从 Keras model 中间移除一层

Question

I have tried to remove the layer number 10 from a pre trained VGG16 model so as to keep the pre-trained weights intact. However I am running into an unknown error. Here is my code:

from keras import Model
from keras.layers import Dropout
from keras.applications.vgg16 import VGG16

model = VGG16(include_top=False, input_shape=(H, W, 3), weights='imagenet')

# Disassemble layers
layers = [l for l in model.layers]

# Now stack everything back
# Note: If you are going to fine tune the model, do not forget to
#       mark other layers as un-trainable
    
x = layers[9].output
x = layers[11](x)
x = layers[12](x)
x = layers[13](x)
    
# Final touch
result_model = Model(input=layers[0].input, output=x)
result_model.summary()

I get the foollowing error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-60-2ec8c5fab52d> in <module>()
      8 # Important: the number of filters should be the same!
      9 # Note: the receiptive field of two 3x3 convolutions is 5x5.
---> 10 dropout1 = Dropout(0.5)(layers[-3].output)
     11 dropout2 = Dropout(0.5)(layers[-2].output)
     12 dropout3 = Dropout(0.5)(layers[-1].output)

/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py in output(self)
    843         if len(self._inbound_nodes) > 1:
    844             raise AttributeError('Layer ' + self.name +
--> 845                                  ' has multiple inbound nodes, '
    846                                  'hence the notion of "layer output" '
    847                                  'is ill-defined. '

AttributeError: Layer block4_conv1 has multiple inbound nodes, hence the notion of "layer output" is ill-defined. Use `get_output_at(node_index)` instead.

Answer 1

The problem is that when you call a layer with a new input node, the old input node is not deleted. You can see that when you call layers[11].inbound_nodes which has a length of 2 after running your code.

My solution is to manually remove the old inbound and outbound nodes. This is not so nice because these lists are protected variables.

layers[9]._outbound_nodes = []
x = layers[9].output

for layer in layers[11:]:
    layer._inbound_nodes = []
    x = layer(x)
    layer._outbound_nodes = []