圖片讀取ImageDataGenerator()

ImageDataGenerator()是keras.preprocessing.image模塊中的圖片生成器蝇摸，同時也可以在batch中對數(shù)據(jù)進行增強，擴充數(shù)據(jù)集大小律歼，增強模型的泛化能力险毁。比如進行旋轉(zhuǎn)畔况，變形慧库，歸一化等等。

keras.preprocessing.image.ImageDataGenerator(featurewise_center=False, samplewise_center=False, featurewise_std_normalization=False, samplewise_std_normalization=False, zca_whitening=False, zca_epsilon=1e-06, rotation_range=0.0, width_shift_range=0.0, height_shift_range=0.0, brightness_range=None, shear_range=0.0, zoom_range=0.0, channel_shift_range=0.0, fill_mode='nearest', cval=0.0, horizontal_flip=False, vertical_flip=False, rescale=None, preprocessing_function=None, data_format=None, validation_split=0.0)

參數(shù):

• featurewise_center: Boolean. 對輸入的圖片每個通道減去每個通道對應(yīng)均值域庇。
• samplewise_center: Boolan. 每張圖片減去樣本均值, 使得每個樣本均值為0听皿。
• featurewise_std_normalization(): Boolean()
• samplewise_std_normalization(): Boolean()
• zca_epsilon(): Default 12-6
• zca_whitening: Boolean. 去除樣本之間的相關(guān)性
• rotation_range(): 旋轉(zhuǎn)范圍
• width_shift_range(): 水平平移范圍
• height_shift_range(): 垂直平移范圍
• shear_range(): float, 透視變換的范圍
• zoom_range(): 縮放范圍
• fill_mode: 填充模式, constant, nearest, reflect
• cval: fill_mode == 'constant'的時候填充值
• horizontal_flip(): 水平反轉(zhuǎn)
• vertical_flip(): 垂直翻轉(zhuǎn)
• preprocessing_function(): user提供的處理函數(shù)
• data_format(): channels_first或者channels_last
• validation_split(): 多少數(shù)據(jù)用于驗證集

方法:

• apply_transform(x, transform_parameters):根據(jù)參數(shù)對x進行變換
• fit(x, augment=False, rounds=1, seed=None): 將生成器用于數(shù)據(jù)x,從數(shù)據(jù)x中獲得樣本的統(tǒng)計參數(shù), 只有featurewise_center, featurewise_std_normalization或者zca_whitening為True才需要
• flow(x, y=None, batch_size=32, shuffle=True, sample_weight=None, seed=None, save_to_dir=None, save_prefix='', save_format='png', subset=None) ):按batch_size大小從x,y生成增強數(shù)據(jù)
• flow_from_directory()從路徑生成增強數(shù)據(jù),和flow方法相比最大的優(yōu)點在于不用一次將所有的數(shù)據(jù)讀入內(nèi)存當(dāng)中,這樣減小內(nèi)存壓力，這樣不會發(fā)生OOM吗冤，血的教訓(xùn)。
• get_random_transform(img_shape, seed=None): 返回包含隨機圖像變換參數(shù)的字典
• random_transform(x, seed=None): 進行隨機圖像變換, 通過設(shè)置seed可以達到同步變換覆致。
• standardize(x): 對x進行歸一化

實例：
mnist分類數(shù)據(jù)增強

 from keras.preprocessing.image import ImageDataGenerator
 from keras.datasets import mnist
 from keras.datasets import cifar10
 from keras.utils import np_utils
 import numpy as np
 import matplotlib.pyplot as plt
 num_classes = 10
 (x_train, y_train), (x_test, y_test) = mnist.load_data()
 x_train = np.expand_dims(x_train, axis = 3)
 y_train = np_utils.to_categorical(y_train, num_classes)
 y_test = np_utils.to_categorical(y_test, num_classes)
 
 datagen = ImageDataGenerator(
     featurewise_center=True,
     featurewise_std_normalization=True,
     rotation_range=20,
     width_shift_range=0.2,
     height_shift_range=0.2,
     horizontal_flip=True)
 
 # compute quantities required for featurewise normalization
 # (std, mean, and principal components if ZCA whitening is applied)
 datagen.fit(x_train)
 
 data_iter = datagen.flow(x_train, y_train, batch_size=8)
 
while True:
     x_batch, y_batch = data_iter.next()
     for i in range(8):
         print(i//4)
         plt.subplot(2,4,i+1)
         plt.imshow(x_batch[i].reshape(28,28), cmap='gray')
     plt.show()

5FGhke.png

portrait分割數(shù)據(jù)增強,需要對image和mask同步處理:
featurewise結(jié)果：

from keras.preprocessing.image import ImageDataGenerator
from keras.datasets import mnist
from keras.datasets import cifar10
from keras.utils import np_utils
import numpy as np
import matplotlib.pyplot as plt
num_classes = 10
seed = 1
# featurewise需要數(shù)據(jù)集的統(tǒng)計信息儡羔，因此需要先讀入一個x_train汰蜘，用于對增強圖像的均值和方差處理族操。
x_train = np.load('images-224.npy')
imagegen = ImageDataGenerator(
    featurewise_center=True,
    featurewise_std_normalization=True,
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True)

maskgen = ImageDataGenerator(
     rescale = 1./255,
     rotation_range=20,
     width_shift_range=0.2,
     height_shift_range=0.2,
     horizontal_flip=True)

# compute quantities required for featurewise normalization
# (std, mean, and principal components if ZCA whitening is applied)
imagegen.fit(x_train)
image_iter = imagegen.flow_from_directory('../data/images',target_size=(224,224), class_mode=None, batch_size=8, seed=seed)
mask_iter = maskgen.flow_from_directory('../data/masks', color_mode='rgb', target_size=(224,224), class_mode=None, batch_size=8, seed=seed)
data_iter = zip(image_iter, mask_iter)
while True:
    for x_batch, y_batch in data_iter:
        for i in range(8):
            print(i//4)
            plt.subplot(2,8,i+1)
            plt.imshow(x_batch[i].reshape(224,224,3))
            plt.subplot(2,8,8+i+1)
            plt.imshow(y_batch[i].reshape(224,224, 3), cmap='gray')
        plt.show()

5FGPpR.png

samplewise結(jié)果:

from keras.preprocessing.image import ImageDataGenerator
from keras.datasets import mnist
from keras.datasets import cifar10
from keras.utils import np_utils
import numpy as np
import matplotlib.pyplot as plt
num_classes = 10
seed = 1
imagegen = ImageDataGenerator(
    samplewise_center=True,
    samplewise_std_normalization=True,
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True)
maskgen = ImageDataGenerator(
     rescale = 1./255,
     rotation_range=20,
     width_shift_range=0.2,
     height_shift_range=0.2,
     horizontal_flip=True)
image_iter = imagegen.flow_from_directory('../data/images',target_size=(224,224), class_mode=None, batch_size=8, seed=seed)
mask_iter = maskgen.flow_from_directory('../data/masks', color_mode='rgb', target_size=(224,224), class_mode=None, batch_size=8, seed=seed)
data_iter = zip(image_iter, mask_iter)
while True:
    for x_batch, y_batch in data_iter:
        for i in range(8):
            print(i//4)
            plt.subplot(2,8,i+1)
            plt.imshow(x_batch[i].reshape(224,224,3))
            plt.subplot(2,8,8+i+1)
            plt.imshow(y_batch[i].reshape(224,224, 3), cmap='gray')
        plt.show()

5FGa1r.png

data/
...images/
........./images
...masks/
........./masks

只有這樣提供才能保證正確讀取圖片姐赡，沒有子目錄會檢測不到圖片柠掂。
此外正如github上的issue:https://github.com/keras-team/keras/pull/3052/commits/81fb0fa7c332b1b9d2669d68797fda041de17088

for subdir in sorted(os.listdir(directory)):
                if os.path.isdir(os.path.join(directory, subdir)):
                    classes.append(subdir)

flow_from_directory()會從路徑推測label, 在進行映射之前涯贞，會先對路徑進行排序，具體順序是alphanumerically州疾， 也是os.listdir()對子目錄排序的結(jié)果严蓖。這樣你才知道具體來說哪個路徑的類對應(yīng)哪個label氧急。
原圖:

5FGcqY.png