keras提供了VGG19在ImageNet上的预训练权重模型文件，其他可用的模型还有VGG16、Xception、ResNet50、InceptionV3 4个。

VGG19在keras中的定义：

1.  
   def VGG19(include_top=True, weights='imagenet',
2.  
   input_tensor=None, input_shape=None,
3.  
   pooling=None,
4.  
   classes=1000)
5.  
    

参数介绍：

include_top: 是否包含最后的3个全连接层

weights： 定义为‘imagenet’，表示加载在imagenet数据库上训练的预训练权重，定义为None则不加载权重，参数随机初始化

1）使用VGG19预训练模型分类图片的例子

1.  
   # coding: utf-8
2.  
   from keras.applications.vgg19 import VGG19
3.  
   from keras.preprocessing import image
4.  
   from keras.applications.vgg19 import preprocess_input
5.  
   from keras.models import Model
6.  
   import numpy as np
7.  
    
8.  
   base_model = VGG19(weights='imagenet', include_top=True)
9.  
   img_path = 'cat.jpg'
10.  
    img = image.load_img(img_path, target_size=(224, 224)) # 加载图像，归一化大小
11.  
    x = image.img_to_array(img) # 序列化
12.  
    x = np.expand_dims(x, axis=0) # 扩展维度[batch_size,H,W,Channel]
13.  
    x = preprocess_input(x) # 预处理，默认mode="caffe"
14.  
    """
15.  
    def preprocess_input(x, data_format=None, mode='caffe'):
16.  
    Preprocesses a tensor or Numpy array encoding a batch of images.
17.  
     
18.  
    # Arguments
19.  
    x: Input Numpy or symbolic tensor, 3D or 4D.
20.  
    data_format: Data format of the image tensor/array.
21.  
    mode: One of "caffe", "tf".
22.  
    - caffe: will convert the images from RGB to BGR,
23.  
    then will zero-center each color channel with
24.  
    respect to the ImageNet dataset,
25.  
    without scaling.
26.  
    - tf: will scale pixels between -1 and 1,
27.  
    sample-wise.
28.  
     
29.  
    # Returns
30.  
    Preprocessed tensor or Numpy array.
31.  
     
32.  
    # Raises
33.  
    ValueError: In case of unknown `data_format` argument.
34.  
    """
35.  
    out = base_model.predict(x) # 预测结果，1000维的向量
36.  
    print(out.shape) # 预测结果out是一个1000维的向量，代表了预测结果分别属于10000个分类的概率，形状是（1,1000）

2) 使用VGG19预训练模型提取VGG19网络中任意层的输出特征的例子

上个例子可以看到keras对VGG网络的封装异常好，简单几行代码就可以分类图片。keras中VGG预训练参数模型另一个更常用的应用是可以提取VGG网络中任意一层的特征。
以下例子提取的是VGG19网络中第5个卷积层的输出特征（也是第1个全连接层的输入特征）

1.  
   # coding: utf-8
2.  
   from keras.applications.vgg19 import VGG19
3.  
   from keras.preprocessing import image
4.  
   from keras.applications.vgg19 import preprocess_input
5.  
   from keras.models import Model
6.  
   import numpy as np
7.  
    
8.  
   base_model = VGG19(weights='imagenet', include_top=False)
9.  
   model = Model(inputs=base_model.input, outputs=base_model.get_layer('block5_pool').output)
10.  
    img_path = 'cat.jpg'
11.  
    img = image.load_img(img_path, target_size=(224, 224))
12.  
    x = image.img_to_array(img)
13.  
    x = np.expand_dims(x, axis=0)
14.  
    x = preprocess_input(x)
15.  
    block5_pool_features = model.predict(x)
16.  
    print(block5_pool_features.shape) #(1, 7, 7, 512)

也可以设置为加载最后3个全连接层的VGG19网络，就可以获取最后3个全连接层的输出了：

1.  
   # coding: utf-8
2.  
   from keras.applications.vgg19 import VGG19
3.  
   from keras.preprocessing import image
4.  
   from keras.applications.vgg19 import preprocess_input
5.  
   from keras.models import Model
6.  
   import numpy as np
7.  
   base_model = VGG19(weights='imagenet', include_top=True)
8.  
   model = Model(inputs=base_model.input, outputs=base_model.get_layer('fc2').output)
9.  
   img_path = 'cat.jpg'
10.  
    img = image.load_img(img_path, target_size=(224, 224))
11.  
    x = image.img_to_array(img)
12.  
    x = np.expand_dims(x, axis=0)
13.  
    x = preprocess_input(x)
14.  
    fc2 = model.predict(x)
15.  
    print(fc2.shape) #(1, 4096)

加了全连接层，所以base_model.get_layer('fc2') 里参数也可以是 flatten、fc1、fc2和predictions 。

VGG19各个模块在keras中定义的名称可以通过以下代码获得，然后根据名称可以轻松获取该层特征：：

1.  
   from keras.applications.vgg19 import VGG19
2.  
   from keras.preprocessing import image
3.  
   from keras.applications.vgg19 import preprocess_input
4.  
   from keras.models import Model
5.  
   import numpy as np
6.  
   import os
7.  
   import tensorflow as tf
8.  
   os.environ["CUDA_VISIBLE_DEVICES"] = "0"
9.  
   gpu_options = tf.GPUOptions(allow_growth=True)
10.  
    sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
11.  
    from keras.utils import plot_model
12.  
    from matplotlib import pyplot as plt
13.  
     
14.  
    #【0】VGG19模型，加载预训练权重
15.  
    base_model = VGG19(weights='imagenet')
16.  
     
17.  
    #【1】创建一个新model, 使得它的输出(outputs)是 VGG19 中任意层的输出(output)
18.  
    model = Model(inputs=base_model.input, outputs=base_model.get_layer('block4_pool').output)
19.  
    print(model.summary()) # 打印模型概况,如下所示
20.  
    plot_model(model,to_file = 'a simple convnet.png') # 画出模型结构图，并保存成图片，如下图所示
21.  
    """
22.  
    _________________________________________________________________
23.  
    Layer (type) Output Shape Param #
24.  
    =================================================================
25.  
    input_1 (InputLayer) (None, 224, 224, 3) 0
26.  
    _________________________________________________________________
27.  
    block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
28.  
    _________________________________________________________________
29.  
    block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
30.  
    _________________________________________________________________
31.  
    block1_pool (MaxPooling2D) (None, 112, 112, 64) 0
32.  
    _________________________________________________________________
33.  
    block2_conv1 (Conv2D) (None, 112, 112, 128) 73856
34.  
    _________________________________________________________________
35.  
    block2_conv2 (Conv2D) (None, 112, 112, 128) 147584
36.  
    _________________________________________________________________
37.  
    block2_pool (MaxPooling2D) (None, 56, 56, 128) 0
38.  
    _________________________________________________________________
39.  
    block3_conv1 (Conv2D) (None, 56, 56, 256) 295168
40.  
    _________________________________________________________________
41.  
    block3_conv2 (Conv2D) (None, 56, 56, 256) 590080
42.  
    _________________________________________________________________
43.  
    block3_conv3 (Conv2D) (None, 56, 56, 256) 590080
44.  
    _________________________________________________________________
45.  
    block3_conv4 (Conv2D) (None, 56, 56, 256) 590080
46.  
    _________________________________________________________________
47.  
    block3_pool (MaxPooling2D) (None, 28, 28, 256) 0
48.  
    _________________________________________________________________
49.  
    block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160
50.  
    _________________________________________________________________
51.  
    block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808
52.  
    _________________________________________________________________
53.  
    block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808
54.  
    _________________________________________________________________
55.  
    block4_conv4 (Conv2D) (None, 28, 28, 512) 2359808
56.  
    _________________________________________________________________
57.  
    block4_pool (MaxPooling2D) (None, 14, 14, 512) 0
58.  
    =================================================================
59.  
    Total params: 10,585,152
60.  
    Trainable params: 10,585,152
61.  
    Non-trainable params: 0
62.  
    _________________________________________________________________
63.  
     
64.  
    """

 

3）最终的特征提取代码

1.  
   # -*- coding: UTF-8 -*-
2.  
   #-------------------------------------------
3.  
   #任 务：利用VGG19提取任意中间层特征
4.  
   #数 据：网上下载的测试图片‘elephant.jpg’
5.  
   #-------------------------------------------
6.  
    
7.  
   from keras.applications.vgg19 import VGG19
8.  
   from keras.preprocessing import image
9.  
   from keras.applications.vgg19 import preprocess_input
10.  
    from keras.models import Model
11.  
    import numpy as np
12.  
    import os
13.  
    import tensorflow as tf
14.  
    os.environ["CUDA_VISIBLE_DEVICES"] = "6"
15.  
    gpu_options = tf.GPUOptions(allow_growth=True)
16.  
    sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
17.  
    from keras.utils import plot_model
18.  
    from matplotlib import pyplot as plt
19.  
     
20.  
    #【0】VGG19模型，加载预训练权重
21.  
    base_model = VGG19(weights='imagenet')
22.  
     
23.  
    #【1】创建一个新model, 使得它的输出(outputs)是 VGG19 中任意层的输出(output)
24.  
    model = Model(inputs=base_model.input, outputs=base_model.get_layer('block4_pool').output)
25.  
    print(model.summary()) # 打印模型概况
26.  
    plot_model(model,to_file = 'a simple convnet.png') # 画出模型结构图，并保存成图片
27.  
     
28.  
    '''
29.  
    _________________________________________________________________
30.  
    Layer (type) Output Shape Param #
31.  
    =================================================================
32.  
    input_1 (InputLayer) (None, 224, 224, 3) 0
33.  
    _________________________________________________________________
34.  
    block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
35.  
    _________________________________________________________________
36.  
    block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
37.  
    _________________________________________________________________
38.  
    block1_pool (MaxPooling2D) (None, 112, 112, 64) 0
39.  
    _________________________________________________________________
40.  
    block2_conv1 (Conv2D) (None, 112, 112, 128) 73856
41.  
    _________________________________________________________________
42.  
    block2_conv2 (Conv2D) (None, 112, 112, 128) 147584
43.  
    _________________________________________________________________
44.  
    block2_pool (MaxPooling2D) (None, 56, 56, 128) 0
45.  
    _________________________________________________________________
46.  
    block3_conv1 (Conv2D) (None, 56, 56, 256) 295168
47.  
    _________________________________________________________________
48.  
    block3_conv2 (Conv2D) (None, 56, 56, 256) 590080
49.  
    _________________________________________________________________
50.  
    block3_conv3 (Conv2D) (None, 56, 56, 256) 590080
51.  
    _________________________________________________________________
52.  
    block3_conv4 (Conv2D) (None, 56, 56, 256) 590080
53.  
    _________________________________________________________________
54.  
    block3_pool (MaxPooling2D) (None, 28, 28, 256) 0
55.  
    _________________________________________________________________
56.  
    block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160
57.  
    _________________________________________________________________
58.  
    block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808
59.  
    _________________________________________________________________
60.  
    block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808
61.  
    _________________________________________________________________
62.  
    block4_conv4 (Conv2D) (None, 28, 28, 512) 2359808
63.  
    _________________________________________________________________
64.  
    block4_pool (MaxPooling2D) (None, 14, 14, 512) 0
65.  
    =================================================================
66.  
    Total params: 10,585,152
67.  
    Trainable params: 10,585,152
68.  
    Non-trainable params: 0
69.  
    _________________________________________________________________
70.  
     
71.  
    '''
72.  
     
73.  
    #【2】从网上下载一张图片，保存在当前路径下
74.  
    img_path = 'elephant.jpg'
75.  
    img = image.load_img(img_path, target_size=(224, 224)) # 加载图片并resize成224x224
76.  
     
77.  
    #【3】将图片转化为4d tensor形式
78.  
    x = image.img_to_array(img)
79.  
    x = np.expand_dims(x, axis=0)
80.  
     
81.  
    #【4】数据预处理
82.  
    x = preprocess_input(x) #去均值中心化，preprocess_input函数详细功能见注释
83.  
    """
84.  
    def preprocess_input(x, data_format=None, mode='caffe'):
85.  
    Preprocesses a tensor or Numpy array encoding a batch of images.
86.  
     
87.  
    # Arguments
88.  
    x: Input Numpy or symbolic tensor, 3D or 4D.
89.  
    data_format: Data format of the image tensor/array.
90.  
    mode: One of "caffe", "tf".
91.  
    - caffe: will convert the images from RGB to BGR,
92.  
    then will zero-center each color channel with
93.  
    respect to the ImageNet dataset,
94.  
    without scaling.
95.  
    - tf: will scale pixels between -1 and 1,
96.  
    sample-wise.
97.  
     
98.  
    # Returns
99.  
    Preprocessed tensor or Numpy array.
100.  
      
101.  
     # Raises
102.  
     ValueError: In case of unknown `data_format` argument.
103.  
     """
104.  
     #【5】提取特征
105.  
     block4_pool_features = model.predict(x)
106.  
     print(block4_pool_features.shape) #(1, 14, 14, 512)

示例图片：