Tensorflow tf.cond评估两个山脚

此代码段应计算整个培训阶段的用户输入平均值和测试阶段的输出平均值

这是输出结果：

import tensorflow as tf
import numpy as np

isTrain = tf.placeholder(tf.bool)
user_input = tf.placeholder(tf.float32)

# ema = tf.train.ExponentialMovingAverage(decay=.5)

with tf.device('/cpu:0'):
    beta = tf.Variable(tf.ones([1]))

    batch_mean = beta.assign(user_input)
    ema = tf.train.ExponentialMovingAverage(decay=0.5)
    ema_apply_op = ema.apply([batch_mean])
    ema_mean = ema.average(batch_mean)

    def mean_var_with_update():
        with tf.control_dependencies([ema_apply_op]):
            return tf.identity(batch_mean)

    mean = tf.cond(isTrain,
        mean_var_with_update,
        lambda: (ema_mean))

# ======= End Here ==========
saver = tf.train.Saver()
init = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init)

u_input = [[2], [3], [4] ]
for u in u_input:
    aa = sess.run([mean], feed_dict={user_input:u, isTrain: True })
    print("Train", aa)

for u in u_input:
    aa = sess.run([ema_mean], feed_dict={user_input:u, isTrain: False })
    print("Test correct", aa)

for u in u_input:
    aa = sess.run([mean], feed_dict={user_input:u, isTrain: False })
    print("Test", aa)

然而，

ema\u mean

在调用

sess.run（[mean]）

时总是得到评估，即使

isTrain=False

---

代码中有错误吗？tensorflow版本是0.7.1

我添加了一些日志语句，

ema\u mean

似乎只有在

isTrain

为false时才会得到评估

('Train', [array([ 2.], dtype=float32)])
('Train', [array([ 3.], dtype=float32)])
('Train', [array([ 4.], dtype=float32)])
('Test correct', [array([ 3.], dtype=float32)])
('Test correct', [array([ 3.], dtype=float32)])
('Test correct', [array([ 3.], dtype=float32)])
('Test', [array([ 2.5], dtype=float32)])
('Test', [array([ 2.75], dtype=float32)])
('Test', [array([ 3.375], dtype=float32)])

你看

tf.reset_default_graph()

isTrain = tf.placeholder(tf.bool)
user_input = tf.placeholder(tf.float32)

# ema = tf.train.ExponentialMovingAverage(decay=.5)

with tf.device('/cpu:0'):
    beta = tf.Variable(tf.ones([1]))

    batch_mean = beta.assign(user_input)
    ema = tf.train.ExponentialMovingAverage(decay=0.5)
    ema_apply_op = ema.apply([batch_mean])
    ema_mean = ema.average(batch_mean)

    def mean_var_with_update():
        with tf.control_dependencies([ema_apply_op]):
            return tf.Print(tf.identity(batch_mean), ["mean_var_with_update"])
            #return tf.identity(batch_mean)

    mean = tf.Print(tf.cond(isTrain,
        mean_var_with_update,
        lambda: (tf.Print(ema_mean, ["ema_mean"]))),
                    ["evaluating mean", isTrain])

# ======= End Here ==========
saver = tf.train.Saver()
init = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init)

u_input = [[2], [3], [4] ]
for u in u_input:
    aa = sess.run([mean], feed_dict={user_input:u, isTrain: True })
    print("Train", aa)

for u in u_input:
    aa = sess.run([ema_mean], feed_dict={user_input:u, isTrain: False })
    print("Test correct", aa)

for u in u_input:
    aa = sess.run([mean], feed_dict={user_input:u, isTrain: False })
    print("Test", aa)

---

请注意，

Print

语句是在对所有输入进行求值之后进行求值的，因此外部Print语句是最后一次打印的

我认为这与。条件中的

tf.control\u依赖项将把依赖项添加到
tf.cond
本身

因此，尝试在具有更新功能的
mean\u var\u函数中创建
ema\u apply\u op
。
如果
isTrain=False
则不应评估批处理平均值
如果
isTrain=True
则不应评估
ema\u平均值
，无论
isTrain
值如何，都会评估这两个值。这是正确的行为吗？@24hours:可能是。尝试在带有更新功能的
mean\u var\u内创建
ema\u apply\u op
。
[mean_var_with_update]
[evaluating mean][True]
[mean_var_with_update]
[evaluating mean][True]
[mean_var_with_update]
[evaluating mean][True]
[ema_mean]
[evaluating mean][False]
[ema_mean]
[evaluating mean][False]
[ema_mean]
[evaluating mean][False]