Torch.ge not producing reasonable result with torch.int tensor

1

It seems that when using torch.ge on a torch.int tensor, the result is not reasonable.
What’s the reason?

Pytorch version: 0.4.0
Host: Ubuntu 16.04
Python: Anaconda

Test code:

import torch
import numpy as np

# Create a torch.int tensor from numpy
a_np = np.asarray([0, 1, 1, 0, 0])
a = torch.from_numpy(a_np)

# Create a torch.float tensor directly
b = torch.FloatTensor([0, 1, 1, 0, 0])

# Perform torch.ge on a and b
print(torch.ge(a, 0.5))   # Output: tensor([ 1,  1,  1,  1,  1], dtype=torch.uint8)
print(torch.ge(b, 0.5))   # Output: tensor([ 0,  1,  1,  0,  0], dtype=torch.uint8)

The workaround is to call a.float() after torch.from_numpy(.), then it performs exactly like the case with b.

2

Likely we convert the scalar to the dtype of the tensor somewhere in the calling path, then 0.5 becomes 0. Please file a bug report / feature request to github.com/pytorch/pytorch.

3

Thank you for the response. Bug reported here.