TorchEnvironmentProvider

[Dom1L]

Hi,

this PR contains an environment provider based on the neighborpair implementation of TorchANI. I was talking about it in issue #96. It can be used either on a CPU or GPU and provides a speedup compared to the ASE implementation (tested on a TitanV with a 352 atom structure):

• 16.3 ms Torch
• 1.12 s ASE
  The switch to numpy after the calculation of the indices does not seem to bother the speed, as I also tried a Torch version of this.

A downside is that it requires a fairly new Pytorch version that contains new functions (for example cartesian products). I used Pytorch 1.1.0.dev20190411 for this one. So one would need to add that to the requirements to pass the tests. I didn't do it because I wasn't sure how you want to handle that.

I also added the TorchEnvironmentProvider to the existing tests for the environments and made sure it passes them. However, the returned neighborlists are sorted in contrast to the ASE ones, so that might trigger some asserts when testing with larger structures. I am also not an expert with neighborlists, so I'm not sure whether the different Torch and ASE implementations cause any problems. But as it is right now, it might be a nice option to have.

Hope it helps and if there is anything else let me know

[ktschuett]

We have to figure out how to do this with the torch version. Do you know when this will be released? Perhaps, we hold this back until then...

[Dom1L]

I totally agree, this was also my biggest concern.
A quick solution would be to just avoid these new functions, but I think holding it until the 1.1 release is completely fine, as this is not an urgent feature. I just don't know when it's coming though...