Stress Calculation in ForceRegressionTask
#302
Conversation
|
I personally prefer |
| volume = torch.linalg.det(cell).abs().unsqueeze(-1) | ||
| stress = virials / volume.view(-1, 1, 1) | ||
| stress = torch.where( | ||
| torch.abs(stress) < 1e10, stress, torch.zeros_like(stress) |
There was a problem hiding this comment.
What is this hard coded value?
For something like this, I don't actually know if it's actually more performant to just clip values, instead of using torch.where where you have to allocate a zeros_like tensor.
There was a problem hiding this comment.
this is an inspired bit from here. im assuming it is helpful in tasks which compute losses from stress. happy to remove the line and reconsider adding it back if/when we do stress related training.
|
For now this is fine, but I have to say that |
I agree, it could use some organizing. |
This PR adds the ability to perform stress calculations inside the
ForceRegressionTask, and is inspired by the implementation here: https://github.com/ACEsuit/mace/blob/main/mace/modules/utils.py. This functionality if usefull in downstream tasks that require this parameter, and will be helpful for future MatSciML tasks such as aStressRegressionTaskor similar.