这是一个与维度无关的独立解决方案。

这将把任何非负整数的N维数组arr转换为一个N+1维数组one_hot，其中one_hot[i_1，…，i_N,c] = 1表示arr[i_1，…，i_N] = c.可以通过np恢复输入。argmax (one_hot, 1)

def expand_integer_grid(arr, n_classes):
    """

    :param arr: N dim array of size i_1, ..., i_N
    :param n_classes: C
    :returns: one-hot N+1 dim array of size i_1, ..., i_N, C
    :rtype: ndarray

    """
    one_hot = np.zeros(arr.shape + (n_classes,))
    axes_ranges = [range(arr.shape[i]) for i in range(arr.ndim)]
    flat_grids = [_.ravel() for _ in np.meshgrid(*axes_ranges, indexing='ij')]
    one_hot[flat_grids + [arr.ravel()]] = 1
    assert((one_hot.sum(-1) == 1).all())
    assert(np.allclose(np.argmax(one_hot, -1), arr))
    return one_hot

下面是我根据上面的答案和我自己的用例写的一个示例函数:

def label_vector_to_one_hot_vector(vector, one_hot_size=10):
    """
    Use to convert a column vector to a 'one-hot' matrix

    Example:
        vector: [[2], [0], [1]]
        one_hot_size: 3
        returns:
            [[ 0.,  0.,  1.],
             [ 1.,  0.,  0.],
             [ 0.,  1.,  0.]]

    Parameters:
        vector (np.array): of size (n, 1) to be converted
        one_hot_size (int) optional: size of 'one-hot' row vector

    Returns:
        np.array size (vector.size, one_hot_size): converted to a 'one-hot' matrix
    """
    squeezed_vector = np.squeeze(vector, axis=-1)

    one_hot = np.zeros((squeezed_vector.size, one_hot_size))

    one_hot[np.arange(squeezed_vector.size), squeezed_vector] = 1

    return one_hot

label_vector_to_one_hot_vector(vector=[[2], [0], [1]], one_hot_size=3)

我添加了一个简单的补全函数，只使用numpy操作符:

   def probs_to_onehot(output_probabilities):
        argmax_indices_array = np.argmax(output_probabilities, axis=1)
        onehot_output_array = np.eye(np.unique(argmax_indices_array).shape[0])[argmax_indices_array.reshape(-1)]
        return onehot_output_array

它以一个概率矩阵作为输入:例如:

[[0.03038822 0.65810204 0.16549407 0.3797123] . [0.02771272 0.2760752 0.3280924 0.33458805]

它会返回

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

我认为简短的答案是否定的。对于n维的更一般的情况，我想到了这个:

# For 2-dimensional data, 4 values
a = np.array([[0, 1, 2], [3, 2, 1]])
z = np.zeros(list(a.shape) + [4])
z[list(np.indices(z.shape[:-1])) + [a]] = 1

我想知道是否有更好的解决方案——我不喜欢我必须在最后两行创建这些列表。不管怎样，我用timeit做了一些测量，似乎基于numpy的(索引/范围)和迭代版本的表现是一样的。

def one_hot(n, class_num, col_wise=True):
  a = np.eye(class_num)[n.reshape(-1)]
  return a.T if col_wise else a

# Column for different hot
print(one_hot(np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 9, 9, 9, 8, 7]), 10))
# Row for different hot
print(one_hot(np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 9, 9, 9, 8, 7]), 10, col_wise=False))

如果使用tensorflow，则存在one_hot():

import tensorflow as tf
import numpy as np

a = np.array([1, 0, 3])
depth = 4
b = tf.one_hot(a, depth)
# <tf.Tensor: shape=(3, 3), dtype=float32, numpy=
# array([[0., 1., 0.],
#        [1., 0., 0.],
#        [0., 0., 0.]], dtype=float32)>