook stepHook)	Train until maxSteps or until EMA loss goes below targetEmaLoss (if provided).
TrainingResult	train(int maxSteps, int batchSize, int logEvery, double emaBeta, Double targetEmaLoss, Trainer.StepHook stepHook)	Train until maxSteps or until EMA loss goes below targetEmaLoss (if provided).
double	trainStep(int batchSize)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

Trainer

public Trainer(Trainer.StepFunction stepFn)

Method Details

trainStep

public double trainStep(int batchSize)

train

public TrainingResult train(int maxSteps,
 int batchSize,
 int logEvery,
 double emaBeta,
 Double targetEmaLoss)

Train until maxSteps or until EMA loss goes below targetEmaLoss (if provided). Uses the default periodic backend release cadence.

train

public TrainingResult train(int maxSteps,
 int batchSize,
 int logEvery,
 double emaBeta,
 Double targetEmaLoss,
 int releaseEverySteps)

train

public TrainingResult train(int maxSteps,
 int batchSize,
 int logEvery,
 double emaBeta,
 Double targetEmaLoss,
 Trainer.StepHook stepHook)

Train until maxSteps or until EMA loss goes below targetEmaLoss (if provided). Uses the default periodic backend release cadence.

train

public TrainingResult train(int maxSteps,
 int batchSize,
 int logEvery,
 double emaBeta,
 Double targetEmaLoss,
 int releaseEverySteps,
 Trainer.StepHook stepHook)

Train until maxSteps or until EMA loss goes below targetEmaLoss (if provided). releaseEverySteps invalid input: '<'= 0 disables periodic release, but final release still runs.