|
1
|
|
|
<?php |
|
2
|
|
|
|
|
3
|
|
|
declare (strict_types = 1); |
|
4
|
|
|
|
|
5
|
|
|
namespace Phpml\NeuralNetwork\Training; |
|
6
|
|
|
|
|
7
|
|
|
use Phpml\NeuralNetwork\Network; |
|
8
|
|
|
use Phpml\NeuralNetwork\Node\Neuron; |
|
9
|
|
|
use Phpml\NeuralNetwork\Training; |
|
10
|
|
|
use Phpml\NeuralNetwork\Training\Backpropagation\Sigma; |
|
11
|
|
|
|
|
12
|
|
|
class Backpropagation implements Training |
|
13
|
|
|
{ |
|
14
|
|
|
/** |
|
15
|
|
|
* @var Network |
|
16
|
|
|
*/ |
|
17
|
|
|
private $network; |
|
18
|
|
|
|
|
19
|
|
|
/** |
|
20
|
|
|
* @var int |
|
21
|
|
|
*/ |
|
22
|
|
|
private $theta; |
|
23
|
|
|
|
|
24
|
|
|
/** |
|
25
|
|
|
* @var array |
|
26
|
|
|
*/ |
|
27
|
|
|
private $sigmas; |
|
28
|
|
|
|
|
29
|
|
|
/** |
|
30
|
|
|
* @param Network $network |
|
31
|
|
|
* @param int $theta |
|
32
|
|
|
*/ |
|
33
|
|
|
public function __construct(Network $network, int $theta = 1) |
|
34
|
|
|
{ |
|
35
|
|
|
$this->network = $network; |
|
36
|
|
|
$this->theta = $theta; |
|
37
|
|
|
} |
|
38
|
|
|
|
|
39
|
|
|
/** |
|
40
|
|
|
* @param array $samples |
|
41
|
|
|
* @param array $targets |
|
42
|
|
|
* @param float $desiredError |
|
43
|
|
|
* @param int $maxIterations |
|
44
|
|
|
*/ |
|
45
|
|
|
public function train(array $samples, array $targets, float $desiredError = 0.001, int $maxIterations = 10000) |
|
46
|
|
|
{ |
|
47
|
|
|
for ($i = 0; $i < $maxIterations; ++$i) { |
|
48
|
|
|
$resultsWithinError = $this->trainSamples($samples, $targets, $desiredError); |
|
49
|
|
|
|
|
50
|
|
|
if ($resultsWithinError == count($samples)) { |
|
51
|
|
|
break; |
|
52
|
|
|
} |
|
53
|
|
|
} |
|
54
|
|
|
} |
|
55
|
|
|
|
|
56
|
|
|
/** |
|
57
|
|
|
* @param array $samples |
|
58
|
|
|
* @param array $targets |
|
59
|
|
|
* @param float $desiredError |
|
60
|
|
|
* |
|
61
|
|
|
* @return int |
|
62
|
|
|
*/ |
|
63
|
|
|
private function trainSamples(array $samples, array $targets, float $desiredError): int |
|
64
|
|
|
{ |
|
65
|
|
|
$resultsWithinError = 0; |
|
66
|
|
|
foreach ($targets as $key => $target) { |
|
67
|
|
|
$result = $this->network->setInput($samples[$key])->getOutput(); |
|
68
|
|
|
|
|
69
|
|
|
if ($this->isResultWithinError($result, $target, $desiredError)) { |
|
70
|
|
|
++$resultsWithinError; |
|
71
|
|
|
} else { |
|
72
|
|
|
$this->trainSample($samples[$key], $target); |
|
73
|
|
|
} |
|
74
|
|
|
} |
|
75
|
|
|
|
|
76
|
|
|
return $resultsWithinError; |
|
77
|
|
|
} |
|
78
|
|
|
|
|
79
|
|
|
/** |
|
80
|
|
|
* @param array $sample |
|
81
|
|
|
* @param array $target |
|
82
|
|
|
*/ |
|
83
|
|
|
private function trainSample(array $sample, array $target) |
|
84
|
|
|
{ |
|
85
|
|
|
$this->network->setInput($sample)->getOutput(); |
|
86
|
|
|
$this->sigmas = []; |
|
87
|
|
|
|
|
88
|
|
|
$layers = $this->network->getLayers(); |
|
89
|
|
|
$layersNumber = count($layers); |
|
90
|
|
|
|
|
91
|
|
|
for ($i = $layersNumber; $i > 1; --$i) { |
|
92
|
|
|
foreach ($layers[$i - 1]->getNodes() as $key => $neuron) { |
|
93
|
|
|
if ($neuron instanceof Neuron) { |
|
94
|
|
|
$sigma = $this->getSigma($neuron, $target, $key, $i == $layersNumber); |
|
95
|
|
|
foreach ($neuron->getSynapses() as $synapse) { |
|
96
|
|
|
$synapse->changeWeight($this->theta * $sigma * $synapse->getNode()->getOutput()); |
|
97
|
|
|
} |
|
98
|
|
|
} |
|
99
|
|
|
} |
|
100
|
|
|
} |
|
101
|
|
|
} |
|
102
|
|
|
|
|
103
|
|
|
/** |
|
104
|
|
|
* @param Neuron $neuron |
|
105
|
|
|
* @param array $target |
|
106
|
|
|
* @param int $key |
|
107
|
|
|
* @param bool $lastLayer |
|
108
|
|
|
* |
|
109
|
|
|
* @return float |
|
110
|
|
|
*/ |
|
111
|
|
|
private function getSigma(Neuron $neuron, array $target, int $key, bool $lastLayer): float |
|
112
|
|
|
{ |
|
113
|
|
|
$neuronOutput = $neuron->getOutput(); |
|
114
|
|
|
$sigma = $neuronOutput * (1 - $neuronOutput); |
|
115
|
|
|
|
|
116
|
|
|
if ($lastLayer) { |
|
117
|
|
|
$sigma *= ($target[$key] - $neuronOutput); |
|
118
|
|
|
} else { |
|
119
|
|
|
$sigma *= $this->getPrevSigma($neuron); |
|
120
|
|
|
} |
|
121
|
|
|
|
|
122
|
|
|
$this->sigmas[] = new Sigma($neuron, $sigma); |
|
123
|
|
|
|
|
124
|
|
|
return $sigma; |
|
125
|
|
|
} |
|
126
|
|
|
|
|
127
|
|
|
/** |
|
128
|
|
|
* @param Neuron $neuron |
|
129
|
|
|
* |
|
130
|
|
|
* @return float |
|
131
|
|
|
*/ |
|
132
|
|
|
private function getPrevSigma(Neuron $neuron): float |
|
133
|
|
|
{ |
|
134
|
|
|
$sigma = 0.0; |
|
135
|
|
|
|
|
136
|
|
|
foreach ($this->sigmas as $neuronSigma) { |
|
137
|
|
|
$sigma += $neuronSigma->getSigmaForNeuron($neuron); |
|
138
|
|
|
} |
|
139
|
|
|
|
|
140
|
|
|
return $sigma; |
|
141
|
|
|
} |
|
142
|
|
|
|
|
143
|
|
|
/** |
|
144
|
|
|
* @param array $result |
|
145
|
|
|
* @param array $target |
|
146
|
|
|
* @param float $desiredError |
|
147
|
|
|
* |
|
148
|
|
|
* @return bool |
|
149
|
|
|
*/ |
|
150
|
|
|
private function isResultWithinError(array $result, array $target, float $desiredError) |
|
151
|
|
|
{ |
|
152
|
|
|
foreach ($target as $key => $value) { |
|
153
|
|
|
if ($result[$key] > $value + $desiredError || $result[$key] < $value - $desiredError) { |
|
154
|
|
|
return false; |
|
155
|
|
|
} |
|
156
|
|
|
} |
|
157
|
|
|
|
|
158
|
|
|
return true; |
|
159
|
|
|
} |
|
160
|
|
|
} |
|
161
|
|
|
|
php-ai /
php-ml
