TensorFlow Keras ImageDataGenerator VGG16 Bottleneck

猫狗识别  

#%%

from keras.applications.vgg16 import VGG16
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential, load_model
from keras.layers import Dropout, Flatten, Dense
from keras.optimizers import Adam

#%%

# 载入VGG网络，不包括全连接层
model = VGG16(weights='imagenet', include_top=False)

#%%

model.summary()

#%%

# 对图片进行随机处理，以扩大数据集
datagen = ImageDataGenerator(
    # 随机旋转角度
    rotation_range=40,
    # 随机水平平移
    width_shift_range=0.2,
    # 随机竖直平移
    height_shift_range=0.2,
    # 数值归一化
    rescale=1. / 255,
    # 随机裁剪
    shear_range=0.2,
    # 随机放大
    zoom_range=0.2,
    # 水平翻转
    horizontal_flip=True,
    # 填充方式
    fill_mode='nearest'
)

#%%

batch_size = 32
# 训练集生成器
train_generator = datagen.flow_from_directory(
    # 训练集路径
    'image/train',
    # 设置生成的图片大小
    target_size=(150, 150),
    # 批次大小
    batch_size=batch_size,
    # 不生成标签
    # class_mode=None,
    # 不随机打乱
    shuffle=False
)
# 测试集生成器
test_generator = datagen.flow_from_directory(
    # 训练集路径
    'image/test',
    # 设置生成的图片大小
    target_size=(150, 150),
    # 批次大小
    batch_size=batch_size,
    # 不生成标签
    # class_mode=None,
    # 不随机打乱
    shuffle=False
)

#%%

# 得到经过卷积层和池化层后的训练集数据
# 相当于model.predict
bottleneck_features_train = model.predict_generator(
    # 训练集生成器
    train_generator,
    # batch数量 变换1次
    train_generator.samples / batch_size * 1)
# (400, 4, 4, 512)
# (image数量, row, col, 层数)
print(bottleneck_features_train.shape)
bottleneck_features_test = model.predict_generator(
    # 测试集生成器
    test_generator,
    # batch数量 变换1次
    test_generator.samples / batch_size * 1)
# (200, 4, 4, 512)
print(bottleneck_features_test.shape)

#%%

# 保存这些数据 训练集和测试集bottleneck结果
np.save(open("bottleneck_features_train.npy", "wb"), bottleneck_features_train)
np.save(open("bottleneck_features_test.npy", "wb"), bottleneck_features_test)

#%%

# 读取训练集和测试集bottleneck结果
train_data = np.load(open('bottleneck_features_train.npy', 'rb'))
test_data = np.load(open('bottleneck_features_test.npy', 'rb'))
print(train_data.shape)
print(test_data.shape)

#%%

'''labels = np.array([0] * 20 + [1] * 20)
train_labels = np.array([])
for i in range(10):
    train_labels = np.concatenate((train_labels, labels))

labels = np.array([0] * 10 + [1] * 10)
test_labels = np.array([])
for i in range(10):
    test_labels = np.concatenate((test_labels, labels))'''

#%%

train_labels = np.zeros((train_data.shape[0], 2))
for i in range(int(train_data.shape[0] / 2)):
    train_labels[i, 0] = 1
for i in range(int(train_data.shape[0] / 2), train_data.shape[0]):
    train_labels[i, 1] = 1
test_labels = np.zeros((test_data.shape[0], 2))
for i in range(int(test_data.shape[0] / 2)):
    test_labels[i, 0] = 1
for i in range(int(test_data.shape[0] / 2), test_data.shape[0]):
    test_labels[i, 1] = 1
#%%

model = Sequential()
model.add(Flatten(input_shape=train_data.shape[1:]))
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(2, activation='softmax'))

adam = Adam(lr=1e-4)

#%%

model.compile(optimizer=adam, loss="categorical_crossentropy", metrics=['accuracy'])

#%%

model.summary()

#%%

model.fit(train_data, train_labels, epochs=20)

#%%

model.save("bottleneck_fc_model.h5")

#%%

model = load_model("bottleneck_fc_model.h5")

#%%

test_pred = model.predict(test_data)