Link Search Menu Expand Document
SNN

Training 🏋

This part is about how to train a StraightforwardNeuralNetwork. There are 2 ways to do this, synchronous with the train method or async with the methods startTrainingAsync and stopTrainingAsync.

Declaration

void train(Data& data, Wait wait, int batchSize = 1, int evaluationFrequency = 1);

Arguments

  • data: The dataset used to train the neural network
  • wait: Stop training condition
  • batchsize: The size of batch
  • evaluationFrequency: The number of epochs between each evaluation of the neural network on the testing set

Example

The synchronous version

StraightforwardNeuralNetwork neuralNetwork({Input(4), FullyConnected(20), FullyConnected(3)});
neuralNetwork.train(data, 20_s || 0.9_acc); // train neural network on data until 90% accuracy or 20s

The asynchronous version

StraightforwardNeuralNetwork neuralNetwork({Input(4), FullyConnected(20), FullyConnected(3)});
neuralNetwork.startTrainingAsync(data); // start training on a new parallel thread
neuralNetwork.waitFor(20_s || 0.9_acc); // wait until the network reach 90% accuracy or 20s
neuralNetwork.stopTrainingAsync(); // stop training, evaluate the network and delete the thread

It is better to use the train method unless you do something else during the training such as displaying a real-time graph.

List of all conditions for stop the train or WaitFor methods:

  • _ep: The number of epochs to reach. A epoch is one learn about the whole dataset.
  • _acc: The minimum accuracy (between 0 and 1) to reach.
  • _mae: The maximum MAE (mean absolute error) to reach.
  • _ms: Time before the learning stops in milliseconds.
  • _s: Time before the learning stops in seconds.
  • _min: Time before the learning stops in minutes.

You can use one or several conditions with the operator && or || are used but you can’t mix the both operator.

Batch size and evaluation frequency

You can add bach size and change the evaluation frequency. The default value for batch size and evaluation frequency is 1. You can set the evaluation frequency to 0 to never evaluate the network during learning, therefore the stop condition must be a duration.

neuralNetwork.train(data, 1.5_mae, 32, 2); // set the bach size to 32 and the evaluation frequency to 2

Tip

You can set the verbose enum to minimal for display the accuracy and the epoch number during training.