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:

  def create_base_network(input_dim):
      seq = Sequential()
      seq.add(Conv2D(64, 5, activation='relu', padding='same', name='conv1', input_shape=input_dim))
      seq.add(Conv2D(128, 4, activation='relu', padding='same', name='conv2'))
      seq.add(Conv2D(256, 4, activation='relu', padding='same', name='conv3'))
      seq.add(Conv2D(64, 1, activation='relu', padding='same', name='conv4'))
      seq.add(MaxPooling2D(2, 2, name='pool1'))
      seq.add(Flatten(name='fla1'))
      seq.add(Dense(512, activation='relu', name='dense1'))
      seq.add(Reshape((1, 512), name='reshape'))

整体代码,该模型存在多个输入(6个):

	def create_base_network(input_dim):
        seq = Sequential()
        seq.add(Conv2D(64, 5, activation='relu', padding='same', name='conv1', input_shape=input_dim))
        seq.add(Conv2D(128, 4, activation='relu', padding='same', name='conv2'))
        seq.add(Conv2D(256, 4, activation='relu', padding='same', name='conv3'))
        seq.add(Conv2D(64, 1, activation='relu', padding='same', name='conv4'))
        seq.add(MaxPooling2D(2, 2, name='pool1'))
        seq.add(Flatten(name='fla1'))
        seq.add(Dense(512, activation='relu', name='dense1'))
        seq.add(Reshape((1, 512), name='reshape'))
           return seq
           
     base_network = create_base_network(img_size)
     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)
     print('the shape of base1:', base_network(input_1).shape)   # (, 1, 512)
     out_all = Concatenate(axis=1)([base_network(input_1), base_network(input_2), base_network(input_3), base_network(input_4), base_network(input_5), base_network(input_6)])
     print('****', out_all.shape)   # (, 6, 512)
     lstm_layer = LSTM(128, name = 'lstm')(out_all)

     out_puts = Dense(3, activation = 'softmax', name = 'out')(lstm_layer)
     model = Model([input_1,input_2,input_3,input_4,input_5,input_6], out_puts)
     model.summary()

网络模型:
关于Keras里的Sequential(序列模型)转化为Model(函数模型)的问题

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_base_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 = Dense(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)

    base_network_1 = create_base_network(input_1)
    base_network_2 = create_base_network(input_2)
    base_network_3 = create_base_network(input_3)
    base_network_4 = create_base_network(input_4)
    base_network_5 = create_base_network(input_5)
    base_network_6 = create_base_network(input_6)
    # print('the shape of base1:', base_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
        [base_network_1, base_network_2, base_network_3, base_network_4, base_network_5, base_network_6])
    print('****', out_all.shape)   # (, 6, 512)
    lstm_layer = LSTM(128, name = 'lstm')(out_all)

    out_puts = Dense(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()

结果模型输出如下:
关于Keras里的Sequential(序列模型)转化为Model(函数模型)的问题
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base_network = create_base_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:
base_network_1 = create_base_network(input_1)
...
base_network_6 = create_base_network(input_6)

前面定义的 def create_base_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:

# 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 = Dense(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)

base_network_1 = x8(x7(x6(x5(x4(x3(x2(x1(input_1))))))))
base_network_2 = x8(x7(x6(x5(x4(x3(x2(x1(input_2))))))))
base_network_3 = x8(x7(x6(x5(x4(x3(x2(x1(input_3))))))))
base_network_4 = x8(x7(x6(x5(x4(x3(x2(x1(input_4))))))))
base_network_5 = x8(x7(x6(x5(x4(x3(x2(x1(input_5))))))))
base_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
      [base_network_1, base_network_2, base_network_3, base_network_4, base_network_5, base_network_6])

# lstm layer
lstm_layer = LSTM(128, name = 'lstm3')(out_all)
# dense layer
out_layer = Dense(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()

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。。。

上一篇:文件版本管理方案设想


下一篇:快速幂