您可以使用字典映射和F.one_hot 的组合来实现此目的,F.one_hot 是执行密集到one-hot 转换的内置函数。
-
构造字典映射str到索引。我们将插入一个额外的字符"0",以便为短于最大长度的单词添加一些填充:
>>> alphabet = dict(zip('0abcdefghijklmnopqrstuvwxyz', range(27)))
-
找到最大长度并填充较短的单词:
>>> max_length = max(len(w) for w in words)
-
用适当数量的0 字符填充所有单词:
>>> padded = [w + '0'*(max_length - len(w)) for w in words]
['cd0', 'abc']
-
现在通过遍历所有填充的单词和字符来构造张量:
>>> dense = torch.tensor([[alphabet[c] for c in w] for w in padded])
tensor([[3, 4, 0],
[1, 2, 3]])
最后申请torch.functional.one_hot:
>>> x = F.one_hot(dense, num_classes=27)
tensor([[[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, 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, 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, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]])
一直以来,我们一直在使用包含 26 + 1 值的字典,其中额外的字符 "0" 用于对填充值进行一次性编码。我们现在可以删除对应"0"的第一列:
>>> x[..., 1:]
tensor([[[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, 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, 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, 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, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]])