torch.equals bug

[PetrochukM]

🐛 Bug

Torch equals does not work.

To Reproduce

>>> import torch
>>> t =torch.tensor([0.3000, 0.3000, 0.0000, 0.0000])
>>> t1 =torch.tensor([0.3000, 0.3000, 0.0000, 0.0000])
>>> torch.equal(t, t1)
True
>>> t1 =torch.tensor([0.3000, 0.3000, 0.0000, 0.0000]) + 1 - 1
>>> torch.equal(t, t1)
False
>>> t1
tensor([0.3000, 0.3000, 0.0000, 0.0000])

Expected behavior

That equality still works after some operations.

Environment

Collecting environment information...
PyTorch version: 1.0.0
Is debug build: No
CUDA used to build PyTorch: None

OS: Mac OSX 10.14.1
GCC version: Could not collect
CMake version: Could not collect

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] Could not collect

[fmassa]

I believe this is floating point error, does this work for double tensors?

[PetrochukM]

No.

>>> import torch
>>> t =torch.tensor([0.3000, 0.3000, 0.0000, 0.0000], dtype=torch.float64)
>>> t1 =torch.tensor([0.3000, 0.3000, 0.0000, 0.0000], dtype=torch.float64)
>>> torch.equal(t, t1)
True
>>> t1 = t1 + 1.0 - 1.0
>>> torch.equal(t, t1)
False
>>> t1
tensor([0.3000, 0.3000, 0.0000, 0.0000], dtype=torch.float64)

It'd be hard to imagine this being something other than a floating point error though.

Is there any support for almost_equal?

[fmassa]

I think there is something in torch.testing

[PetrochukM]

It's not publically documented, yeah? It'd be nice to have that public for writing tests. Using numpy.testing at the moment.

[eickenberg]

Definitely a floating-point issue, since it works with [0.1250, 0.1250, 0.0000, 0.0000] which has a finite expansion in binary (1/2**3), contrary to 0.3.

Maybe a robust workaround is something like torch.norm of difference tested to be under a threshold

[PetrochukM]

Thanks guys!