Importing solve function from PyTorch fails for edge case version

Describe the bug

Hi there,

We’re running the PyTorch version labeled 1.7.1+cu110, but it is giving us issues when importing Kornia because of the following check in _compat.py:

if version.parse(torch.__version__) > version.parse("1.7.1"):
    # TODO: remove the type: ignore once Python 3.6 is deprecated.
    # It turns out that Pytorch has no attribute `torch.linalg` for
    # Python 3.6 / PyTorch 1.7.0, 1.7.1
    from torch.linalg import solve  # type: ignore
else:
    from torch import solve as _solve

    # NOTE: in previous versions `torch.solve` accepted arguments in another order.
    def solve(A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
        return _solve(B, A).solution

This check appears to believe that our version is > 1.7.1 and therefore attempts to import the wrong solve function.

Reproduction steps

1. Get PyTorch 1.7.1+anything (cu110, cu101, cu92, cpu) 
2. Attempt to import Kornia

Expected behavior

Importing Kornia with this version should not fail on a PyTorch version check. Perhaps it would make sense to split on + before checking the version.

Environment

- PyTorch Version (e.g., 1.0): 1.7.1+cu110
- OS (e.g., Linux): Ubuntu 20.04
- How you installed PyTorch (`conda`, `pip`, source): `pip`
- Build command you used (if compiling from source): -
- Python version: 3.7.5
- CUDA/cuDNN version: -
- GPU models and configuration: - 
- Any other relevant information: -

Additional context

No response

[mdjong1]

We have pytest as one of our dev dependencies, so that might be why it didn’t fail there

[edgarriba]

@pmeier you are right, my fault to include in main dependencies list. Problem of always going under development setup 😃 @mdjong1 this is the right place to include https://github.com/kornia/kornia/blob/master/setup.py#L115