合并具有不同输入形状的不同模型的输出

时间:2017-02-24 17:56:45

标签: machine-learning keras keras-layer

我是Keras的新手。我正在尝试合并Keras中三个预训练模型的输出层。每个模型都有两个独立的输入,但具有不同的尺寸,以及Dense图层输出。

    model1 = MyModel1() #returns keras.engine.training.Model
    model2 = MyModel2() #returns keras.engine.training.Model
    model3 = MyModel3() #returns keras.engine.training.Model

    x = merge([model1(model1.input),
               model2(model2.input),
               model3(model3.input)],
              mode='concat', concat_axis=1)

    # add some trainable layers here...

    # and a final softmax layer
    x = Dense(2, activation='softmax')(x)

    return Model(input=[model1.input,
                        model2.input,
                        model3.input],
                 output=x)

由于model?.input返回一个Tensors列表,这不起作用。我尝试了不同的东西,似乎什么都没有用。这个问题有一个简单的解决方案吗?

修改 来自indraforyou的改进工作解决方案,用于每个模型中的多个输入。

    from keras.models import Model
    from keras.layers import Input, Dense, merge


    def MyModel1():
        inp1 = Input(batch_shape=(None,32,))
        inp2 = Input(batch_shape=(None,32))
        x = Dense(8)(inp1)
        y = Dense(8)(inp2)
        merged = merge([x, y], mode='concat', concat_axis=-1)
        out = Dense(8)(merged)
        return Model(input=[inp1,inp2], output=out)

    def MyModel2():
        inp1 = Input(batch_shape=(None,10,))
        inp2 = Input(batch_shape=(None,10,))
        x = Dense(4)(inp1)
        y = Dense(4)(inp2)
        merged = merge([x, y], mode='concat', concat_axis=-1)
        out = Dense(4)(merged)
        return Model(input=[inp1,inp2], output=out)

    def MyModel3():
        inp1 = Input(batch_shape=(None,12,))
        inp2 = Input(batch_shape=(None,12,))
        x = Dense(6)(inp1)
        y = Dense(6)(inp1)
        merged = merge([x, y], mode='concat', concat_axis=-1)
        out = Dense(6)(merged)
        return Model(input=[inp1,inp2], output=out)

    model1 = MyModel1()
    model2 = MyModel2()
    model3 = MyModel3()

    x = merge([model1.output,
               model2.output,
               model3.output],
              mode='concat', concat_axis=-1)

    x = Dense(2, activation='softmax')(x)

    merged =  Model(input=[model1.input[0], model1.input[1],
                           model2.input[0], model2.input[1],
                           model3.input[0], model3.input[1]],
                    output=x)

    merged.summary()

1 个答案:

答案 0 :(得分:4)

models对象不是可调用函数。这应解决问题:

x = merge([model1.output,
           model2.output,
           model3.output],
          mode='concat', concat_axis=1)

更新工作代码

from keras.models import Model
from keras.layers import Input, Dense, merge


def MyModel1():
  inp = Input(batch_shape=(None,32,))
  out = Dense(8)(inp)
  return Model(input=inp, output=out)

def MyModel2():
  inp = Input(batch_shape=(None,10,))
  out = Dense(4)(inp)
  return Model(input=inp, output=out)

def MyModel3():
  inp = Input(batch_shape=(None,12,))
  out = Dense(6)(inp)
  return Model(input=inp, output=out)

model1 = MyModel1()
model2 = MyModel2()
model3 = MyModel3()

x = merge([model1.output,
           model2.output,
           model3.output],
          mode='concat', concat_axis=1)

x = Dense(2, activation='softmax')(x)

merged =  Model(input=[model1.input,
                    model2.input,
                    model3.input],
             output=x)

merged.summary()