How to transform a multi-hot tensor to a one-hot tensor?

vision
1

for example,

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

and i want get

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

what is the elegant way to achieve this operation in pytorch?

2

Hi Yu!

Use nonzero() to find locations of the 1s in your multi-hot tensor,
and then (after some fiddling to get the indices in the right form) use
one_hot() to produce your desired one-hot result:

>>> import torch
>>> torch.__version__
'1.10.0'
>>> multihot = torch.tensor ([[1, 0, 1, 0, 0, 0], [0, 0, 1, 0, 1, 0]])
>>> count = multihot[0].sum()   # number of ones in first row                     >>> assert  (multihot.sum (dim = 1) == count).all()   # verify same number of ones in each row
>>> onehot = torch.nn.functional.one_hot (multihot.nonzero()[:, 1].view (multihot.size (0), count))
>>> onehot
tensor([[[1, 0, 0, 0, 0],
         [0, 0, 1, 0, 0]],

        [[0, 0, 1, 0, 0],
         [0, 0, 0, 0, 1]]])

Best.

K. Frank

3

Hi K.Frank!
Thanks for your reply.
But i notice that the shape of multihot is 2, 6, but the one hot is 2, 2, 5.
the final dim is not equl in above.
Best,
Yu

4

Hi Yu!

Yes, you are correct. That was my oversight.

We can explicitly tell one_hot() the number of classes, rather than
having it “guess:”

>>> onehot = torch.nn.functional.one_hot (multihot.nonzero()[:, 1].view (multihot.size (0), count), num_classes = multihot.size (1))
>>> onehot
tensor([[[1, 0, 0, 0, 0, 0],
         [0, 0, 1, 0, 0, 0]],

        [[0, 0, 1, 0, 0, 0],
         [0, 0, 0, 0, 1, 0]]])

Note the use of num_classes = multihot.size (1).

Best.

K. Frank