tf.where

与​​np.where​​类似，有两种使用方式

  • ​​np.where(condition,x,y)​​condition中为True的为x，为False的为y
  • ​​np.where(condition)​​ 返回是满足条件的位置索引，如果是一个二维矩阵，返回两个一维array

​​np.where(condition,x,y)​​​x,y是可以取单独的数，但是​​tf.where(condition,x,y)​​x,y的维度要与condition一致

x = np.array([[0.3,0.4],[0.6,0.8]])
x = np.where(x[:,:]>0.5)
x
(array([1, 1], dtype=int64), array([0, 1], dtype=int64))

x = np.array([[0.3,0.4],[0.6,0.8]])
x = np.where(x[:,:]>0.5,1,0)
x
array([[0, 0],
       [1, 1]])
a = tf.constant([[1,3],[2,4]])
b = tf.where(a>2,1,0)

ValueError: Shapes must be equal rank, but are 0 and 2 for 'Select_4' (op: 'Select') with input shapes: [2,2], [], [].
a = tf.constant([[1,3],[2,4]])
a1 = tf.ones_like(a)
a0 = tf.zeros_like(a)
b = tf.where(a>2,a1,a0)
with tf.Session() as sess:
    print(b.eval（))
[[0 1]
 [0 1]]