如何将顺序值放入 sequence_mask？

Question

我有一个顺序值...

v =  tf.constant([0,1,2,3,4,5,6,7,8,9,10,11,12,13,14], dtype=tf.int32).

和序列掩码..

[
[True, True, False, False, False],
[True, True, True, True, False],
[True, True, True, False, False],
[True, True, True, True, True],
[True, False, False, False, False],
]

如果我想用张量'v'的元素填充sequence_mark以获得这样的结果..

[
[0, 1, 0, 0, 0],
[2, 3, 4, 5, 0],
[6, 7, 8, 0, 0],
[9, 10, 11, 12, 13],
[14, 0, 0, 0, 0],
]

我能怎么做。

Answer 1

你可以这样做：

import tensorflow as tf

# Sequence lengths
s = tf.constant([2, 4, 3, 5, 1])
# Make mask
m = tf.sequence_mask(s, 5)
# Mask as integers
m_int = tf.dtypes.cast(m, tf.int32)
# Cumsum over mask, starting from 0
c = tf.cumsum(tf.reshape(m_int, [-1]), exclusive=True)
# Reshape cumsum to original shape and apply mask
result = tf.reshape(c, tf.shape(m)) * m_int
with tf.Session() as sess:
    print(sess.run(result))
    # [[ 0  1  0  0  0]
    #  [ 2  3  4  5  0]
    #  [ 6  7  8  0  0]
    #  [ 9 10 11 12 13]
    #  [14  0  0  0  0]]

如果v是一个不只是简单序列的值数组，您可以执行以下操作：

# v is a vector of values
result = tf.reshape(tf.gather(v, c), tf.shape(m)) * m_int