<?PHP
ini_set('memory_limit', '20480M'); // Increase the memory limit to 20480MB (20GB)
include_once("../../CLASSES/Headers.php");
use NameSpaceNumpyLight\NumpyLight;
use NameSpaceRandomGenerator\RandomGenerator;
use NameSpaceActivationRelu\Activation_Relu;
use NameSpaceOptimizerSGD\Optimizer_SGD;
use NameSpaceOptimizerAdagrad\Optimizer_Adagrad;
use NameSpaceOptimizerRMSprop\Optimizer_RMSprop;
function load_mnist_dataset($dataset, $path) {
$labels = [];
$dir = $path . '/' . $dataset;
// Check if the main directory exists and is readable
if (is_readable($dir) && ($dir_content = scandir($dir))) {
foreach ($dir_content as $item) {
if ($item !== '.' && $item !== '..' && is_dir($dir . '/' . $item)) {
$labels[] = $item;
}
}
}
$X = [];
$y = [];
foreach ($labels as $label) {
$label_path = $dir . '/' . $label;
if (is_readable($label_path) && ($files = scandir($label_path))) {
foreach ($files as $file) {
if ($file !== '.' && $file !== '..') {
$filePath = $label_path . '/' . $file;
if (is_readable($filePath) && !is_dir($filePath)) {
$imageProcessor = new ImageProcessor($filePath);
$imageData = $imageProcessor->getImageGrayscaleArray(["rows" => 28, "cols" => 28]);
$X[] = $imageData;
$y[] = $label;
}
}
}
}
}
return ["X" => $X, "y" => $y];
}
function create_data_mnist($path) {
// Load both sets separately
$trainData = load_mnist_dataset('train', $path);
$X = $trainData['X'];
$y = $trainData['y'];
$testData = load_mnist_dataset('test', $path);
$X_test = $testData['X'];
$y_test = $testData['y'];
// And return all the data
return [$X,$y,$X_test,$y_test];
}
$mnist_data = create_data_mnist("fashion_mnist_images");
list($X, $y, $X_test, $y_test) = $mnist_data;
$keys = range(0, NumpyLight::shape($X)[0] - 1);
NumpyLight::random()->shuffle($keys);
$X = NumpyLight::rearrangeMatrix($X, $keys);
$y = NumpyLight::rearrangeMatrix($y, $keys);
# Scale and reshape samples
// X = (X.reshape(X.shape[ 0 ], - 1 ).astype(np.float32) - 127.5 ) / 127.5;
// $X = NumpyLight::reshape($X,[NumpyLight::shape($X)[0],-1]);
$X = NumpyLight::divide(
NumpyLight::subtract(
NumpyLight::reshape(
$X,
[NumpyLight::shape($X)[0],-1])
,
127.5)
,127.5);
$X_test = NumpyLight::divide(
NumpyLight::subtract(
NumpyLight::reshape(
$X_test,
[NumpyLight::shape($X_test)[0],-1])
,
127.5)
,127.5);
$validation = [$X_test, $y_test];
echo "\n\nModel Init\n\n";
$Model = new Model();
$Model->add(new Layer_Dense(NumpyLight::shape($X)[1],64));
$Model->add(new Activation_Relu());
$Model->add(new Layer_Dense(64,64));
$Model->add(new Activation_Relu());
$Model->add(new Layer_Dense(64,64));
$Model->add(new Activation_Softmax());
$Model->set(
$loss_function = new Loss_CategoricalCrossentropy(),
$optimizer = new Optimizer_Adam($learning_rate = 0.001, $decay = 1e-3),
$accuracy = new Accuracy_Categorical()
);
$Model->finalize();
$Model->train($X, $y,$epoch = 200, $batch_size = 128,$print_every = 100,$validation_data = $validation);
$Model->save();
// $ModelParameters = $Model->get_parameters();
// echo "\n\nModel from Params\n\n";
// $Model = new Model();
// $Model->add(new Layer_Dense(NumpyLight::shape($X)[1],64));
// $Model->add(new Activation_Relu());
// $Model->add(new Layer_Dense(64,64));
// $Model->add(new Activation_Relu());
// $Model->add(new Layer_Dense(64,64));
// $Model->add(new Activation_Softmax());
// $Model->set(
// $loss_function = new Loss_CategoricalCrossentropy(),
// $optimizer = new Optimizer_Adam($learning_rate = 0.001, $decay = 1e-3),
// $accuracy = new Accuracy_Categorical()
// );
// $Model->finalize();
// $Model->set_parameters($ModelParameters);
// $Model->evaluate($X_test, $y_test);
?>
|
Download
