Créer son propre générateur de données

Joseph Konka @joekakone

03/03/2022 · 3 min

tensorflow keras datagenerator

L'API ImageDataGenerator de Keras peut-être limitée dans certaines situations nécessitant la création d'un générateur de données personalisé, c'est l'objet de ce nouveau tutoriel.

Créer un générateur de données

Nous avons besoin d'une fonction qui permet de lire les données

def load_image(path, size):
    """custom function to load image"""
    image = Image.open(path)
    image = image.convert('RGB')
    image = image.resize(size)
    image = np.array(image)*(1./255)

    return image

class DataGenerator(tf.keras.utils.Sequence):
    """Custom generator"""
    def __init__(self, x_set, y_set, batch_size=32, target_size=(150, 150)):
        self.x, self.y = x_set, y_set
        self.batch_size = batch_size
        self.target_size = target_size

    def __len__(self):
        return math.ceil(len(self.x) / self.batch_size)

    def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) *
        self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) *
        self.batch_size]

        return np.array([
            load_image(file_name, target_size)
               for file_name in batch_x]), np.array(batch_y)

Appeler la classe#

train_ds = DataGenerator(train_images, train_labels, batch_size=BATCH_SIZE, target_size=(WIDTH, HEIGHT))
val_ds = DataGenerator(val_images, val_labels, batch_size=BATCH_SIZE, target_size=(WIDTH, HEIGHT))

Code complet#

def load_image(path, size):
    """custom function to load image"""
    image = Image.open(path)
    image = image.convert('RGB')
    image = image.resize(size)
    image = np.array(image)*(1./255)

    return image

class DataGenerator(tf.keras.utils.Sequence):
    """Custom generator"""
    def __init__(self, x_set, y_set, batch_size=32, target_size=(150, 150)):
        self.x, self.y = x_set, y_set
        self.batch_size = batch_size
        self.target_size = target_size

    def __len__(self):
        return math.ceil(len(self.x) / self.batch_size)

    def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) *
        self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) *
        self.batch_size]

        return np.array([
            load_image(file_name, target_size)
               for file_name in batch_x]), np.array(batch_y)

.

Commentaires

Inscrivez-vous à une formation complète

Voir le catalogue