使用 Keras Sequential 模型实现快捷方式

Question

I have implemented shortcut with the Keras functional model this way:

inputs = ...

# shortcut path
shortcut = ShortcutLayer()(inputs)

# main path
outputs = MainLayer()(inputs)

# add main and shortcut together
outputs = Add()([outputs, shortcut])

Is it possible to convert this to a Sequential model, so that I don't need to know inputs in advance?

Basically, what I want to achieve looks like:

def my_model_with_shortcut():
    # returns a Sequential model equivalent to the functional one above

model = my_model_with_shortcut()

inputs = ...
outputs = model(inputs)

Answer 1

I would try the following;

def my_model_with_shortcut():
    def _create_shortcut(inputs):
         # here create model as in case you know inputs, e.g.:
         aux = Dense(10, activation='relu')(inputs)
         output = Dense(10, activation='relu')(aux)
         return output  
    return _create_shortcut

Now you should have your scenario possible.

Answer 2

Usually I create a callable object that maintains the model. I do this when I want this segment to be reused many times with the same weights .

You can try Marcin's approach, but you will create a new model every time you call the function. It may be enough for you, depending on your purposes. My answer has these differences:

• Advantage: you reuse the same model with the same weights as many times as you want
• Disadvantage: after the first input, all new inputs cannot have different shapes (a consequence of sharing the weights)

Code:

from keras.models import Model
from keras.layers import *
import numpy as np

class ModelSegment():
    def __init__(self):

        self.shortcutLayer = Dense(10,activation='sigmoid')
        self.mainLayer = Dense(10,activation='sigmoid')
        self.addLayer = Add()

        self.model = None

    def __call__(self,inputs):

        if self.model is None:
            shortcut = self.shortcutLayer(inputs)
            outputs = self.mainLayer(inputs)
            outputs = self.addLayer([outputs,shortcut])

            #creating this model is not really necessary
            #you can skip the following code line and discard the if clause
            #but since you stated you wanted a "model" I added it here
            self.model = Model(inputs,outputs)

            return outputs
        else:
            return self.model(inputs)



segment = ModelSegment()

inputTensor = Input((13,))
inputTensor2 = Input((13,))
out = segment(inputTensor)
out2 = segment(inputTensor2)

m1 = Model(inputTensor,out)
m2 = Model(inputTensor2,out2)

m1.summary()
m2.summary()

w1 = m1.get_weights()
w2 = m2.get_weights()

for arr1,arr2 in zip(w1,w2):
    print((arr1==arr2).all())