Force loss function
If I'm reading the loss function code, the force residuals are being divided by the number of atoms. Since force is already a per-atom quantity, this should not be the case.
Comments (6)
-
-
Now that changes seem to be working, tests are broken.
-
Here is a photo of our discussion today:
-
reporter Thanks! That makes sense. But now I have another question: I can see where the code divides by the number of atoms here. But where does it divide by the number of images?
-
reporter Nevermind -- we found it. I understand now that
force_loss
is the SSR andforce_rmse
is the RMSE, and the RMSE is only calculated withincheck_convergence
, where the info is reported to the user. -
reporter - changed status to resolved
Re-arrange loss function code.
Should be functionally identical, but a bit more efficient. (But probably wasn't rate-limiting.) Also makes code more logical to read in my opinion. Fixes Issue
#218.→ <<cset e5beac6ddbd4>>
- Log in to comment
I think PR #36 should solve this.