Logistic regression

MNIST. Binary classification: 0 vs 1

MNIST case with RubixML

In real projects, people rarely implement logistic regression from scratch. It is much more convenient to use a machine learning library. The same 0-vs-1 MNIST case with RubixML becomes shorter and closer to production-style code:

Example of use


                
<?php

use Rubix\ML\Classifiers\LogisticRegression;
use Rubix\ML\Datasets\Labeled;
use Rubix\ML\Datasets\Unlabeled;
use Rubix\ML\Extractors\CSV;

try {
    function mnistRows(string $file): iterable {
        foreach (new CSV($file, false) as $row) {
            if (!isset($row[0])) {
                continue;
            }

            $label = (int) $row[0];

            if ($label !== 0 && $label !== 1) {
                continue;
            }

            $pixels = array_map(static fn ($value): float => (float) $value, array_slice($row, 1));

            yield array_merge($pixels, [$label === 1 ? 'one' : 'zero']);
        }
    }

    $trainRows = mnistRows('train.csv');
    $testRows = mnistRows('test.csv');

    $dataset = Labeled::fromIterator($trainRows);
    $testDataset = Labeled::fromIterator($testRows);
} catch (Exception $e) {
    echo '<div class="alert alert-danger" role="alert">' . htmlspecialchars($e->getMessage(), ENT_QUOTES, 'UTF-8') . '</div>';
    exit;
}

echo 'Train samples handled: ' . number_format($dataset->numSamples()) . PHP_EOL;
echo 'Test samples handled: ' . number_format($testDataset->numSamples()) . PHP_EOL . PHP_EOL;

$model = new LogisticRegression(epochs: 5);
$model->train($dataset);

echo 'Number of epochs: ' . $model->params()['epochs'] . PHP_EOL . PHP_EOL;

$correct = 0;

foreach ($testDataset->samples() as $i => $x) {
    $prediction = $model->predict(new Unlabeled([$x]))[0];

    if ($prediction === $testDataset->labels()[$i]) {
        $correct++;
    }
}

$accuracy = $correct / $testDataset->numSamples();
echo 'Accuracy: ' . round($accuracy * 100, 2) . '%';

Sample digit of 0

Probability of digit 0: 1

Predicted digit: 0

Sample of digit 1

Probability of digit 1: 0.978

Predicted digit: 1

Result: Memory: 0 Mb Time running: < 0.001 sec.


                Train samples handled: 12,666
Test samples handled: 2,116

Number of epochs: 5

Accuracy: 99.95%