函数调用后，张量流变量的名称发生了变化

Question

我正在尝试编写自己的批处理规范化代码。因此，我测试下面的代码。为了跟踪在线平均均值和方差，我将它们作为参数传递到 getsta() 函数中。但是，我发现"avg_mean"和"avg_variance"的名称发生了变化。虽然我可以稍后手动强制更改它们的名称，但似乎 Tensorflow 将每个名称视为一个单独的变量。

def getsta(x,avg_mean,avg_variance):
print('getsta start...')
decay=0.9
mean = tf.get_variable(
'mean', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
variance = tf.get_variable(
'howvariance', [1], tf.float32,
initializer=tf.constant_initializer(1.0, tf.float32))
'''
if (avg_mean == 0.0): 
avg_mean = tf.get_variable(
'avg_mean', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
if (avg_variance == 0.0):
avg_variance = tf.get_variable(
'avg_variance', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
'''
mean, variance = tf.nn.moments(x, [0], name='moments')
avg_mean -= (1.0 - decay) * (avg_mean - mean)
avg_variance -= (1.0 - decay) * (avg_variance - variance)
return x, mean, variance, avg_mean, avg_variance
def train():
x1 = tf.constant([1,2,3,4], tf.float32)
x2 = tf.constant([5,6,7,8], tf.float32)
x3 = tf.constant([1,3,5,7], tf.float32)
x4 = tf.constant([4,8,12,16], tf.float32)
with tf.variable_scope("AAA") as scopeA:
avg_mean1 = tf.get_variable(
'test1', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
avg_variance1 = tf.get_variable(
'test2', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
print([avg_mean1.name, avg_variance1.name])
y1, mean1, variance1, avg_mean1, avg_variance1 = getsta(x1,avg_mean1,avg_variance1)
print([avg_mean1.name, avg_variance1.name])
scopeA.reuse_variables()
print([avg_mean1.name, avg_variance1.name])
y1, mean1, variance1, avg_mean1, avg_variance1 = getsta(x2,avg_mean1,avg_variance1)
print([avg_mean1.name, avg_variance1.name])
avg_mean1 = tf.identity(avg_mean1, name='avg_mean')
avg_variance1 = tf.identity(avg_variance1, name='avg_variance')
print([avg_mean1.name, avg_variance1.name])
with tf.variable_scope("BBB") as scopeB:
avg_mean2 = tf.get_variable(
'test1', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
avg_variance2 = tf.get_variable(
'test2', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
print([avg_mean2.name, avg_variance2.name])
y2, mean2, variance2, avg_mean2, avg_variance2 = getsta(x3, avg_mean2, avg_variance2)
print([avg_mean2.name, avg_variance2.name])
scopeB.reuse_variables()
print([avg_mean2.name, avg_variance2.name])
y2, mean2, variance2, avg_mean2, avg_variance2 = getsta(x4, avg_mean2, avg_variance2)
print([avg_mean2.name, avg_variance2.name])
avg_mean2 = tf.identity(avg_mean2, name='avg_mean')
avg_variance2 = tf.identity(avg_variance2, name='avg_variance')
print([avg_mean2.name, avg_variance2.name])
saver = tf.train.Saver()
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
print(sess.run([y1, mean1, variance1, avg_mean1, avg_variance1]))
print(sess.run([y2, mean2, variance2, avg_mean2, avg_variance2]))
allVars = tf.global_variables()
values = sess.run(allVars)
for var, val in zip(allVars, values):
print(var.name, val)
saver.save(sess, "/tmp/test_EMA/var_save.ckpt")

运行上面的 train() 函数后，我得到的输出为

['AAA/test1:0', 'AAA/test2:0']
getsta start...
['AAA/sub_1:0', 'AAA/sub_3:0']
['AAA/sub_1:0', 'AAA/sub_3:0']
getsta start...
['AAA/sub_5:0', 'AAA/sub_7:0']
['AAA/avg_mean:0', 'AAA/avg_variance:0']
['BBB/test1:0', 'BBB/test2:0']
getsta start...
['BBB/sub_1:0', 'BBB/sub_3:0']
['BBB/sub_1:0', 'BBB/sub_3:0']
getsta start...
['BBB/sub_5:0', 'BBB/sub_7:0']
['BBB/avg_mean:0', 'BBB/avg_variance:0']
[array([ 5.,  6.,  7.,  8.], dtype=float32), 6.5, 1.25, array([ 0.875], dtype=float32), array([ 0.23750001], dtype=float32)]
[array([  4.,   8.,  12.,  16.], dtype=float32), 10.0, 20.0, array([ 1.36000001], dtype=float32), array([ 2.45000005], dtype=float32)]
AAA/test1:0 [ 0.]
AAA/test2:0 [ 0.]
AAA/mean:0 [ 0.]
AAA/howvariance:0 [ 1.]
BBB/test1:0 [ 0.]
BBB/test2:0 [ 0.]
BBB/mean:0 [ 0.]
BBB/howvariance:0 [ 1.]

我有几个问题：

1. 有没有办法在不更改名称的情况下记录"avg_mean"？我想在加载模型时加载最终更新的值。
2. 是否可以在 getsta() 函数中定义"avg_mean"，同时仍然将其保留为静态变量，即它的初始值来自上次的调用，但并不总是从 0 开始。

3. 当我打印出所有变量时，为什么名称和值似乎都不像输出中的最后八行那样正确？我希望输出是

   AAA/avg_mean:0 [0.875]
   AAA/avg_variance:0 [0.23750001]
   BBB/avg_mean:0 [1.36000001]
   BBB/avg_variance:0 [2.45000005]
   

提前感谢您的帮助！

--------------------更新--------------------

我修改了我的代码如下

def train():
x1 = tf.constant([1,2,3,4], tf.float32)
x2 = tf.constant([5,6,7,8], tf.float32)
x3 = tf.constant([1,3,5,7], tf.float32)
x4 = tf.constant([4,8,12,16], tf.float32)
with tf.variable_scope("AAA") as scopeA:
avg_mean1 = tf.get_variable(
'avg_mean', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
avg_variance1 = tf.get_variable(
'avg_variance', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
y1, mean1, variance1 = getsta(x1,avg_mean1,avg_variance1)
scopeA.reuse_variables()
y1, mean1, variance1 = getsta(x2,avg_mean1,avg_variance1)
print([avg_mean1.name, avg_variance1.name])
with tf.variable_scope("BBB") as scopeB:
avg_mean2 = tf.get_variable(
'avg_mean', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
avg_variance2 = tf.get_variable(
'avg_variance', [1], tf.float32,
initializer=tf.constant_initializer(0.0, tf.float32))
y2, mean2, variance2 = getsta(x3, avg_mean2, avg_variance2)
scopeB.reuse_variables()
y2, mean2, variance2 = getsta(x4, avg_mean2, avg_variance2)
print([avg_mean2.name, avg_variance2.name])
saver = tf.train.Saver()
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
print(sess.run([y1, mean1, variance1, avg_mean1, avg_variance1]))
print(sess.run([y2, mean2, variance2, avg_mean2, avg_variance2]))
allVars = tf.global_variables()
values = sess.run(allVars)
for var, val in zip(allVars, values):
print(var.name, val)
saver.save(sess, "/tmp/test_EMA/var_save.ckpt")

现在变量的名称是固定的。但是，输出(变量的值)似乎不正确。输出是

[array([ 5.,  6.,  7.,  8.], dtype=float32), 6.5, 1.25, array([ 0.], dtype=float32), array([ 0.], dtype=float32)]
[array([  4.,   8.,  12.,  16.], dtype=float32), 10.0, 20.0, array([ 0.], dtype=float32), array([ 0.], dtype=float32)]
AAA/avg_mean:0 [ 0.]
AAA/avg_variance:0 [ 0.]
AAA/mean:0 [ 0.]
AAA/variance:0 [ 1.]
BBB/avg_mean:0 [ 0.]
BBB/avg_variance:0 [ 0.]
BBB/mean:0 [ 0.]

我应该如何修改我的代码以获得正确的结果？谢谢。

Answer 1

以下是代码中需要修复的问题：

1-在下一行中，您将avg_variance1替换为getsta返回的operation。你绝不能那样做。当您创建变量avg_variance1并将其传递给getsta变量将在计算图中的所有位置更新时，TensorFlow 将处理这个问题，您无需显式执行此操作。只要你指的是variable的name你就没问题。TensorFlow变量不像常规的Python变量。

y1, mean1, variance1, avg_mean1, avg_variance1 = getsta(x1,avg_mean1,avg_variance1)

(您需要为avg_mean1、avg_mean2等修复此问题)

2-avg_mean2、avg_variance2、avg_mean2和avg_variance2是Tensor而不是variable。因此，它们不在您要打印的variables列表中。