从 Keras 层获取权重

Question

I am trying to obtain the weights from the following Dense layer:

x = Dense(1024)(Flatten()(previous_layer))

If I try to do the following:

x = Dense(1024)
weights = x.get_weights()

this works fine, but my understanding is that these weights will be useless, as we have not supplied any input to the layer.

However, if I try to do the following:

x = Dense(1024)(Flatten()(previous_layer))
weights = x.get_weights()

this doesn't work, as x is now a Tensor object and does not have the get_weights method:

'Tensor' object has no attribute 'get_weights'

What am I doing wrong?

Answer 1

There's a difference between the layer ( Dense(n) ) and the output tensor you get when applying this layer to some input tensor ( Dense(n)(input) ). You need to store the layer in a variable, not just the output tensor:

>>> import keras
>>> input_layer = keras.layers.Input((2,))
>>> layer = keras.layers.Dense(3) # create a layer
>>> print(layer)
<keras.layers.core.Dense object at 0x7f03ca9d4d68>
>>> print(layer.get_weights()) # the layer does not have weights yet
[]
>>> output_tensor = layer(input_layer) # apply the layer to the input tensor
>>> print(output_tensor)
Tensor("dense_1/BiasAdd:0", shape=(?, 3), dtype=float32)
>>> print(layer.get_weights()) # now get the weights
[array([[-0.84973848, -0.19682372, -0.14602524],
       [ 0.70318353, -0.1578933 , -0.94751853]], dtype=float32),
 array([ 0.,  0.,  0.], dtype=float32)]

Answer 2

lstm_output = keras.layers.LSTM(100,return_sequences = True)(x_5)

output_location_final = Dense(2,activation='tanh')

#lstm_output is a tensor

result = output_location_final(lstm_output)

#result is a layer

weight = output_location_final.get_weights()

print(weight)