🐛 Bug

The weight of positive examples should default to 1.. However, there are inputs for which the result is finite without specifying pos_weight and nan with pos_weight=torch.tensor([1.]).

To Reproduce

Steps to reproduce the behavior:

import torch, math 
import torch.nn.functional as F

logits = torch.tensor([0., -120.])
target = torch.tensor([0., 1.])

# passes
assert math.isfinite(
    F.binary_cross_entropy_with_logits(logits, target).item()
)

pos_weight = torch.tensor([1., 1.])

# fails
assert math.isfinite(
    F.binary_cross_entropy_with_logits(logits, target, pos_weight=pos_weight).item()
)

Note that for moderate inputs (-10 instead of -120) this works as expected

Expected behavior

Both cases should lead to the same result, because the latter is just explicitly passing the default value.

Environment

PyTorch version: 1.0.0.dev20181108
Is debug build: No
CUDA used to build PyTorch: None

OS: Ubuntu 18.04.1 LTS
GCC version: (Ubuntu 7.3.0-27ubuntu1~18.04) 7.3.0
CMake version: version 3.10.2

Python version: 3.6
Is CUDA available: No
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA

Versions of relevant libraries:
[pip] Could not collect
[conda] pytorch-nightly-cpu 1.0.0.dev20181108 py3.6_cpu_0 pytorch
[conda] torch-dimcheck 0.0.1
[conda] torchvision 0.2.1

Additional context

I find the same issue when running on GPU