【小白学习keras教程】八、Sequential Model和模型函数API两种模型建立方法

@Author：Runsen

@[toc]

Load dataset

• digits dataset in scikit-learn
• url: http://scikit-learn.org/stable/auto_examples/datasets/plot_digits_last_image.html

from sklearn.datasets import load_digits
from tensorflow.keras.utils import plot_model
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.layers import Input, Dense, concatenate, Activation
data = load_digits()
X_data = data.images
y_data = data.target
# flatten X_data
X_data = X_data.reshape(X_data.shape[0], X_data.shape[1]*X_data.shape[2])

• 在Keras中创建深度学习模型的最简单方法是使用顺序模型API

• 顺序模型API只允许按顺序连接层，

• doc: https://keras.io/getting-started/sequential-model-guide/

Sequential Model API

方法1

• 将层列表传递到Sequential（）中
• 模型小（即浅）时使用效率高

model = Sequential([Dense(10, input_shape = X_data.shape[1:]), Dense(10, activation = 'softmax')])
model.summary()

plot_model(model)

方法2

• 有时，如果模型越来越深，就很难在图层列表中丢失
• 在这种情况下，可以使用**add（）**函数逐行添加层，这样做，还可以一步一步地跟踪模型的输出形状

model = Sequential()
model.add(Dense(10, input_shape = X_data.shape[1:], activation = 'relu'))
print(model.output_shape)

(None, 10)

model.add(Dense(50, activation = 'relu'))
model.add(Dense(50, activation = 'relu'))
print(model.output_shape)

(None, 50)

model.add(Dense(10, activation = 'sigmoid'))
print(model.output_shape)

(None, 10)

model.summary()

plot_model(model)

• 可以通过设置图层的“名称”来跟踪图层

model = Sequential()
model.add(Dense(10, input_shape = X_data.shape[1:], activation = 'relu', name = 'Input_layer'))
model.add(Dense(50, activation = 'relu', name = 'First_hidden_layer'))
model.add(Dense(10, activation = 'softmax', name = 'Output_layer'))
model.summary()

plot_model(model)

模型函数API

• 通过顺序API创建模型简单易行，但不可能创建复杂的模型结构
• 例如，初始或剩余网络结构不可能使用顺序API实现，因为它们需要层合并和多个输出等操作
• 在这种情况下，可以利用函数API
• 通过定义输入和输出创建模型

单输入输出

• 只有单一输入和输出的模型
• 这种结构也可以使用顺序API创建

# creating layers
input_layer = Input(shape = X_data.shape[1:])
activation_1 = Activation('relu')(input_layer)
hidden_layer = Dense(50)(activation_1)
activation_2 = Activation('relu')(hidden_layer)
output_layer = Dense(10, activation = 'softmax')(activation_2)
# creating model
model = Model(inputs = input_layer, outputs = output_layer)
model.summary()

plot_model(model)

合并图层

• 有时，需要合并层（例如，GoogleNet或ResNet）

1.连接

• concatenate()简单地合并两个或更多层的结果
• 例如，假设有两个层要连接，其结果是
  [x1，x2，…，xn]和[y1，y2，…，yn]。然后，连接层将是[x1，…，xn，…，y1，…，yn]

# creating layers
input_layer = Input(shape = X_data.shape[1:])
activation_1 = Activation('relu')(input_layer)
hidden_layer_1 = Dense(50, activation = 'relu')(activation_1)
hidden_layer_2 = Dense(50, activation = 'relu')(activation_1)
concat_layer = concatenate([hidden_layer_1, hidden_layer_2])
print(hidden_layer_1.shape)
print(hidden_layer_2.shape)
print(concat_layer.shape)

(None, 50)
(None, 50)
(None, 100)

model = Model(inputs = input_layer, outputs = concat_layer)
plot_model(model)

2. add, subtract, multiply, average, maximum

• 这些层对两个或更多层的所有对应元素执行元素操作

• 因此，保持了输入层的维数

# creating layers
input_layer = Input(shape = X_data.shape[1:])
activation_1 = Activation('relu')(input_layer)
hidden_layer_1 = Dense(50, activation = 'relu')(activation_1)
hidden_layer_2 = Dense(50, activation = 'relu')(activation_1)
add_layer = add([hidden_layer_1, hidden_layer_2])
print(hidden_layer_1.shape)
print(hidden_layer_2.shape)
print(add_layer.shape)

(None, 50)
(None, 50)
(None, 50)

model = Model(inputs = input_layer, outputs = add_layer)
plot_model(model)

3. dot

• dot()在两层结果之间执行内积运算

• 应定义“轴”来执行操作

# creating layers
input_layer = Input(shape = X_data.shape[1:])
activation_1 = Activation('relu')(input_layer)
hidden_layer_1 = Dense(50, activation = 'relu')(activation_1)
hidden_layer_2 = Dense(50, activation = 'relu')(activation_1)
dot_layer = dot([hidden_layer_1, hidden_layer_2], axes = -1)
print(hidden_layer_1.shape)
print(hidden_layer_2.shape)
print(dot_layer.shape)

(None, 50)
(None, 50)
(None, 1)

model = Model(inputs = input_layer, outputs = dot_layer)
plot_model(model)