如何在字符级别对单个热门编码？

Question

我想将一个句子转换为一个单热矢量数组.这些向量将是字母表的一个热门表示.它看起来如下:

"hello" # h=7, e=4 l=11 o=14

会成为

[[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
 [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]

不幸的是,来自sklearn的OneHotEncoder不作为输入字符串.

Answer 1

这是递归神经网络中的常见任务tensorflow,如果您想使用它,那么就有一个特定的功能用于此目的.

In [24]: alphabets = {'a' : 0, 'b': 1, 'c':2, 'd':3, 'e':4, 'f':5, 'g':6, 'h':7, 'i':8, 'j':9, 'k':10, 'l':11, 'm':12, 'n':13, 'o':14}

In [25]: idxs = [alphabets[ch] for ch in 'hello']

In [26]: idxs
Out[26]: [7, 4, 11, 11, 14]

# @divakar's approach
In [153]: idxs = np.fromstring("hello",dtype=np.uint8)-97

In [13]: one_hot = tf.one_hot(idxs, 26, dtype=tf.uint8)

In [14]: sess = tf.InteractiveSession()

In [15]: one_hot.eval()
Out[15]: 
array([[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=uint8)

Answer 2

只需将传递的字符串中的字母与给定的字母表进行比较:

def string_vectorizer(strng, alphabet=string.ascii_lowercase):
    vector = [[0 if char != letter else 1 for char in alphabet] 
                  for letter in strng]
    return vector

请注意,使用自定义字母表(例如"defbcazk"时,将按照每个元素出现在原始列表中的顺序排列列).

输出string_vectorizer('hello'):

[[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
 [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]