Convert int into one-hot format

4

Hi,

You can use the scatter_ method to achieve this.
I would also advise to create the y_onehot tensor once and then just fill it:

import torch

batch_size = 5
nb_digits = 10
# Dummy input that HAS to be 2D for the scatter (you can use view(-1,1) if needed)
y = torch.LongTensor(batch_size,1).random_() % nb_digits
# One hot encoding buffer that you create out of the loop and just keep reusing
y_onehot = torch.FloatTensor(batch_size, nb_digits)

# In your for loop
y_onehot.zero_()
y_onehot.scatter_(1, y, 1)

print(y)
print(y_onehot)
56 Likes