In Torch how do I create a 1-hot tensor from a list of integer labels?

Question

I have a byte tensor of integer class labels, e.g. from the MNIST data set.

 1
 7
 5
[torch.ByteTensor of size 3]

How do use it to create a tensor of 1-hot vectors?

 1  0  0  0  0  0  0  0  0  0
 0  0  0  0  0  0  1  0  0  0
 0  0  0  0  1  0  0  0  0  0
[torch.DoubleTensor of size 3x10]

I know I could do this with a loop, but I'm wondering if there's any clever Torch indexing that will get it for me in a single line.

Answer 1

indices = torch.LongTensor{1,7,5}:view(-1,1)
one_hot = torch.zeros(3, 10)
one_hot:scatter(2, indices, 1)

You can find the documentation for scatter in the torch/torch7 github readme (in the master branch).

Answer 2

An alternate method is to shuffle rows from an identity matrix:

indicies = torch.LongTensor{1,7,5}
one_hot = torch.eye(10):index(1, indicies)

This was not my idea, I found it in karpathy/char-rnn.