圖形依賴關係如何與`tf.cond`配合使用？

Question

假設True和False情況下存在單獨的相關性：

tensorflow是否假定兩個依賴項都是必需的，因此會處理True和False情況下的完整子圖？或者它：

1. 工藝布爾表達式的依賴關係，然後
2. 過程的依賴關係無論是真正的異或tf.cond(...)的虛假的一面呢？

Answer 1

的tf.cond()功能被設計評估布爾表達式之後執行只有一個真或假分支，。例如，如果你寫的東西像下面這樣：

v = tf.Variable(0) 
condition = tf.placeholder(tf.bool, shape=[]) 

op_to_run = tf.cond(condition, lambda: v.assign_add(1), lambda: v.assign_sub(1)) 

sess = tf.Session() 
sess.run(tf.global_variables_initializer()) 

print(sess.run(v)) # ==> "0" 

for _ in range(3): 
    sess.run(op_to_run, feed_dict={condition: True}) 

print(sess.run(v)) # ==> "3" 

for _ in range(5): 
    sess.run(op_to_run, feed_dict={condition: False}) 

print(sess.run(v)) # ==> "-2" 

不過請注意，您必須確保任何副作用的運算定義內的lambda（或等價的功能），你傳遞給tf.cond() 。如果副作用的OPS被外界的tf.cond()定義他們會無條件地執行：

v = tf.Variable(0) 
condition = tf.placeholder(tf.bool, shape=[]) 

# N.B. DO NOT DO THIS! Both side-effecting ops are defined outside the `tf.cond()` 
# so they will both execute, regardless of the condition. 
inc_op = v.assign_add(1) 
dec_op = v.assign_sub(1) 

op_to_run = tf.cond(condition, lambda: inc_op, lambda: dec_op) 

sess = tf.Session() 
sess.run(tf.global_variables_initializer()) 

print(sess.run(v)) # ==> "0" 

# Both the `assign_add()` and `assign_sub()` will run, cancelling each other out. 
for _ in range(3): 
    sess.run(op_to_run, feed_dict={condition: True}) 

print(sess.run(v)) # ==> "0" 

# Both the `assign_add()` and `assign_sub()` will run, cancelling each other out. 
for _ in range(5): 
    sess.run(op_to_run, feed_dict={condition: False}) 

print(sess.run(v)) # ==> "0"