使用训练好的模型层在keras中创建另一个模型

Question

我用Keras创建了一个模型，如下所示：

    m = Sequential()
    m.add(Dense(912, activation='relu', input_shape=(943, 1)))
    m.add(Dense(728, activation='relu'))
    m.add(Dense(528, activation='relu'))
    m.add(Flatten())
    m.add(Dense(500, activation='relu', name="bottleneck"))
    m.add(Dense(528, activation='relu'))
    m.add(Dense(728, activation='relu'))
    m.add(Dense(943, activation='linear'))

    m.compile(loss='mean_squared_error', optimizer='SGD')
    m.summary()

现在我想把bottleneck层添加到下面的创建网络中：

    model = Sequential()
    model.add(Dense(930, activation='relu', input_shape=(943, 1)))
    model.add(Flatten())
    model.add(m.get_layer('bottleneck'))
    model.add(m.get_layer('bottleneck'))
    model.add(m.get_layer('bottleneck'))
    model.add(m.get_layer('bottleneck'))
    model.add(Flatten())
    model.add(Dense(100, activation='linear'))

但是在训练模型m之后，在启动错误时抛出错误：

ValueError: Input 0 is incompatible with layer bottleneck: expected axis -1 of input shape to have value 497904 but got shape (None, 876990)

Answer 1

错误消息试图告诉您，与第一个模型相比，第二个模型中“瓶颈”层的输入形状不同。

为了重用一个层，你需要匹配该层的输入数量。在您的示例中，第一个模型对该层有497904个输入，但您正尝试在下一个模型中使用它，该模型具有一个具有876990个输入的输入层。

我怀疑你想要更多这样的东西(注意，我在每种情况下都立即变平了，这样我们就可以更好地掌握每一层的输入数量)：

m = Sequential()
m.add(Flatten(input_shape=(943, 1)))
m.add(Dense(912, activation='relu'))
m.add(Dense(728, activation='relu'))
m.add(Dense(528, activation='relu'))
m.add(Dense(500, activation='relu', name="bottleneck"))
m.add(Dense(528, activation='relu'))
m.add(Dense(728, activation='relu'))
m.add(Dense(943, activation='linear'))

m.compile(loss='mean_squared_error', optimizer='SGD')
m.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten (Flatten)            (None, 943)               0         
_________________________________________________________________
dense (Dense)                (None, 912)               860928    
_________________________________________________________________
dense_1 (Dense)              (None, 728)               664664    
_________________________________________________________________
dense_2 (Dense)              (None, 528)               384912    
_________________________________________________________________
bottleneck (Dense)           (None, 500)               264500    
_________________________________________________________________
dense_3 (Dense)              (None, 528)               264528    
_________________________________________________________________
dense_4 (Dense)              (None, 728)               385112    
_________________________________________________________________
dense_5 (Dense)              (None, 943)               687447    
=================================================================
Total params: 3,512,091
Trainable params: 3,512,091
Non-trainable params: 0

请注意，我们的瓶颈层的输入具有形状(None,528)。现在，在第二个模型中，我们可以这样做：

model = Sequential()
model.add(Dense(930, activation='relu', input_shape=(943, 1)))
model.add(Flatten())
model.add(Dense(528, activation='relu'))
model.add(m.get_layer('bottleneck'))
model.add(Flatten())
model.add(Dense(100, activation='linear'))
model.summary()

Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_9 (Dense)              (None, 943, 930)          1860      
_________________________________________________________________
flatten_3 (Flatten)          (None, 876990)            0         
_________________________________________________________________
dense_10 (Dense)             (None, 528)               463051248 
_________________________________________________________________
bottleneck (Dense)           (None, 500)               264500    
_________________________________________________________________
flatten_4 (Flatten)          (None, 500)               0         
_________________________________________________________________
dense_11 (Dense)             (None, 100)               50100     
=================================================================
Total params: 463,367,708
Trainable params: 463,367,708
Non-trainable params: 0