这是一种可能的解决方案:
import tensorflow as tf
# Input data
nums = tf.placeholder(tf.int32, [None, None])
rows = tf.shape(nums)[0]
# Number of zeros on each row
zero_mask = tf.cast(tf.equal(nums, 0), tf.int32)
num_zeros = tf.reduce_sum(zero_mask, axis=1)
# Random values
r = tf.random_uniform([rows], 0, 1, dtype=tf.float32)
# Multiply by the number of zeros to decide which of the zeros you pick
zero_idx = tf.cast(tf.floor(r * tf.cast(num_zeros, r.dtype)), tf.int32)
# Find the indices of the smallest values, which should be the zeros
_, zero_pos = tf.nn.top_k(-nums, k=tf.maximum(tf.reduce_max(num_zeros), 1))
# Select the corresponding position of each row
result = tf.gather_nd(zero_pos, tf.stack([tf.range(rows), zero_idx], axis=1))
# Test
with tf.Session() as sess:
x = [[0,0,0,1,1,1],
[1,2,1,0,1,0]]
print(sess.run(result, feed_dict={nums: x}))
print(sess.run(result, feed_dict={nums: x}))
print(sess.run(result, feed_dict={nums: x}))
示例输出:
[1 3]
[2 5]
[0 3]
如果某行没有任何零,那么它将选择索引 0,尽管您可以制作一个掩码来过滤具有以下内容的那些:
has_zeros = tf.reduce_any(tf.equal(nums, 0), axis=1)