Online

Learners that implement the Online interface can be trained in batches. Learners of this type are great for when you either have a continuous stream of data or a dataset that is too large to fit into memory. In addition, partial training allows the model to evolve over time.

Partially Train

To partially train an Online learner pass it a training set to its partial() method:

public partial(Dataset $dataset) : void

$folds = $dataset->fold(3);

$estimator->train($folds[0]);

$estimator->partial($folds[1]);

$estimator->partial($folds[2]);

Note

Learner will continue to train as long as you are using the partial() method, however, calling train() on a trained or partially trained learner will reset it back to baseline first.

Streaming Training

For datasets that are too large to fit into memory all at once, you can stream the records from disk using an Extractor and train in batches with the chunked() dataset factory method. In the example below, we stream the records of a large NDJSON file from disk and partially train an MLP Regressor.

use Rubix\ML\Datasets\Labeled;
use Rubix\ML\Extractors\NDJSON;
use Rubix\ML\Regressors\MLPRegressor;
use Rubix\ML\NeuralNet\Layers\Dense;
use Rubix\ML\NeuralNet\Layers\Activation;
use Rubix\ML\NeuralNet\ActivationFunctions\SiLU;
use Rubix\ML\NeuralNet\Optimizers\Adam;
use Rubix\ML\NeuralNet\Optimizers\Schedulers\Constant;

$estimator = new MLPRegressor([
    new Dense(128),
    new Activation(new SiLU()),
    new Dense(64),
    new Activation(new SiLU()),
    new Dense(1),
], optimizer: new Adam(new Constant(0.001)));

$extractor = new NDJSON('too-large.jsonl');

foreach (Labeled::chunked($extractor, 8192) as $batch) {
    $estimator->partial($batch);
}

Because partial() warms up a learner on the first call, there is no need to call train() beforehand. Every batch must be of the same shape and feature order, and since each batch is validated on construction, the first batch will fail fast if a record is malformed.