Question about Softmin in different version

1

There is a question about Softmin function in different version of Pytorch.

In version 0.4.0, if I do the following.

import torch.nn.functional as F

a = torch.randn(3, 4)
b = F.softmin(a, dim=-1).sum(dim=-1)

b is equal to tensor([ 1.,  1.,  1.])

In version 0.4.1,

a = torch.randn(3, 4)
b = F.softmin(a, dim=-1).sum(dim=-1)

b is equal to tensor([-1.0000, -1.0000, -1.0000])

Is this a bug ?

I did’t have version 1.0 installed, if anyone knows way, please let me know.

2

Hello Kelley!

This appears to be a known and fixed bug: See this github issue:

(For what it’s worth, when I run your softmin test on pytorch
version 1.0.1, I get the right answer, that is, the softmin values
sum to 1.)

Best.

K. Frank

1 Like