1
|
|
|
<?php |
2
|
|
|
|
3
|
|
|
declare(strict_types=1); |
4
|
|
|
|
5
|
|
|
namespace Phpml\NeuralNetwork\Network; |
6
|
|
|
|
7
|
|
|
use Phpml\Estimator; |
8
|
|
|
use Phpml\Exception\InvalidArgumentException; |
9
|
|
|
use Phpml\Helper\Predictable; |
10
|
|
|
use Phpml\IncrementalEstimator; |
11
|
|
|
use Phpml\NeuralNetwork\ActivationFunction; |
12
|
|
|
use Phpml\NeuralNetwork\Layer; |
13
|
|
|
use Phpml\NeuralNetwork\Node\Bias; |
14
|
|
|
use Phpml\NeuralNetwork\Node\Input; |
15
|
|
|
use Phpml\NeuralNetwork\Node\Neuron; |
16
|
|
|
use Phpml\NeuralNetwork\Node\Neuron\Synapse; |
17
|
|
|
use Phpml\NeuralNetwork\Training\Backpropagation; |
18
|
|
|
|
19
|
|
|
abstract class MultilayerPerceptron extends LayeredNetwork implements Estimator, IncrementalEstimator |
20
|
|
|
{ |
21
|
|
|
use Predictable; |
22
|
|
|
|
23
|
|
|
/** |
24
|
|
|
* @var array |
25
|
|
|
*/ |
26
|
|
|
protected $classes = []; |
27
|
|
|
|
28
|
|
|
/** |
29
|
|
|
* @var ActivationFunction |
30
|
|
|
*/ |
31
|
|
|
protected $activationFunction; |
32
|
|
|
|
33
|
|
|
/** |
34
|
|
|
* @var Backpropagation |
35
|
|
|
*/ |
36
|
|
|
protected $backpropagation = null; |
37
|
|
|
|
38
|
|
|
/** |
39
|
|
|
* @var int |
40
|
|
|
*/ |
41
|
|
|
private $inputLayerFeatures; |
42
|
|
|
|
43
|
|
|
/** |
44
|
|
|
* @var array |
45
|
|
|
*/ |
46
|
|
|
private $hiddenLayers = []; |
47
|
|
|
|
48
|
|
|
/** |
49
|
|
|
* @var float |
50
|
|
|
*/ |
51
|
|
|
private $learningRate; |
52
|
|
|
|
53
|
|
|
/** |
54
|
|
|
* @throws InvalidArgumentException |
55
|
|
|
*/ |
56
|
|
|
public function __construct(int $inputLayerFeatures, array $hiddenLayers, array $classes, int $iterations = 10000, ?ActivationFunction $activationFunction = null, float $learningRate = 1) |
57
|
|
|
{ |
58
|
|
|
if (empty($hiddenLayers)) { |
59
|
|
|
throw InvalidArgumentException::invalidLayersNumber(); |
60
|
|
|
} |
61
|
|
|
|
62
|
|
|
if (count($classes) < 2) { |
63
|
|
|
throw InvalidArgumentException::invalidClassesNumber(); |
64
|
|
|
} |
65
|
|
|
|
66
|
|
|
$this->classes = array_values($classes); |
67
|
|
|
$this->iterations = $iterations; |
|
|
|
|
68
|
|
|
$this->inputLayerFeatures = $inputLayerFeatures; |
69
|
|
|
$this->hiddenLayers = $hiddenLayers; |
70
|
|
|
$this->activationFunction = $activationFunction; |
71
|
|
|
$this->learningRate = $learningRate; |
|
|
|
|
72
|
|
|
|
73
|
|
|
$this->initNetwork(); |
74
|
|
|
} |
75
|
|
|
|
76
|
|
|
public function train(array $samples, array $targets): void |
77
|
|
|
{ |
78
|
|
|
$this->reset(); |
79
|
|
|
$this->initNetwork(); |
80
|
|
|
$this->partialTrain($samples, $targets, $this->classes); |
81
|
|
|
} |
82
|
|
|
|
83
|
|
|
/** |
84
|
|
|
* @throws InvalidArgumentException |
85
|
|
|
*/ |
86
|
|
|
public function partialTrain(array $samples, array $targets, array $classes = []): void |
87
|
|
|
{ |
88
|
|
|
if (!empty($classes) && array_values($classes) !== $this->classes) { |
89
|
|
|
// We require the list of classes in the constructor. |
90
|
|
|
throw InvalidArgumentException::inconsistentClasses(); |
91
|
|
|
} |
92
|
|
|
|
93
|
|
|
for ($i = 0; $i < $this->iterations; ++$i) { |
94
|
|
|
$this->trainSamples($samples, $targets); |
95
|
|
|
} |
96
|
|
|
} |
97
|
|
|
|
98
|
|
|
public function setLearningRate(float $learningRate): void |
99
|
|
|
{ |
100
|
|
|
$this->learningRate = $learningRate; |
101
|
|
|
$this->backpropagation->setLearningRate($this->learningRate); |
102
|
|
|
} |
103
|
|
|
|
104
|
|
|
/** |
105
|
|
|
* @param mixed $target |
106
|
|
|
*/ |
107
|
|
|
abstract protected function trainSample(array $sample, $target); |
108
|
|
|
|
109
|
|
|
/** |
110
|
|
|
* @return mixed |
111
|
|
|
*/ |
112
|
|
|
abstract protected function predictSample(array $sample); |
113
|
|
|
|
114
|
|
|
protected function reset(): void |
115
|
|
|
{ |
116
|
|
|
$this->removeLayers(); |
117
|
|
|
} |
118
|
|
|
|
119
|
|
|
private function initNetwork(): void |
120
|
|
|
{ |
121
|
|
|
$this->addInputLayer($this->inputLayerFeatures); |
122
|
|
|
$this->addNeuronLayers($this->hiddenLayers, $this->activationFunction); |
123
|
|
|
$this->addNeuronLayers([count($this->classes)], $this->activationFunction); |
124
|
|
|
|
125
|
|
|
$this->addBiasNodes(); |
126
|
|
|
$this->generateSynapses(); |
127
|
|
|
|
128
|
|
|
$this->backpropagation = new Backpropagation($this->learningRate); |
129
|
|
|
} |
130
|
|
|
|
131
|
|
|
private function addInputLayer(int $nodes): void |
132
|
|
|
{ |
133
|
|
|
$this->addLayer(new Layer($nodes, Input::class)); |
134
|
|
|
} |
135
|
|
|
|
136
|
|
|
private function addNeuronLayers(array $layers, ?ActivationFunction $activationFunction = null): void |
137
|
|
|
{ |
138
|
|
|
foreach ($layers as $neurons) { |
139
|
|
|
$this->addLayer(new Layer($neurons, Neuron::class, $activationFunction)); |
140
|
|
|
} |
141
|
|
|
} |
142
|
|
|
|
143
|
|
|
private function generateSynapses(): void |
144
|
|
|
{ |
145
|
|
|
$layersNumber = count($this->layers) - 1; |
146
|
|
|
for ($i = 0; $i < $layersNumber; ++$i) { |
147
|
|
|
$currentLayer = $this->layers[$i]; |
148
|
|
|
$nextLayer = $this->layers[$i + 1]; |
149
|
|
|
$this->generateLayerSynapses($nextLayer, $currentLayer); |
150
|
|
|
} |
151
|
|
|
} |
152
|
|
|
|
153
|
|
|
private function addBiasNodes(): void |
154
|
|
|
{ |
155
|
|
|
$biasLayers = count($this->layers) - 1; |
156
|
|
|
for ($i = 0; $i < $biasLayers; ++$i) { |
157
|
|
|
$this->layers[$i]->addNode(new Bias()); |
158
|
|
|
} |
159
|
|
|
} |
160
|
|
|
|
161
|
|
|
private function generateLayerSynapses(Layer $nextLayer, Layer $currentLayer): void |
162
|
|
|
{ |
163
|
|
|
foreach ($nextLayer->getNodes() as $nextNeuron) { |
164
|
|
|
if ($nextNeuron instanceof Neuron) { |
165
|
|
|
$this->generateNeuronSynapses($currentLayer, $nextNeuron); |
166
|
|
|
} |
167
|
|
|
} |
168
|
|
|
} |
169
|
|
|
|
170
|
|
|
private function generateNeuronSynapses(Layer $currentLayer, Neuron $nextNeuron): void |
171
|
|
|
{ |
172
|
|
|
foreach ($currentLayer->getNodes() as $currentNeuron) { |
173
|
|
|
$nextNeuron->addSynapse(new Synapse($currentNeuron)); |
174
|
|
|
} |
175
|
|
|
} |
176
|
|
|
|
177
|
|
|
private function trainSamples(array $samples, array $targets): void |
178
|
|
|
{ |
179
|
|
|
foreach ($targets as $key => $target) { |
180
|
|
|
$this->trainSample($samples[$key], $target); |
181
|
|
|
} |
182
|
|
|
} |
183
|
|
|
} |
184
|
|
|
|
In PHP it is possible to write to properties without declaring them. For example, the following is perfectly valid PHP code:
Generally, it is a good practice to explictly declare properties to avoid accidental typos and provide IDE auto-completion: