关于Keras ⾥的Sequential （序列模型）转化为Model （函数模型）的问题⽂章⽬录
前⾔
想在keras模型上加上注意⼒机制，于是把keras的序列模型转化为函数模型，结果发现参数维度不⼀致的问题，结果也变差了。跟踪问题后续发现是转为函数模型后，⽹络共享层出现了问题。
⼀、序列模型
该部分采⽤的是add添加⽹络层，由于存在多次重复调⽤相同⽹络层的情况，因此封装成⼀个⾃定义函数：
竹外桃花三两枝的意思整体代码，该模型存在多个输⼊（6个）： def create_ba_network (input_dim ): q = Sequential () q .add (Conv2D (64, 5, activation ='relu', padding ='same', name ='conv1', input_shape =input_dim )) q .add (Conv2D (128, 4, activation ='relu', padding ='same', name ='conv2')) q .add (Conv2D (256, 4, activation ='relu', padding ='same', name ='conv3')) q .add (Conv2D (64, 1, activation ='relu', padding ='same', name ='conv4')) q .add (MaxPooling2D (2, 2, name ='pool1')) q .add (Flatten (name ='fla1')) q .add (Den (512, activation ='relu', name ='den1')) q .add (Reshape ((1, 512), name ='reshape'))
1
2
3
4
5论语20篇
6世纪雷锋
7
8
9
10 model
⽹络模型：
哄男朋友睡觉的故事
⼆、改为函数模型
1.错误代码
第⼀次更改⽹络模型后，虽然运⾏未报错，但参数变多，模型性能也下降了，如下： def create_ba_network (input_dim ): x = Conv2D (64, 5, activation ='relu', padding ='same')(input
_dim ) x = Conv2D (128, 4, activation ='relu', padding ='same')(x ) x = Conv2D (256, 4, activation ='relu', padding ='same')(x ) x = Conv2D (64, 1, activation ='relu', padding ='same')(x ) x = MaxPooling2D (2, 2)(x ) x = Flatten ()(x ) x = Den (512, activation ='relu')(x ) x = Reshape ((1, 512))(x ) return x input_1 = Input (shape =img_size ) input_2 = Input (shape =img_size ) input_3 = Input (shape =img_size ) input_4 = Input (shape =img_size ) input_5 = Input (shape =img_size ) input_6 = Input (shape =img_size ) ba_network_1 = create_ba_network (input_1) ba_network_2 = create_ba_network (input_2) ba_network_3 = create_ba_network (input_3) ba_network_4 = create_ba_network (input_4) ba_network_5 = create_ba_network (input_5) ba_network_6 = create_ba_network (input_6) # print ('the shape of ba1:', ba_network (input_1).shape ) # (, 1, 512) out_all = Concatenate (axis = 1)( # 维度不变, 维度拼接，第⼀维度变为原来的6倍 [ba_network_1, ba_network_2, ba_network_3, ba_network_4, ba_network_5, ba_network_6]) print ('****', out_all .shape ) # (, 6, 512) lstm_layer = LSTM (128, name = 'lstm')(out_all ) out_puts = Den (3, activation = 'softmax', name = 'out')(lstm_layer ) model = Model (inputs = [input_1, input_2, input_3, input_4, input_5, input_6], outputs = out_puts ) # 6个输⼊ model .summary ()
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
结果模型输出如下：
可以看到，模型的参数变为了原来的6倍多，改了很多次，后来发现，原来是因为序列模型中的ba_network =
create_ba_network(img_size)相当于已将模型实例化成了⼀个model，后续调⽤时只传⼊参数，⽽不更改模型结构。
⽽改为Model API后：
怀孕能吃葡萄干吗ba_network_1 = create_ba_network(input_1)
...
ba_network_6 = create_ba_network(input_6)
前⾯定义的 def create_ba_network（ inputs），并未进⾏实例化，后续相当于创建了6次相关⽹络层，应该先实例化，应当改为以下部分：
总结
Keras⾥的函数模型，如果想要多个输⼊共享多个⽹络层， 还是得将各个层实例化，不能偷懒。。。#
建⽴⽹络共享层x1 = Conv2D (64, 5, activation = 'relu', padding = 'same', name = 'conv1')x2 = Conv2D (128, 4, activation = 'relu', padding = 'same', name = 'conv2')x3 = Conv2D (256, 4, activation = 'relu', padding = 'same', name = 'conv3')x4 = Conv2D (64, 1, activation = 'relu', padding = 'same', name = 'conv4')x5 = MaxPooling2D (2, 2)x6 = Flatten ()x7 = Den (512, activation = 'relu')x8 = Reshape ((1, 512))input_1 = Input (shape = img_size ) # 得到6个输⼊input_2 = Input (shape = img_size )input_3 = Input (shape = img_size )input_4 = Input (shape = img_size )input_5 = Input (shape = img_size )input_6 = Input (shape = img_size )ba_network_1 = x8(x7(x6(x5(x4(x3(x2(x1(input_1))))))))ba_network_2 = x8(x7(x6(x5(x4(x3(x2(x1(input_2))))))))ba_network_3 = x8(x7(x6(x5(x4(x3(x2(x1(input_3))))))))ba_network_4 = x8(x7(x6(x5(x4(x3(x2(x1(input_4))))))))ba_network_5 = x8(x7(x6(x5(x4(x3(x2(x1(input_5))))))))ba_network_6 = x8(x7(x6(x5(x4(x3(x2(x1(input_6)))))))) # 输⼊连接out_all = Concatenate (axis = 1)( # 维度不变, 维度拼接，第⼀维度变为原来的6倍 [ba_network_1, ba_network_2, ba_network_3, ba_network_4, ba_network_5, ba_network_6])# lstm layer lstm_layer = LSTM (128, name = 'lstm3')(out_all )# den layer out_layer = Den (3, activation = 'softmax', name = 'out')(lstm_layer )model = Model (inputs = [input_1, input_2, input_3, input_4, input_5, input_6], outputs = out_layer ) # 6个输⼊model .summary ()1
2华侨大学全国排名
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
好书读后感30
31
32
乐乘33
34
