Keras 使用自定义图像生成器进行增强

Question

我正在使用自定义图像生成器按此处所述批量从磁盘读取数据，https:// keras.io/examples/vision/oxford_pets_image_segmentation/

确切的生成器如下所示：

from tensorflow import keras
import numpy as np
from tensorflow.keras.preprocessing.image import load_img
import tensorflow

class OxfordPets(keras.utils.Sequence):
    """Helper to iterate over the data (as Numpy arrays)."""

    def __init__(self, batch_size, img_size, input_img_paths, target_img_paths):
        self.batch_size = batch_size
        self.img_size = img_size
        self.input_img_paths = input_img_paths
        self.target_img_paths = target_img_paths

    def __len__(self):
        return len(self.target_img_paths) // self.batch_size

    def __getitem__(self, idx):
        """Returns tuple (input, target) correspond to batch #idx."""
        i = idx * self.batch_size
        batch_input_img_paths = self.input_img_paths[i : i + self.batch_size]
        batch_target_img_paths = self.target_img_paths[i : i + self.batch_size]
        x = np.zeros((self.batch_size,) + self.img_size + (3,), dtype="float32")
        for j, path in enumerate(batch_input_img_paths):
            img = load_img(path, target_size=self.img_size)
            x[j] = img
        y = np.zeros((self.batch_size,) + self.img_size + (1,), dtype="uint8")
        for j, path in enumerate(batch_target_img_paths):
            img = load_img(path, target_size=self.img_size, color_mode="grayscale")
            y[j] = np.expand_dims(img, 2)
            # Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2:
            y[j] -= 1
        return x, y

这非常有效，并且很有帮助，因为我的数据太大而无法读入内存。它需要两条路径，分别是输入图像 (batch_input_img_paths) 和蒙版 (batch_target_img_paths) 的路径。

我想修改这个生成器以使用数据增强。我正在尝试这个：

class OxfordPets(keras.utils.Sequence)： “”“帮助迭代数据（作为 Numpy 数组）。”“”

    def __init__(self, batch_size, img_size, input_img_paths, target_img_paths):
        self.batch_size = batch_size
        self.img_size = img_size
        self.input_img_paths = input_img_paths
        self.target_img_paths = target_img_paths

    def __len__(self):
        return len(self.target_img_paths) // self.batch_size
    
    def __data_augmentation(self, img):
         ''' function for apply some data augmentation '''
         img = tensorflow.keras.preprocessing.image.random_shift(img, 0.2, 0.2)
         img = tensorflow.keras.preprocessing.image.random_zoom(img, 0.2)
         img = tensorflow.keras.preprocessing.image.random_shear(img, 0.2)
         img = tensorflow.keras.preprocessing.image.random_rotation(img, 40)
         img = tensorflow.image.random_flip_left_right(img)
         img = tensorflow.image.random_flip_up_down(img)
         return img

    def __getitem__(self, idx):
        """Returns tuple (input, target) correspond to batch #idx."""
        i = idx * self.batch_size
        batch_input_img_paths = self.input_img_paths[i : i + self.batch_size]
        batch_target_img_paths = self.target_img_paths[i : i + self.batch_size]
        x = np.zeros((self.batch_size,) + self.img_size + (3,), dtype="float32")
        for j, path in enumerate(batch_input_img_paths):
            img = load_img(path, target_size=self.img_size)

            #apply augmentation
            img = self.__data_augmentation(img)

            x[j] = img
        y = np.zeros((self.batch_size,) + self.img_size + (1,), dtype="uint8")
        for j, path in enumerate(batch_target_img_paths):
            img = load_img(path, target_size=self.img_size, color_mode="grayscale")
            y[j] = np.expand_dims(img, 2)

             #apply augmentation
            img = self.__data_augmentation(img)

            # Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2:
            y[j] -= 1
        return x, y

但这会返回：

AttributeError: 'Image' object has no attribute 'shape'

我不确定的另一件事是是否将完全相同的增强应用于数据和标签。也许为此需要随机种子？我想做的另一件事是不仅将增强图像输入到最终网络中，而且通过增强来增加训练规模，以便输入一些真实图像和一些增强图像，例如仅增强 30% 的图像输入图像。

编辑：

我认为我的增强工作是这样的：

class OxfordPets(keras.utils.Sequence):
    """Helper to iterate over the data (as Numpy arrays)."""

    def __init__(self, batch_size, img_size, input_img_paths, target_img_paths):
        self.batch_size = batch_size
        self.img_size = img_size
        
        #augmentation
        self.augmentor = tensorflow.keras.preprocessing.image.ImageDataGenerator(
        rotation_range=40,
        width_shift_range=0.2,
        height_shift_range=0.2,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True)

        self.input_img_paths = input_img_paths
        self.target_img_paths = target_img_paths
       
    def __len__(self):
        return len(self.target_img_paths) // self.batch_size


    def __getitem__(self, idx):
        """Returns tuple (input, target) correspond to batch #idx."""
        i = idx * self.batch_size
        batch_input_img_paths = self.input_img_paths[i : i + self.batch_size]
        batch_target_img_paths = self.target_img_paths[i : i + self.batch_size]
        x = np.zeros((self.batch_size,) + self.img_size + (3,), dtype="float32")
        for j, path in enumerate(batch_input_img_paths):
            img = load_img(path, target_size=self.img_size)
           
            x[j] = img
        y = np.zeros((self.batch_size,) + self.img_size + (1,), dtype="uint8")
        for j, path in enumerate(batch_target_img_paths):
            img = load_img(path, target_size=self.img_size, color_mode="grayscale")

            
            y[j] = np.expand_dims(img, 2)
            # Ground truth labels are 1, 2, 3. Subtract one to make them 0, 1, 2:
            y[j] -= 1
        return x, y

但这并不能解决我实际增加数据集大小的问题，我相信它现在只发送增强图像，而没有真正的数据。