1
|
|
|
<?php |
2
|
|
|
|
3
|
|
|
declare(strict_types=1); |
4
|
|
|
|
5
|
|
|
namespace Phpml\NeuralNetwork\Training; |
6
|
|
|
|
7
|
|
|
use Phpml\NeuralNetwork\Node\Neuron; |
8
|
|
|
use Phpml\NeuralNetwork\Training\Backpropagation\Sigma; |
9
|
|
|
|
10
|
|
|
class Backpropagation |
11
|
|
|
{ |
12
|
|
|
/** |
13
|
|
|
* @var float |
14
|
|
|
*/ |
15
|
|
|
private $learningRate; |
16
|
|
|
|
17
|
|
|
/** |
18
|
|
|
* @var array |
19
|
|
|
*/ |
20
|
|
|
private $sigmas = []; |
21
|
|
|
|
22
|
|
|
/** |
23
|
|
|
* @var array |
24
|
|
|
*/ |
25
|
|
|
private $prevSigmas = []; |
26
|
|
|
|
27
|
|
|
public function __construct(float $learningRate) |
28
|
|
|
{ |
29
|
|
|
$this->setLearningRate($learningRate); |
30
|
|
|
} |
31
|
|
|
|
32
|
|
|
public function setLearningRate(float $learningRate): void |
33
|
|
|
{ |
34
|
|
|
$this->learningRate = $learningRate; |
35
|
|
|
} |
36
|
|
|
|
37
|
|
|
public function getLearningRate(): float |
38
|
|
|
{ |
39
|
|
|
return $this->learningRate; |
40
|
|
|
} |
41
|
|
|
|
42
|
|
|
/** |
43
|
|
|
* @param mixed $targetClass |
44
|
|
|
*/ |
45
|
|
|
public function backpropagate(array $layers, $targetClass): void |
46
|
|
|
{ |
47
|
|
|
$layersNumber = count($layers); |
48
|
|
|
|
49
|
|
|
// Backpropagation. |
50
|
|
|
for ($i = $layersNumber; $i > 1; --$i) { |
51
|
|
|
$this->sigmas = []; |
52
|
|
|
foreach ($layers[$i - 1]->getNodes() as $key => $neuron) { |
53
|
|
|
if ($neuron instanceof Neuron) { |
54
|
|
|
$sigma = $this->getSigma($neuron, $targetClass, $key, $i == $layersNumber); |
55
|
|
|
foreach ($neuron->getSynapses() as $synapse) { |
56
|
|
|
$synapse->changeWeight($this->learningRate * $sigma * $synapse->getNode()->getOutput()); |
57
|
|
|
} |
58
|
|
|
} |
59
|
|
|
} |
60
|
|
|
|
61
|
|
|
$this->prevSigmas = $this->sigmas; |
62
|
|
|
} |
63
|
|
|
|
64
|
|
|
// Clean some memory (also it helps make MLP persistency & children more maintainable). |
65
|
|
|
$this->sigmas = []; |
66
|
|
|
$this->prevSigmas = []; |
67
|
|
|
} |
68
|
|
|
|
69
|
|
|
private function getSigma(Neuron $neuron, int $targetClass, int $key, bool $lastLayer): float |
70
|
|
|
{ |
71
|
|
|
$neuronOutput = $neuron->getOutput(); |
72
|
|
|
$sigma = $neuron->getDerivative(); |
73
|
|
|
|
74
|
|
|
if ($lastLayer) { |
75
|
|
|
$value = 0; |
76
|
|
|
if ($targetClass === $key) { |
77
|
|
|
$value = 1; |
78
|
|
|
} |
79
|
|
|
|
80
|
|
|
$sigma *= ($value - $neuronOutput); |
81
|
|
|
} else { |
82
|
|
|
$sigma *= $this->getPrevSigma($neuron); |
83
|
|
|
} |
84
|
|
|
|
85
|
|
|
$this->sigmas[] = new Sigma($neuron, $sigma); |
86
|
|
|
|
87
|
|
|
return $sigma; |
88
|
|
|
} |
89
|
|
|
|
90
|
|
|
private function getPrevSigma(Neuron $neuron): float |
91
|
|
|
{ |
92
|
|
|
$sigma = 0.0; |
93
|
|
|
|
94
|
|
|
foreach ($this->prevSigmas as $neuronSigma) { |
95
|
|
|
$sigma += $neuronSigma->getSigmaForNeuron($neuron); |
96
|
|
|
} |
97
|
|
|
|
98
|
|
|
return $sigma; |
99
|
|
|
} |
100
|
|
|
} |
101
|
|
|
|