如何在 TensorFlow 中编码标签？

Question

我需要将我的字符串标签转换为像 [0, 0, ... , 1, ... 0] 这样的向量。
据我所知，这被称为一个热向量。
我有 10 个类，所以有 10 个不同的字符串标签。

有人可以帮忙解决直接和逆变换吗？
我是 tensorflow 的新手，所以请善待。

Answer 1

前进方向很容易，因为有tf.one_hot op：

import tensorflow as tf

original_indices = tf.constant([1, 5, 3])
depth = tf.constant(10)
one_hot_encoded = tf.one_hot(indices=original_indices, depth=depth)

with tf.Session():
  print(one_hot_encoded.eval())

输出：

[[ 0.  1.  0.  0.  0.  0.  0.  0.  0.  0.]
 [ 0.  0.  0.  0.  0.  1.  0.  0.  0.  0.]
 [ 0.  0.  0.  1.  0.  0.  0.  0.  0.  0.]]

反之亦然，tf.where 可以找到非零索引：

def decode_one_hot(batch_of_vectors):
  """Computes indices for the non-zero entries in batched one-hot vectors.

  Args:
    batch_of_vectors: A Tensor with length-N vectors, having shape [..., N].
  Returns:
    An integer Tensor with shape [...] indicating the index of the non-zero
    value in each vector.
  """
  nonzero_indices = tf.where(tf.not_equal(
      batch_of_vectors, tf.zeros_like(batch_of_vectors)))
  reshaped_nonzero_indices = tf.reshape(
      nonzero_indices[:, -1], tf.shape(batch_of_vectors)[:-1])
  return reshaped_nonzero_indices

with tf.Session():
  print(decode_one_hot(one_hot_encoded).eval())

打印：

[1 5 3]