Implement FeatureUnion 🚀 (#382)

omysurya · May 14, 2019 · b500f0b · b500f0b
1 parent ff118eb
commit b500f0b
Show file tree

Hide file tree

Showing 6 changed files with 235 additions and 34 deletions.
diff --git a/README.md b/README.md
@@ -91,6 +91,7 @@ Public datasets are available in a separate repository [php-ai/php-ml-datasets](
     * Regression
 * Workflow
     * [Pipeline](http://php-ml.readthedocs.io/en/latest/machine-learning/workflow/pipeline)
+    * FeatureUnion
 * Neural Network
     * [Multilayer Perceptron Classifier](http://php-ml.readthedocs.io/en/latest/machine-learning/neural-network/multilayer-perceptron-classifier/)
 * Cross Validation
@@ -103,6 +104,9 @@ Public datasets are available in a separate repository [php-ai/php-ml-datasets](
     * [Normalization](http://php-ml.readthedocs.io/en/latest/machine-learning/preprocessing/normalization/)
     * [Imputation missing values](http://php-ml.readthedocs.io/en/latest/machine-learning/preprocessing/imputation-missing-values/)
     * LabelEncoder
+    * LambdaTransformer
+    * NumberConverter
+    * ColumnFilter
 * Feature Extraction
     * [Token Count Vectorizer](http://php-ml.readthedocs.io/en/latest/machine-learning/feature-extraction/token-count-vectorizer/)
         * NGramTokenizer

diff --git a/src/FeatureUnion.php b/src/FeatureUnion.php
@@ -0,0 +1,72 @@
+<?php
+
+declare(strict_types=1);
+
+namespace Phpml;
+
+use Phpml\Exception\InvalidArgumentException;
+
+final class FeatureUnion implements Transformer
+{
+    /**
+     * @var Pipeline[]
+     */
+    private $pipelines = [];
+
+    /**
+     * @var Pipeline[]
+     */
+    public function __construct(array $pipelines)
+    {
+        if ($pipelines === []) {
+            throw new InvalidArgumentException('At least one pipeline is required');
+        }
+
+        $this->pipelines = array_map(static function (Pipeline $pipeline): Pipeline {
+            return $pipeline;
+        }, $pipelines);
+    }
+
+    public function fit(array $samples, ?array $targets = null): void
+    {
+        $originSamples = $samples;
+        foreach ($this->pipelines as $pipeline) {
+            foreach ($pipeline->getTransformers() as $transformer) {
+                $transformer->fit($samples, $targets);
+                $transformer->transform($samples, $targets);
+            }
+            $samples = $originSamples;
+        }
+    }
+
+    public function transform(array &$samples, ?array &$targets = null): void
+    {
+        $this->transformSamples($samples, $targets);
+    }
+
+    public function fitAndTransform(array &$samples, ?array &$targets = null): void
+    {
+        $this->transformSamples($samples, $targets, true);
+    }
+
+    private function transformSamples(array &$samples, ?array &$targets = null, bool $fit = false): void
+    {
+        $union = [];
+        $originSamples = $samples;
+        foreach ($this->pipelines as $pipeline) {
+            foreach ($pipeline->getTransformers() as $transformer) {
+                if ($fit) {
+                    $transformer->fit($samples, $targets);
+                }
+                $transformer->transform($samples, $targets);
+            }
+
+            foreach ($samples as $index => $sample) {
+                $union[$index] = array_merge($union[$index] ?? [], is_array($sample) ? $sample : [$sample]);
+            }
+            $samples = $originSamples;
+        }
+
+        $samples = $union;
+    }
+}
diff --git a/src/Metric/Regression.php b/src/Metric/Regression.php
@@ -28,7 +28,7 @@ public static function meanSquaredLogarithmicError(array $targets, array $predic
 
         $errors = [];
         foreach ($targets as $index => $target) {
-            $errors[] = (log(1 + $target) - log(1 + $predictions[$index])) ** 2;
+            $errors[] = log((1 + $target) / (1 + $predictions[$index])) ** 2;
         }
 
         return Mean::arithmetic($errors);

diff --git a/src/Pipeline.php b/src/Pipeline.php
@@ -4,37 +4,28 @@
 
 namespace Phpml;
 
-class Pipeline implements Estimator
+use Phpml\Exception\InvalidOperationException;
+
+class Pipeline implements Estimator, Transformer
 {
     /**
      * @var Transformer[]
      */
     private $transformers = [];
 
     /**
-     * @var Estimator
+     * @var Estimator|null
      */
     private $estimator;
 
     /**
      * @param Transformer[] $transformers
      */
-    public function __construct(array $transformers, Estimator $estimator)
-    {
-        foreach ($transformers as $transformer) {
-            $this->addTransformer($transformer);
-        }
-
-        $this->estimator = $estimator;
-    }
-
-    public function addTransformer(Transformer $transformer): void
-    {
-        $this->transformers[] = $transformer;
-    }
-
-    public function setEstimator(Estimator $estimator): void
+    public function __construct(array $transformers, ?Estimator $estimator = null)
     {
+        $this->transformers = array_map(static function (Transformer $transformer): Transformer {
+            return $transformer;
+        }, $transformers);
         $this->estimator = $estimator;
     }
 
@@ -46,16 +37,20 @@ public function getTransformers(): array
         return $this->transformers;
     }
 
-    public function getEstimator(): Estimator
+    public function getEstimator(): ?Estimator
     {
         return $this->estimator;
     }
 
     public function train(array $samples, array $targets): void
     {
+        if ($this->estimator === null) {
+            throw new InvalidOperationException('Pipeline without estimator can\'t use train method');
+        }
+
         foreach ($this->transformers as $transformer) {
             $transformer->fit($samples, $targets);
-            $transformer->transform($samples);
+            $transformer->transform($samples, $targets);
         }
 
         $this->estimator->train($samples, $targets);
@@ -66,15 +61,27 @@ public function train(array $samples, array $targets): void
      */
     public function predict(array $samples)
     {
-        $this->transformSamples($samples);
+        if ($this->estimator === null) {
+            throw new InvalidOperationException('Pipeline without estimator can\'t use predict method');
+        }
+
+        $this->transform($samples);
 
         return $this->estimator->predict($samples);
     }
 
-    private function transformSamples(array &$samples): void
+    public function fit(array $samples, ?array $targets = null): void
+    {
+        foreach ($this->transformers as $transformer) {
+            $transformer->fit($samples, $targets);
+            $transformer->transform($samples, $targets);
+        }
+    }
+
+    public function transform(array &$samples, ?array &$targets = null): void
     {
         foreach ($this->transformers as $transformer) {
-            $transformer->transform($samples);
+            $transformer->transform($samples, $targets);
         }
     }
 }
diff --git a/tests/FeatureUnionTest.php b/tests/FeatureUnionTest.php
@@ -0,0 +1,105 @@
+<?php
+
+declare(strict_types=1);
+
+namespace Phpml\Tests;
+
+use Phpml\Exception\InvalidArgumentException;
+use Phpml\FeatureUnion;
+use Phpml\Pipeline;
+use Phpml\Preprocessing\ColumnFilter;
+use Phpml\Preprocessing\Imputer;
+use Phpml\Preprocessing\Imputer\Strategy\MeanStrategy;
+use Phpml\Preprocessing\LabelEncoder;
+use Phpml\Preprocessing\LambdaTransformer;
+use Phpml\Preprocessing\NumberConverter;
+use PHPUnit\Framework\TestCase;
+
+final class FeatureUnionTest extends TestCase
+{
+    public function testFitAndTransform(): void
+    {
+        $columns = ['age', 'income', 'sex'];
+        $samples = [
+            ['23', '100000', 'male'],
+            ['23', '200000', 'female'],
+            ['43', '150000', 'female'],
+            ['33', 'n/a', 'male'],
+        ];
+        $targets = ['1', '2', '1', '3'];
+
+        $union = new FeatureUnion([
+            new Pipeline([
+                new ColumnFilter($columns, ['sex']),
+                new LambdaTransformer(function (array $sample) {
+                    return $sample[0];
+                }),
+                new LabelEncoder(),
+            ]),
+            new Pipeline([
+                new ColumnFilter($columns, ['age', 'income']),
+                new NumberConverter(true),
+                new Imputer(null, new MeanStrategy(), Imputer::AXIS_COLUMN),
+            ]),
+        ]);
+
+        $union->fitAndTransform($samples, $targets);
+
+        self::assertEquals([
+            [0, 23.0, 100000.0],
+            [1, 23.0, 200000.0],
+            [1, 43.0, 150000.0],
+            [0, 33.0, 150000.0],
+        ], $samples);
+        self::assertEquals([1, 2, 1, 3], $targets);
+    }
+
+    public function testFitAndTransformSeparate(): void
+    {
+        $columns = ['age', 'income', 'sex'];
+        $trainSamples = [
+            ['23', '100000', 'male'],
+            ['23', '200000', 'female'],
+            ['43', '150000', 'female'],
+            ['33', 'n/a', 'male'],
+        ];
+        $testSamples = [
+            ['43', '500000', 'female'],
+            ['13', 'n/a', 'male'],
+            ['53', 'n/a', 'male'],
+            ['43', 'n/a', 'female'],
+        ];
+
+        $union = new FeatureUnion([
+            new Pipeline([
+                new ColumnFilter($columns, ['sex']),
+                new LambdaTransformer(function (array $sample) {
+                    return $sample[0];
+                }),
+                new LabelEncoder(),
+            ]),
+            new Pipeline([
+                new ColumnFilter($columns, ['age', 'income']),
+                new NumberConverter(),
+                new Imputer(null, new MeanStrategy(), Imputer::AXIS_COLUMN),
+            ]),
+        ]);
+
+        $union->fit($trainSamples);
+        $union->transform($testSamples);
+
+        self::assertEquals([
+            [1, 43.0, 500000.0],
+            [0, 13.0, 150000.0],
+            [0, 53.0, 150000.0],
+            [1, 43.0, 150000.0],
+        ], $testSamples);
+    }
+
+    public function testNotAllowForEmptyPipelines(): void
+    {
+        $this->expectException(InvalidArgumentException::class);
+
+        new FeatureUnion([]);
+    }
+}
diff --git a/tests/PipelineTest.php b/tests/PipelineTest.php
@@ -11,9 +11,9 @@
 use Phpml\ModelManager;
 use Phpml\Pipeline;
 use Phpml\Preprocessing\Imputer;
+use Phpml\Preprocessing\Imputer\Strategy\MeanStrategy;
 use Phpml\Preprocessing\Imputer\Strategy\MostFrequentStrategy;
 use Phpml\Preprocessing\Normalizer;
-use Phpml\Regression\SVR;
 use Phpml\Tokenization\WordTokenizer;
 use PHPUnit\Framework\TestCase;
 
@@ -32,16 +32,6 @@ public function testPipelineConstruction(): void
         self::assertEquals($estimator, $pipeline->getEstimator());
     }
 
-    public function testPipelineEstimatorSetter(): void
-    {
-        $pipeline = new Pipeline([new TfIdfTransformer()], new SVC());
-
-        $estimator = new SVR();
-        $pipeline->setEstimator($estimator);
-
-        self::assertEquals($estimator, $pipeline->getEstimator());
-    }
-
     public function testPipelineWorkflow(): void
     {
         $transformers = [
@@ -119,6 +109,29 @@ public function testPipelineTransformersWithTargets(): void
         self::assertEquals(['b'], $pipeline->predict([[1, 3, 5]]));
     }
 
+    public function testPipelineAsTransformer(): void
+    {
+        $pipeline = new Pipeline([
+            new Imputer(null, new MeanStrategy()),
+        ]);
+
+        $trainSamples = [
+            [10, 20, 30],
+            [20, 30, 40],
+            [30, 40, 50],
+        ];
+
+        $pipeline->fit($trainSamples);
+
+        $testSamples = [
+            [null, null, null],
+        ];
+
+        $pipeline->transform($testSamples);
+
+        self::assertEquals([[20.0, 30.0, 40.0]], $testSamples);
+    }
+
     public function testSaveAndRestore(): void
     {
         $pipeline = new Pipeline([