TensorFlow, how to reuse variable region name

Question

I defined a class here

class BasicNetwork(object): def __init__(self, scope, task_name, is_train=False, img_shape=(80, 80)): self.scope = scope self.is_train = is_train self.task_name = task_name self.__create_network(scope, img_shape=img_shape) def __create_network(self, scope, img_shape=(80, 80)): with tf.variable_scope(scope): with tf.variable_scope(self.task_name): with tf.variable_scope('input_data'): self.inputs = tf.placeholder(shape=[None, *img_shape, cfg.HIST_LEN], dtype=tf.float32) with tf.variable_scope('networks'): with tf.variable_scope('conv_1'): self.conv_1 = slim.conv2d(activation_fn=tf.nn.relu, inputs=self.inputs, num_outputs=32, kernel_size=[8, 8], stride=4, padding='SAME', trainable=self.is_train) with tf.variable_scope('conv_2'): self.conv_2 = slim.conv2d(activation_fn=tf.nn.relu, inputs=self.conv_1, num_outputs=64, kernel_size=[4, 4], stride=2, padding='SAME', trainable=self.is_train) with tf.variable_scope('conv_3'): self.conv_3 = slim.conv2d(activation_fn=tf.nn.relu, inputs=self.conv_2, num_outputs=64, kernel_size=[3, 3], stride=1, padding='SAME', trainable=self.is_train) with tf.variable_scope('f_c'): self.fc = slim.fully_connected(slim.flatten(self.conv_3), 512, activation_fn=tf.nn.elu, trainable=self.is_train) 

And I want to define two instances of BasicNetwork with different task names. The scope is global. But when I check the output, there is

 ipdb> for i in net_1.layres: print(i) Tensor("global/simple/networks/conv_1/Conv/Relu:0", shape=(?, 20, 20, 32), dtype=float32, device=/device:GPU:2) Tensor("global/simple/networks/conv_2/Conv/Relu:0", shape=(?, 10, 10, 64), dtype=float32, device=/device:GPU:2) Tensor("global/simple/networks/conv_3/Conv/Relu:0", shape=(?, 10, 10, 64), dtype=float32, device=/device:GPU:2) Tensor("global/simple/networks/f_c/fully_connected/Elu:0", shape=(?, 512), dtype=float32, device=/device:GPU:2) ipdb> for i in net_2.layres: print(i) Tensor("global_1/supreme/networks/conv_1/Conv/Relu:0", shape=(?, 20, 20, 32), dtype=float32, device=/device:GPU:2) Tensor("global_1/supreme/networks/conv_2/Conv/Relu:0", shape=(?, 10, 10, 64), dtype=float32, device=/device:GPU:2) Tensor("global_1/supreme/networks/conv_3/Conv/Relu:0", shape=(?, 10, 10, 64), dtype=float32, device=/device:GPU:2) Tensor("global_1/supreme/networks/f_c/fully_connected/Elu:0", shape=(?, 512), dtype=float32, device=/device:GPU:2) 

As you can see in the output, a new global_1 scope has been global_1 , but I want to make it global . I set reuse=True , but later discovered that when there is no realm named global , reuse=True cannot be used. What should I do?

Answer 1

using reuse True, you can get existing variables. Now for reuse of variables hoses should exist on the schedule. If variables with the same name exist, you can reuse them for another operation.

 class BasicNetwork(object): def __init__(self, scope, task_name, reuse, is_train=False, img_shape=(80, 80)): self.scope = scope self.is_train = is_train self.reuse = reuse self.task_name = task_name self.__create_network(scope, reuse=self.reuse, img_shape=img_shape) def __create_network(self, scope, reuse=None, img_shape=(80, 80)): with tf.variable_scope(scope, reuse=reuse): ... # delete this line with tf.variable_scope(self.task_name): # or replace with; with tf.name_scope(self.task_name): trainnet = BasicNetwork('global', taskname, None) # resue the created variables valnet = BasicNetwork('global', taskname, True)