keras 拼接两个函数api 为修改backbone并且使用其权重做准备

约定不等于承诺〃 2022-02-22 15:40 170阅读 0赞

这个代码网络是model2，他用了我之前写的model\_mine的网络结构。

from keras.models import Model
    from keras.layers import Input, Dense, Lambda,add
    
    a = Input(shape=(640, 480, 3))
    b = Input(shape=(23,))
    c = Input(shape=(54,))
    d = Dense(32)(b)
    multi = Lambda(lambda x: x**3)
    e = multi(c)
    
    model_mine = Model(inputs=[a, b, c], outputs=[d, e])
    in1 = Input(shape=(23,))
    in2 = Input(shape=(128,))
    #compile(optimizer, loss=None, metrics=None, loss_weights=None, sample_weight_mode=None, weighted_metrics=None, target_tensors=None)
    [out1, out2] = model_mine(inputs=[a, in1, in2])
    lay23=Dense(32)(out2)
    
    out3=add([out1, lay23])
    model2 = Model(inputs=[a, in1, in2], outputs=[out3])
    print('print model_mine structure')
    model_mine.summary()
    print('print model2 structure')
    model2.summary()

输出是

print model_mine structure
    __________________________________________________________________________________________________
    Layer (type)                    Output Shape         Param #     Connected to                     
    ==================================================================================================
    input_2 (InputLayer)            (None, 23)           0                                            
    __________________________________________________________________________________________________
    input_3 (InputLayer)            (None, 54)           0                                            
    __________________________________________________________________________________________________
    dense_1 (Dense)                 (None, 32)           768         input_2[0][0]                    
    __________________________________________________________________________________________________
    lambda_1 (Lambda)               (None, 54)           0           input_3[0][0]                    
    ==================================================================================================
    Total params: 768
    Trainable params: 768
    Non-trainable params: 0
    __________________________________________________________________________________________________
    print model2 structure
    __________________________________________________________________________________________________
    Layer (type)                    Output Shape         Param #     Connected to                     
    ==================================================================================================
    input_1 (InputLayer)            (None, 640, 480, 3)  0                                            
    __________________________________________________________________________________________________
    input_4 (InputLayer)            (None, 23)           0                                            
    __________________________________________________________________________________________________
    input_5 (InputLayer)            (None, 128)          0                                            
    __________________________________________________________________________________________________
    model_1 (Model)                 multiple             768         input_1[0][0]                    
                                                                     input_4[0][0]                    
                                                                     input_5[0][0]                    
    __________________________________________________________________________________________________
    dense_2 (Dense)                 (None, 32)           4128        model_1[1][1]                    
    __________________________________________________________________________________________________
    add_1 (Add)                     (None, 32)           0           model_1[1][0]                    
                                                                     dense_2[0][0]                    
    ==================================================================================================
    Total params: 4,896
    Trainable params: 4,896
    Non-trainable params: 0
    __________________________________________________________________________________________________
    
    Process finished with exit code 0

那么简单的解释下keras的机制，你就明白了，你建立一个模型后，比如这个model\_mine，其实它本质上就是输入输出，中间的复杂的网络你可以把它当作一个黑盒子，不需要知道。你要是使用它，你就把它当作keras的一个layer就好了。

所以他的网络嫁接非常简单。只需要把网络当成函数，然后输入数出即可。超简单好用的。

但是这里要注意，有个错误可能初学者（比如我）会犯，错误代码：

from keras.models import Model
    from keras.layers import Input, Dense, Lambda,add
    
    a = Input(shape=(640, 480, 3))
    b = Input(shape=(23,))
    c = Input(shape=(54,))
    d = Dense(32)(b)
    multi = Lambda(lambda x: x**3)
    e = multi(c)
    
    model_mine = Model(inputs=[a, b, c], outputs=[d, e])
    #in1 = Input(shape=(23,))
    #in2 = Input(shape=(128,))
    #compile(optimizer, loss=None, metrics=None, loss_weights=None, sample_weight_mode=None, weighted_metrics=None, target_tensors=None)
    [out1, out2] = model_mine(inputs=[a, b, c])
    lay23=Dense(32)(out2)
    
    out3=add([out1, lay23])
    model2 = Model(inputs=[a, out1, out2], outputs=[out3])
    print('print model_mine structure')
    model_mine.summary()
    print('print model2 structure')
    model2.summary()

他会提醒你

UserWarning: Model inputs must come from \`keras.layers.Input\` (thus holding past layer metadata), they cannot be the output of a previous non-Input layer. Here, a tensor specified as input to your model was not an Input tensor, it was generated by layer model\_1.  
Note that input tensors are instantiated via \`tensor = keras.layers.Input(shape)\`.

这什么意思呢。是说，你的模型建立时的输入不能是别的网络的输出，必须是Input()定义的，那你就会奇怪了，那我就是想用别的网络作为我的头部，你咋不让呢。

model这个实例的机制就是你得告诉他从哪里开始(inputs 参数)，从哪里结束(outputs参数)，你还要用layer建立两者中间的关系，让model能够从头顺到尾。

所以正确的代码是，使用Input定义真正的头部，然后使用已经建立的model\_mine处理我的输入tensor，让model\_mine在新的网络中成为一个函数，而不是实例，他和model2的地位是不对等的（想法是这样，但是底层代码大家都是CLASS）。