generate sequence by indices / one-hot encoding

Question

I have a sequence s = [4,3,1,0,5] and num_classes = 6 and I want to generate a Numpy matrix m of shape (len(s), num_classes)

Answer 1

Brocasting still works just do

$ (labels[:,:,:,None]==np.arange(num_classes))+0

Answer 2

You can use broadcasting -

(np.array(s)[:,None]==np.arange(num_classes))+0

Sample run -

In [439]: s
Out[439]: [4, 3, 1, 0, 5]

In [440]: num_classes = 9

In [441]: (np.array(s)[:,None]==np.arange(num_classes))+0
Out[441]: 
array([[0, 0, 0, 0, 1, 0, 0, 0, 0],
       [0, 0, 0, 1, 0, 0, 0, 0, 0],
       [0, 1, 0, 0, 0, 0, 0, 0, 0],
       [1, 0, 0, 0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 1, 0, 0, 0]])

Answer 3

Since you want a single 1 per row, you can fancy-index using arange(len(s)) along the first axis, and using s along the second:

s = [4,3,1,0,5]
n = len(s)
k = 6
m = np.zeros((n, k))
m[np.arange(n), s] = 1
m
=> 
array([[ 0.,  0.,  0.,  0.,  1.,  0.],
       [ 0.,  0.,  0.,  1.,  0.,  0.],
       [ 0.,  1.,  0.,  0.,  0.,  0.],
       [ 1.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  1.]])

m.nonzero()
=> (array([0, 1, 2, 3, 4]), array([4, 3, 1, 0, 5]))

This can be thought of as using index (0,4), then (1,3), then (2,1), (3,0), (4,5).

Answer 4

The accepted answer won't work if your one-hot encoding adds an extra dimension to a multidimensional array. Using broadcasting will give you unexpected results - https://scipy.github.io/old-wiki/pages/Cookbook/Indexing . This solution is elegant but not very efficient.

labels.shape # (80, 256, 10)

def b(labels):
     onehot = np.zeros((a,b,c,num_classes), dtype=float)

     # This is the slow, dumb line:
     (onehot_i, onehot_j, onehot_k) = np.ones(labels.shape).nonzero()

     thehotone = labels[onehot_i, onehot_j, onehot_k]
     onehot[onehot_i, onehot_j, onehot_k, thehotone] = 1
     return onehot