如何将一个模型的中间层的输出用作另一个模型的输入？

Question

我训练了一个模型 A 并尝试使用带有 name="layer_x" 的中间层的输出作为模型 B 的附加输入。

我尝试使用 Keras 文档中的中间层的输出 https://keras.io/getting-started/faq/#how-can-i-obtain-the-output-of-an-intermediate-layer.

模型 A：

inputs = Input(shape=(100,))
dnn = Dense(1024, activation='relu')(inputs)
dnn = Dense(128, activation='relu', name="layer_x")(dnn)
dnn = Dense(1024, activation='relu')(dnn)
output = Dense(10, activation='softmax')(dnn)

模型 B：

input_1 = Input(shape=(200,))
input_2 = Input(shape=(100,)) # input for model A

# loading model A
model_a = keras.models.load_model(path_to_saved_model_a)

intermediate_layer_model = Model(inputs=model_a.input, 
                                 outputs=model_a.get_layer("layer_x").output)

intermediate_output = intermediate_layer_model.predict(data)

merge_layer = concatenate([input_1, intermediate_output])
dnn_layer = Dense(512, activation="relu")(merge_layer)
output = Dense(5, activation="sigmoid")(dnn_layer)
model = keras.models.Model(inputs=[input_1, input_2], outputs=output)

当我调试时，我在这一行得到一个错误：

intermediate_layer_model = Model(inputs=model_a.input, 
                                 outputs=model_a.get_layer("layer_x").output)

File "..", line 89, in set_model
  outputs=self.neural_net_asc.model.get_layer("layer_x").output)
File "C:\WinPython\python-3.5.3.amd64\lib\site-packages\keras\legacy\interfaces.py", line 87, in wrapper
  return func(*args, **kwargs)
File "C:\WinPython\python-3.5.3.amd64\lib\site-packages\keras\engine\topology.py", line 1592, in __init__
  mask = node.output_masks[tensor_index]
AttributeError: 'Node' object has no attribute 'output_masks'

我可以使用get_layer("layer_x").output 访问张量，而output_mask 是None。我是否必须手动设置输出掩码？如果需要，如何设置此输出掩码？

Answer 1

您似乎做错了两件事：

intermediate_output = intermediate_layer_model.predict(data)

当您执行.predict() 时，您实际上是在通过图表传递数据并询问结果是什么。当你这样做时，intermediate_output 将是一个 numpy 数组，而不是你想要的层。

其次，您不需要重新创建新的中间模型。您可以直接使用model_a中您感兴趣的部分。

这是为我“编译”的代码：

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

inputs = Input(shape=(100,))
dnn = Dense(1024, activation='relu')(inputs)
dnn = Dense(128, activation='relu', name="layer_x")(dnn)
dnn = Dense(1024, activation='relu')(dnn)
output = Dense(10, activation='softmax')(dnn)

model_a = Model(inputs=inputs, outputs=output)

# You don't need to recreate an input for the model_a, 
# it already has one and you can reuse it
input_b = Input(shape=(200,))

# Here you get the layer that interests you from model_a, 
# it is still linked to its input layer, you just need to remember it for later
intermediate_from_a = model_a.get_layer("layer_x").output

# Since intermediate_from_a is a layer, you can concatenate it with the other input
merge_layer = concatenate([input_b, intermediate_from_a])
dnn_layer = Dense(512, activation="relu")(merge_layer)
output_b = Dense(5, activation="sigmoid")(dnn_layer)
# Here you remember that one input is input_b and the other one is from model_a
model_b = Model(inputs=[input_b, model_a.input], outputs=output_b)

我希望这是你想做的。

如果有什么不清楚的地方请告诉我:-)