ValueError: Variable ExponentialMovingAverage/ already exists 解決

ExponentialMovingAverage使用方法

import tensorflow as tf
variable_averages = tf.train.ExponentialMovingAverage(learning_rate, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())

發生原因

tf.trainable_variables()會把所有graph中的可訓練的變數輸入ExponentialMovingAverage做出新的變數，例如用變數
“layer1-conv1/weight"做出"layer1-conv1/weight/ExponentialMovingAverage:0”，
假如"layer1-conv1/weight"當時宣告的方式是用tf.get_variable()，則ExponentialMovingAverage會繼承他的不可重複性，因此會出現ExponentialMovingAverage already exitsts。

解決方法

import tensorflow as tf

with tf.variable_scope(name_or_scope=tf.get_variable_scope(),reuse=tf.AUTO_REUSE):
	variable_averages = tf.train.ExponentialMovingAverage(learning_rate, global_step)
	variables_averages_op = variable_averages.apply(tf.trainable_variables())

注意:不要用reuse=True代替reuse=tf.AUTO_REUSE，因為會出現各種神奇的錯誤。

用name_or_scope=tf.get_variable_scope()會使"layer1-conv1/weight"變成
“layer1-conv1/weight/ExponentialMovingAverage:0”，用name_or_scope='test’會使"layer1-conv1/weight"變成"test/layer1-conv1/weight/ExponentialMovingAverage:0"