在具有tensorflow后端的keras中重用内核权重

Question

I want to reuse kernel weight in keras with tensorflow backend.

tf.version = 1.11.0

My code is as following :

from tensorflow.python.keras.layers import  Input
from tensorflow.python.keras import regularizers
def infer3(data_input, Reuse):
    with tf.variable_scope('Network', reuse=Reuse):
        inputs = Input(tensor = data_input)
        network = tf.keras.layers.Conv2D(kernel_size=3, strides=2, filters=64, padding='same',kernel_regularizer=regularizers.l2(1),
                    activation='linear', kernel_initializer="glorot_normal", name='conv1', bias_initializer='zeros')(inputs)

    return network

tf.reset_default_graph() 
input_tensor = tf.placeholder(tf.float32,shape=[BATCH_SIZE,img_H,img_W,1])
output_tensor = tf.placeholder(tf.float32,shape=[BATCH_SIZE,img_H,img_W,1]) 
in_training_mode = tf.placeholder(tf.bool)     


network = infer3(input_tensor,False)
network_test = infer3(input_tensor,True)

When I type " tf.trainable_variables() "

It show message:

[<tf.Variable 'Network/conv1/kernel:0' shape=(3, 3, 1, 64) dtype=float32>,
 <tf.Variable 'Network/conv1/bias:0' shape=(64,) dtype=float32>,
 <tf.Variable 'Network_1/conv1/kernel:0' shape=(3, 3, 1, 64) dtype=float32>,
 <tf.Variable 'Network_1/conv1/bias:0' shape=(64,) dtype=float32>]

How can I reuse kernel weight?

Answer 1

As stated here , you can use get_weights() method on an instance of layer to get its weights. Then you can assign those weights to a variable and initialize another layer of the same type and shape with those weights.