One-Hot Encoding

Turn integer class labels into one-hot rows.

Given labels of shape (n,) with values in [0, num_classes), return an (n, num_classes) array where row i is all zeros except for a 1.0 in column labels[i].

Feeding raw integer labels to a model implies an ordering that is not there: class 3 is not three times class 1, and it is not "between" 2 and 4. One-hot removes that by giving every class its own axis.

Do it with advanced indexing rather than a loop. Allocating zeros and assigning out[np.arange(n), labels] = 1.0 is the idiomatic one-liner, and it is the same indexing trick the cross entropy problem uses to pull out the true class probabilities.

Function signature

def one_hot(labels, num_classes):

Example

Input

one_hot(np.array([0, 2, 1]), 3)

Output

array([[1., 0., 0.],
       [0., 0., 1.],
       [0., 1., 0.]])

Reasoning

Each row has a single 1.0 in the column given by its label: column 0, then column 2, then column 1.

Constraints and notes

• Return a float array of shape (n, num_classes).
• Use advanced indexing, not a loop.