您当前的位置:首页 > IT编程 > Keras
| C语言 | Java | VB | VC | python | Android | TensorFlow | C++ | oracle | 学术与代码 | cnn卷积神经网络 | gnn | 图像修复 | Keras | 数据集 | Neo4j | 自然语言处理 | 深度学习 | 医学CAD | 医学影像 | 超参数 | pointnet |

keras根据层名称来初始化网络

51自学网 2021-12-02 12:11:38
  Keras

keras根据层名称来初始化网络

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
def get_model(input_shape1=[75, 75, 3], input_shape2=[1], weights=None):
 bn_model = 0
 trainable = True
 # kernel_regularizer = regularizers.l2(1e-4)
 kernel_regularizer = None
 activation = 'relu'
 
 img_input = Input(shape=input_shape1)
 angle_input = Input(shape=input_shape2)
 
 # Block 1
 x = Conv2D(64, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block1_conv1')(img_input)
 x = Conv2D(64, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block1_conv2')(x)
 x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
 
 # Block 2
 x = Conv2D(128, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block2_conv1')(x)
 x = Conv2D(128, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block2_conv2')(x)
 x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
 
 # Block 3
 x = Conv2D(256, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block3_conv1')(x)
 x = Conv2D(256, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block3_conv2')(x)
 x = Conv2D(256, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block3_conv3')(x)
 x = MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
 
 # Block 4
 x = Conv2D(512, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block4_conv1')(x)
 x = Conv2D(512, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block4_conv2')(x)
 x = Conv2D(512, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block4_conv3')(x)
 x = MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
 
 # Block 5
 x = Conv2D(512, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block5_conv1')(x)
 x = Conv2D(512, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block5_conv2')(x)
 x = Conv2D(512, (3, 3), activation=activation, padding='same',
    trainable=trainable, kernel_regularizer=kernel_regularizer,
    name='block5_conv3')(x)
 x = MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x)
 
 branch_1 = GlobalMaxPooling2D()(x)
 # branch_1 = BatchNormalization(momentum=bn_model)(branch_1)
 
 branch_2 = GlobalAveragePooling2D()(x)
 # branch_2 = BatchNormalization(momentum=bn_model)(branch_2)
 
 branch_3 = BatchNormalization(momentum=bn_model)(angle_input)
 
 x = (Concatenate()([branch_1, branch_2, branch_3]))
 x = Dense(1024, activation=activation, kernel_regularizer=kernel_regularizer)(x)
 # x = Dropout(0.5)(x)
 x = Dense(1024, activation=activation, kernel_regularizer=kernel_regularizer)(x)
 x = Dropout(0.6)(x)
 output = Dense(1, activation='sigmoid')(x)
 
 model = Model([img_input, angle_input], output)
 optimizer = Adam(lr=1e-5, beta_1=0.9, beta_2=0.999, epsilon=1e-8, decay=0.0)
 model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])
 
 if weights is not None:
  # 将by_name设置成True
  model.load_weights(weights, by_name=True)
  # layer_weights = h5py.File(weights, 'r')
  # for idx in range(len(model.layers)):
  #  model.set_weights()
 print 'have prepared the model.'
 
 return model

补充知识:keras.layers.Dense()方法

keras.layers.Dense()是定义网络层的基本方法,执行的操作是:output = activation(dot(input,kernel)+ bias。

其中activation是激活函数,kernel是权重矩阵,bias是偏向量。如果层输入大于2,在进行初始点积之前会将其展平。

代码如下:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
class Dense(Layer):
 """Just your regular densely-connected NN layer.
 `Dense` implements the operation:
 `output = activation(dot(input, kernel) + bias)`
 where `activation` is the element-wise activation function
 passed as the `activation` argument, `kernel` is a weights matrix
 created by the layer, and `bias` is a bias vector created by the layer
 (only applicable if `use_bias` is `True`).
 Note: if the input to the layer has a rank greater than 2, then
 it is flattened prior to the initial dot product with `kernel`.
 # Example
 ```python
  # as first layer in a sequential model:
  model = Sequential()
  model.add(Dense(32, input_shape=(16,)))
  # now the model will take as input arrays of shape (*, 16)
  # and output arrays of shape (*, 32)
  # after the first layer, you don't need to specify
  # the size of the input anymore:
  model.add(Dense(32))
 ```
 # Arguments
  units: Positive integer, dimensionality of the output space.
  activation: Activation function to use
   (see [activations](../activations.md)).
   If you don't specify anything, no activation is applied
   (ie. "linear" activation: `a(x) = x`).
  use_bias: Boolean, whether the layer uses a bias vector.
  kernel_initializer: Initializer for the `kernel` weights matrix
   (see [initializers](../initializers.md)).
  bias_initializer: Initializer for the bias vector
   (see [initializers](../initializers.md)).
  kernel_regularizer: Regularizer function applied to
   the `kernel` weights matrix
   (see [regularizer](../regularizers.md)).
  bias_regularizer: Regularizer function applied to the bias vector
   (see [regularizer](../regularizers.md)).
  activity_regularizer: Regularizer function applied to
   the output of the layer (its "activation").
   (see [regularizer](../regularizers.md)).
  kernel_constraint: Constraint function applied to
   the `kernel` weights matrix
   (see [constraints](../constraints.md)).
  bias_constraint: Constraint function applied to the bias vector
   (see [constraints](../constraints.md)).
 # Input shape
  nD tensor with shape: `(batch_size, ..., input_dim)`.
  The most common situation would be
  a 2D input with shape `(batch_size, input_dim)`.
 # Output shape
  nD tensor with shape: `(batch_size, ..., units)`.
  For instance, for a 2D input with shape `(batch_size, input_dim)`,
  the output would have shape `(batch_size, units)`.
 """
  
 @interfaces.legacy_dense_support
 def __init__(self, units,
     activation=None,
     use_bias=True,
     kernel_initializer='glorot_uniform',
     bias_initializer='zeros',
     kernel_regularizer=None,
     bias_regularizer=None,
     activity_regularizer=None,
     kernel_constraint=None,
     bias_constraint=None,
     **kwargs):
  if 'input_shape' not in kwargs and 'input_dim' in kwargs:
   kwargs['input_shape'] = (kwargs.pop('input_dim'),)
  super(Dense, self).__init__(**kwargs)
  self.units = units
  self.activation = activations.get(activation)
  self.use_bias = use_bias
  self.kernel_initializer = initializers.get(kernel_initializer)
  self.bias_initializer = initializers.get(bias_initializer)
  self.kernel_regularizer = regularizers.get(kernel_regularizer)
  self.bias_regularizer = regularizers.get(bias_regularizer)
  self.activity_regularizer = regularizers.get(activity_regularizer)
  self.kernel_constraint = constraints.get(kernel_constraint)
  self.bias_constraint = constraints.get(bias_constraint)
  self.input_spec = InputSpec(min_ndim=2)
  self.supports_masking = True
  
 def build(self, input_shape):
  assert len(input_shape) >= 2
  input_dim = input_shape[-1]
  
  self.kernel = self.add_weight(shape=(input_dim, self.units),
          initializer=self.kernel_initializer,
          name='kernel',
          regularizer=self.kernel_regularizer,
          constraint=self.kernel_constraint)
  if self.use_bias:
   self.bias = self.add_weight(shape=(self.units,),
          initializer=self.bias_initializer,
          name='bias',
          regularizer=self.bias_regularizer,
          constraint=self.bias_constraint)
  else:
   self.bias = None
  self.input_spec = InputSpec(min_ndim=2, axes={-1: input_dim})
  self.built = True
  
 def call(self, inputs):
  output = K.dot(inputs, self.kernel)
  if self.use_bias:
   output = K.bias_add(output, self.bias)
  if self.activation is not None:
   output = self.activation(output)
  return output
  
 def compute_output_shape(self, input_shape):
  assert input_shape and len(input_shape) >= 2
  assert input_shape[-1]
  output_shape = list(input_shape)
  output_shape[-1] = self.units
  return tuple(output_shape)
  
 def get_config(self):
  config = {
   'units': self.units,
   'activation': activations.serialize(self.activation),
   'use_bias': self.use_bias,
   'kernel_initializer': initializers.serialize(self.kernel_initializer),
   'bias_initializer': initializers.serialize(self.bias_initializer),
   'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),
   'bias_regularizer': regularizers.serialize(self.bias_regularizer),
   'activity_regularizer': regularizers.serialize(self.activity_regularizer),
   'kernel_constraint': constraints.serialize(self.kernel_constraint),
   'bias_constraint': constraints.serialize(self.bias_constraint)
  }
  base_config = super(Dense, self).get_config()
  return dict(list(base_config.items()) + list(config.items()))

参数说明如下:

units:正整数,输出空间的维数。

activation: 激活函数。如果未指定任何内容,则不会应用任何激活函数。即“线性”激活:a(x)= x)。

use_bias:Boolean,该层是否使用偏向量。

kernel_initializer:权重矩阵的初始化方法。

bias_initializer:偏向量的初始化方法。

kernel_regularizer:权重矩阵的正则化方法。

bias_regularizer:偏向量的正则化方法。

activity_regularizer:输出层正则化方法。

kernel_constraint:权重矩阵约束函数。

bias_constraint:偏向量约束函数。

以上这篇使用keras根据层名称来初始化网络就是小编分享给大家的全部内容


ImageDataGenerator参数详解及用法
Python layers.ActivityRegularization方法代码示例
51自学网,即我要自学网,自学EXCEL、自学PS、自学CAD、自学C语言、自学css3实例,是一个通过网络自主学习工作技能的自学平台,网友喜欢的软件自学网站。
京ICP备13026421号-1