Possible bug when combining numpy function with Variable()

[manuelhaussmann]

Pytorch shows the following inconsistent behaviour.

import numpy as np
import torch
from torch.autograd import Variable

1.0 + Variable(torch.ones(1)) 
# returns as expected
# Variable containing:
#  2
# [torch.FloatTensor of size 1]

np.sum(1.0) + Variable(torch.ones(1)) 
# returns an unexpected (depth of array is 32)
# array([[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[Variable containing:
#  2
# [torch.FloatTensor of size 1]
# ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]], dtype=object)

# Switching their order 
Variable(torch.ones(1)) + np.sum(1.0)
# returns the expected
# Variable containing:
#  2
# [torch.FloatTensor of size 1]

[apaszke]

While I agree that the behavior is very counterintuitive there's little what we can do (except for monkey-patching NumPy, which doesn't seem like an appealing option).