第五章 深度学习用于计算机视觉

""" 3.2.使用预训练的卷积神经网络--模型微调 """
# 1.将VGG16卷积基实例化
from keras.applications import VGG16

conv_base = VGG16(weights = 'imagenet',
                  include_top = False,
                  input_shape = (150, 150, 3)) 



# 2.在卷积基上添加一个密集连接分类器
from keras import models
from keras import layers

model = models.Sequential()
model.add(conv_base)            # 添加的conv_base模型
model.add(layers.Flatten())     # 用Flatten压平 
model.add(layers.Dense(256, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
model.summary()

print('这是在冻结卷积基之前可训练的权重的数量:', len(model.trainable_weights))    #30
# 冻结卷积基(冻结一个或多个层是指在训练过程中保持其权重不变,防止卷积基之前学到的表示被修改)

conv_base.trainable = False    # 冻结   只有添加的两个Dense层的权重才会被训练
print('这是在冻结卷积基之后可训练的权重的数量:', len(model.trainable_weights))    #4

# 3.利用冻结的卷积基端到端地训练模型
from keras.preprocessing.image import ImageDataGenerator
from keras import optimizers
import os

base_dir = 'F:\数据集\dogs-vs-csts\猫狗大战'             # 保存较小数据集的目录
train_dir = os.path.join(base_dir, 'train')              # 划分后的训练集
validation_dir = os.path.join(base_dir, 'validation')    # 划分后的验证集
test_dir = os.path.join(base_dir, 'test')                # 划分后的测试集

train_datagen = ImageDataGenerator(     #　数据增强
    rescale = 1./255,                        # 调整尺寸
    rotation_range = 40,                     # 角度值0～180
    width_shift_range = 0.2,                 # 水平移动的范围
    height_shift_range = 0.2,                # 垂直移动的范围
    shear_range = 0.2,                       # 切错变化
    zoom_range = 0.2,                        # 缩放
    horizontal_flip = True,                  # 随机将一半图像水平翻转
    fill_mode = 'nearest')                   # 填充创新像素

test_datagen = ImageDataGenerator(rescale = 1./255)   # 注意，不能增强验证数据

train_generator = train_datagen.flow_from_directory(
    train_dir,                       # 目标目录
    target_size = (150,150),          # 将所有图像的大小调整为150×150
    batch_size = 20,
    class_mode = 'binary')

# 以文件夹路径为参数,生成经过数据提升/归一化后的数据,在一个无限循环中无限产生batch数据
validation_generator = train_datagen.flow_from_directory(
    validation_dir,                  # 目标目录
    target_size = (150,150),          # 将所有图片的大小调整为150×150
    batch_size = 20,                 # batch数据的大小，默认32
    class_mode = 'binary')           # 因为使用了binary_crossentropy损失，所以需要用二进制标签

model.compile(loss = 'binary_crossentropy',           # 配置模型
              optimizer = optimizers.RMSprop(1e-5),
              metrics = ['acc'])

history = model.fit_generator(               # 利用批量生成器拟合模型
    train_generator,
    steps_per_epoch = 100,
    epochs = 30,
    validation_data = validation_generator,
    validation_steps = 50)


conv_base.summary()      # 查看卷积层的结构


# In[]:第二部分
# 4.冻结直到某一层的所有层
conv_base.trainable = True          # 解冻

set_trainable = False
for layer in conv_base.layers:
    if layer.name == 'block5_conv1':
        set_trainable = True        
    if set_trainable:
        layer.trainable = True        # 解冻block5_conv1层之后的
    else:
        layer.trainable = False       # 冷冻


# 5.微调模型
model.compile(loss = 'binary_crossentropy',
              optimizer = optimizers.RMSprop(lr = 1e-5),
              metrics = ['acc'])

history = model.fit_generator(               # 利用批量生成器拟合模型
    train_generator,
    steps_per_epoch = 100,
    epochs = 100,
    validation_data = validation_generator, 
    validation_steps = 50)


# In[]:
# 6.绘制训练过程中的损失曲线和精度曲线
import matplotlib.pyplot as plt

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(1, len(acc)+1)

plt.plot(epochs, acc, 'r', label='Training acc')            # 绘制训练过程中的精度曲线
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and Validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'r', label='Training loss')          # 绘制训练过程中的损失曲线
plt.plot(epochs, val_loss, 'b', label='Validation val_loss')
plt.title('Training and Validation loss')
plt.legend()
plt.show()


# 7.使曲线变得平滑(将每个损失和精度都替换为指数移动平均值，从而让曲线变得平滑)
def smooth_curve(points, factor=0.8):
    smoothed_points = []
    for point in points:
        if smoothed_points:
            previous = smoothed_points[-1]
            smoothed_points.append(previous * factor + point * (1 - factor))
        else:
            smoothed_points.append(point)
    return smoothed_points

plt.plot(epochs, smooth_curve(acc), 'r', label='Smoothed training acc')
plt.plot(epochs, smooth_curve(val_acc), 'b', label='Smoothed validation acc')
plt.title('Training and Validation accurary')
plt.legend()
plt.figure()

plt.plot(epochs, smooth_curve(loss), 'r', label='Smoothed training loss')
plt.plot(epochs, smooth_curve(val_acc), 'b', label='Smoothed validation loss')
plt.title('Training and Validation loss')
plt.legend()
plt.show()


# 8.在测试数据上评估模型
test_generator = test_datagen.flow_from_directory(
    test_dir,
    target_size=(150, 150),
    batch_size=20,
    class_mode='binary')

test_loss, test_acc = model.evaluate_generator(test_generator, steps=50)   # 评估测试集
print('test acc:', test_acc)