通过额外的输入和层扩展已经构建的张量流模型

Question

我尝试用另一层扩展我已经构建的模型之一，以对模型先前输出的减法和附加输入求和。我所有的尝试都可以使用 Lambda 层进行总结：

    new_input = Input(shape=(20, 3))
    sub_layer = Lambda(lambda x: Reshape((20,1))(backend.sum(x[0] - x[1], axis=-1)))((model.output,new_input))
    model = Model(inputs=[model.input, new_input], outputs=sub_layer.output)

但无论我做什么，我总是收到此错误：

Could not build a TypeSpec for KerasTensor(type_spec=TensorSpec(shape=(None, 20, 1), dtype=tf.float32, name=None), name='tf.reshape/Reshape:0', description="created by layer 'tf.reshape'") of unsupported type <class 'keras.engine.keras_tensor.KerasTensor'>.

Answer 1

不确定你的预训练模型是什么样的，但你可以尝试这样的事情：

import tensorflow as tf

inputs = tf.keras.layers.Input((10,))
outputs = tf.keras.layers.Dense(20, activation='relu')(inputs)
outputs = tf.keras.layers.RepeatVector(3)(outputs)
outputs = tf.keras.layers.Reshape((20,3))(outputs)
model1 = tf.keras.Model(inputs, outputs)

new_input = tf.keras.layers.Input(shape=(20, 3))
sub_layer = tf.keras.layers.Lambda(lambda x: tf.keras.layers.Reshape((20,1))(tf.keras.backend.sum(x[0] - x[1], axis=-1)))((model1.output,new_input))
model2 = tf.keras.Model(inputs=[model1.input, new_input], outputs=sub_layer)
print(model2.summary())

Model: "model_4"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 input_7 (InputLayer)           [(None, 10)]         0           []                               
                                                                                                  
 dense_3 (Dense)                (None, 20)           220         ['input_7[0][0]']                
                                                                                                  
 repeat_vector_1 (RepeatVector)  (None, 3, 20)       0           ['dense_3[0][0]']                
                                                                                                  
 reshape (Reshape)              (None, 20, 3)        0           ['repeat_vector_1[0][0]']        
                                                                                                  
 input_8 (InputLayer)           [(None, 20, 3)]      0           []                               
                                                                                                  
 lambda_3 (Lambda)              (None, 20, 1)        0           ['reshape[0][0]',                
                                                                  'input_8[0][0]']                
                                                                                                  
==================================================================================================
Total params: 220
Trainable params: 220
Non-trainable params: 0
__________________________________________________________________________________________________
None

通过额外的输入和层扩展已经构建的张量流模型

Extend already build tensorflow model by additional input and layer

python

tensorflow

keras

tensorflow2.0

tf.keras