Because I need to write some TF Code for , Forget about sharing mechanism implementation , Reviewed and wrote .

1. Sharing mechanism

TF Mainly provide Variable Scope This unique mechanism to share variables

> # Create or return a variable with a given name  
> tf.get_variable(<name>, <shape>, <initializer>) 
> 
> # Transfer management to get_variable() The scope of the variable name  
> tf.variable_scope(<scope_name>)

2. example

Step 1： Create variables

def conv_relu(input, kernel_shape, bias_shape):
    # Create variable named "weights".
    weights = tf.get_variable("weights", kernel_shape,
        initializer=tf.random_normal_initializer())
    # Create variable named "biases".
    biases = tf.get_variable("biases", bias_shape,
        initializer=tf.constant_initializer(0.0))
    conv = tf.nn.conv2d(input, weights,
        strides=[1, 1, 1, 1], padding='SAME')
    return tf.nn.relu(conv + biases)

Step 2： Define scope ： This part is a reusable module , It can be convolution and other operations .

def my_image_filter(input_images):
    with tf.variable_scope("conv1"):
        # Variables created here will be named "conv1/weights", "conv1/biases".
        relu1 = conv_relu(input_images, [5, 5, 32, 32], [32])
    with tf.variable_scope("conv2"):
        # Variables created here will be named "conv2/weights", "conv2/biases".
        return conv_relu(relu1, [5, 5, 32, 32], [32])

Step3： Reuse
Pay attention here , When reusing a convolution block, be sure to use ：scope.reuse_variables()

with tf.variable_scope("image_filters") as scope:
    result1 = my_image_filter(image1)
    scope.reuse_variables() #  The most important sentence 
    result2 = my_image_filter(image2) #  In this case image2 And image1 You can use the same module 

3. Reference resources ：

1. Official website guidelines ：variable_scope
2. https://www.jianshu.com/p/ab0d38725f88