☰
Free download

Table Of Contents

Built with 🛠 MkDocs - Theme 🖤 Github.

Training a Model

Configure the network trainer, then train only with the prepared training subset.

Basic training

import deepnetts.net.train.opt.OptimizerType;

neuralNet.getTrainer()
        .setStopEpochs(300)
        .setLearningRate(0.001f)
        .setOptimizer(OptimizerType.ADAM)
        .setShuffle(true);

neuralNet.train(trainingSet);
Setting Purpose
setStopEpochs(...) Maximum number of passes over the training data
setLearningRate(...) Size of each parameter update
setOptimizer(...) Algorithm used to update model parameters
setShuffle(true) Changes sample order between epochs

These values are starting points, not universal defaults. A model that does not learn may need a different learning rate, architecture, preprocessing choice, or more training.

Mini-batch training

For a larger dataset, enable batches explicitly:

neuralNet.getTrainer()
        .setStopEpochs(50)
        .setLearningRate(0.001f)
        .setOptimizer(OptimizerType.SGD)
        .setBatchMode(true)
        .setBatchSize(64)
        .setShuffle(true);

neuralNet.train(trainingSet);

Train a CNN

CNNs use the same trainer API, but image batches consume more memory and each epoch usually takes longer than it does for a small tabular model:

network.getTrainer()
        .setStopEpochs(20)
        .setLearningRate(0.001f)
        .setOptimizer(OptimizerType.ADAM)
        .setBatchMode(true)
        .setBatchSize(8)
        .setShuffle(true);

network.train(trainingSet);

Start with a batch size that fits available memory. Compare training and held-out metrics across epochs; a successful smoke test confirms that the pipeline runs, but one epoch is not evidence that the CNN has converged.

Stop on an error target

Training can also stop when it reaches a configured error:

neuralNet.getTrainer()
        .setStopError(0.03f)
        .setStopEpochs(350)
        .setLearningRate(0.001f)
        .setOptimizer(OptimizerType.ADAM);

Keep an epoch limit even when using an error target so a run cannot continue indefinitely.

Common mistakes
  • Training with testSet invalidates the later evaluation.
  • Changing several settings at once makes experiments difficult to compare.
  • A lower training error does not guarantee better performance on unseen data.
  • Training without the same preprocessing used in evaluation creates inconsistent inputs.
  • Augmenting or otherwise changing testSet prevents a clean held-out evaluation.
See it in practice

One epoch is one complete pass through the training dataset.

A fixed seed makes initialization and data ordering reproducible enough to compare repeated development runs.

Was this helpful?