tensorflow 的 control_dependencies函数
tensorflow.control_dependencies( tasklist ）可以把tasklist里的操作作为预定操作。一般是这样使用：
with tensorflow.control_dependencies( tasklist ） ：
trainOp = tensorflow.no_op()
然后执行这个 trainOp （即Tensorflow.Session().run( trainOp)）时,就会把 tasklist所包含的动作先执行掉.
但是在with里面不能直接赋值别的操作,那样的话,就不再执行tasklist里的动作了.比如有个动作 task1 ,它不在tasklist里面,也与tasklist不相关.但是在with代码段里面将task1直接赋值给trainOP,那么不会执行tasklist包含的动作:
with tensorflow.control_dependencies( tasklist ） ：
trainOp = task1
可以这样修改,将新加入的动作和no_动作一起放在list里,达到执行tasklist里包含的动作:
with tensorflow.control_dependencies( tasklist ） ：
trainOp = tensorflow.no_op()
trainOp = [trainOp , task1]
下面是一段验证代码:
import tensorflow as tf
a = tf.Variable( 2 )
selfAdd = tf.Variable( 0 )
selfAddition = tf.assign_add( selfAdd , 3 )
selfSub = tf.Variable( 0 )
selfSubtraction = tf.assign_sub( selfSub , 2 )
b = tf.multiply( a , selfAdd )
with tf.control_dependencies( [ selfAddition ] ) :
train_op = tf.no_op() #待解释语句1
print( train_op )
train_op =[ train_op , selfSubtraction] #待解释语句2
print( train_op )
with tf.Session() as sess :
init = tf.global_variables_initializer()
sess.run( init )
for i in range( 20 ) :
sess.run( train_op ) #待解释语句3
print( "selfAdd:" , sess.run( selfAdd ) )
ra = sess.run(selfAdd )
rb = sess.run(b )
print( '@end selfAdd:' ,ra )
rs = sess.run(selfSub)
print('@end selfSub:' , rs )
print( 'b:' , rb )
sefAddition 是对变量selfAdd进行自加的操作,selfSubtraction是对变量selfSub进行自减的操作. "待解释语句1"建立了一个trainOp,它本身不做什么,但是它会定义预处理动作, 就是control_dependencies调用时实参所指定的 selfAddition. 若想除定义预处理动作之外,还想再加其他动作,可按照"待解释语句2"的方法,在列表里增加动作.这样在"待解释语句3"里每次动执行 selfAddition和selfSubtraction.
必须在no_op之后再加动作,似乎和一般理解不一样,不知道是不是哪里有不对的地方,下面是我的测试代码和输出:
import tensorflow as tf
a = tf.Variable( 2 )
selfAdd = tf.Variable( 0 )
selfAddition = tf.assign_add( selfAdd , 3 )
selfSub = tf.Variable( 0 )
selfSubtraction = tf.assign_sub( selfSub , 2 )
b = tf.multiply( a , selfAdd )
with tf.control_dependencies( [ selfAddition ] ) :
train_op = selfSubtraction
print( train_op )
with tf.Session() as sess :
init = tf.global_variables_initializer()
sess.run( init )
for i in range( 20 ) :
sess.run( train_op )
print( "selfAdd:" , sess.run( selfAdd ) )
ra = sess.run(selfAdd )
rb = sess.run(b )
print( '@end selfAdd:' ,ra )
rs = sess.run(selfSub)
print('@end selfSub:' , rs )
print( 'b:' , rb )
Tensor("AssignSub:0", shape=(), dtype=int32_ref)
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
selfAdd: 0
@end selfAdd: 0
@end selfSub: -40
b: 0经过分析，发现是要在 with tf.control_dependencies 里的代码块里直接定义 operation ，然后执行代码块里的operation就会执行dependeccies里的operation列表。 在外面定义，然后在with 里的代码块中转等于就不会起作用。下面代码1可以起作用，代码2不起作用。
代码1：
import tensorflow as tf
a = tf.Variable( 2 )
selfAdd = tf.Variable( 0 )
selfAddition =tf.assign_add( selfAdd , 3 )
selfSub = tf.Variable( 0 )
#selfSubtraction1 = tf.assign_sub( selfSub , 2 )
#print('op 1：', selfSubtraction1 )
with tf.control_dependencies( [ selfAddition ] ) :
selfSubtraction = tf.assign_sub( selfSub , 2 )
print('op：', selfSubtraction )
with tf.Session() as sess :
init = tf.global_variables_initializer()
sess.run( init )
for i in range( 20 ) :
sess.run( selfSubtraction )
print( "selfAdd:" , sess.run( selfAdd ) )
ra = sess.run(selfAdd )
print( '@end selfAdd:' ,ra )
rs = sess.run(selfSub)
print('@end selfSub:' , rs )
代码2
import tensorflow as tf
a = tf.Variable( 2 )
selfAdd = tf.Variable( 0 )
selfAddition =tf.assign_add( selfAdd , 3 )
selfSub = tf.Variable( 0 )
selfSubtraction1 = tf.assign_sub( selfSub , 2 )
print('op 1：', selfSubtraction1 )
with tf.control_dependencies( [ selfAddition ] ) :
selfSubtraction = selfSubtraction1 # tf.assign_sub( selfSub , 2 )
print('op：', selfSubtraction )
with tf.Session() as sess :
init = tf.global_variables_initializer()
sess.run( init )
for i in range( 20 ) :
sess.run( selfSubtraction )
print( "selfAdd:" , sess.run( selfAdd ) )
ra = sess.run(selfAdd )
print( '@end selfAdd:' ,ra )
rs = sess.run(selfSub)
print('@end selfSub:' , rs )
参考:
https://www.cnblogs.com/qjoanven/p/7736025.html