torch.nn.functional.grid_sample outputs NaN

[DNXie]

🐛 Bug

torch.nn.functional.grid_sample outputs NaN if grid contains large value

To Reproduce

import torch
torch.nn.functional.grid_sample(input=torch.ones([1,1,1,5]), grid=torch.tensor([[[[ 2.9839e+38, -3.2406e+38]]]]))

Output:

tensor([[[[nan]]]])

Expected behavior

Expect a grace exception message if the parameters are invalid or unexpected

Environment

PyTorch version: 1.8.0.dev20210126+cpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 18.04.4 LTS (x86_64)
GCC version: Could not collect
Clang version: Could not collect
CMake version: Could not collect

Python version: 3.7 (64-bit runtime)
Is CUDA available: False
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A

cc @albanD @mruberry @jbschlosser

[mruberry]

I think this is expected. We typically don't (or can't) raise exceptions based on tensor values (although we do in some cases on the CPU).

[albanD]

@mruberry from the doc, any value value of the grid outside [-1, 1] should return the padding. In this case, it is zero padding so it sounds ok to expect a 0 for this kind of grid?

[mruberry]

@mruberry from the doc, any value value of the grid outside [-1, 1] should return the padding. In this case, it is zero padding so it sounds ok to expect a 0 for this kind of grid?

Could be.