Training on dataset with different sets of data

[alikhamze]

Hello,

I have a dataset where one portion includes virial data and the other does not.

If I train deepmd using the full dataset and want it to predict the virial as well, how are the data points without virial data handled? Will having the mixed dataset affect my virial predictions (beyond having fewer virial data points than force and energy data points)?

Thanks,
-Ali

[njzjz]

If a data set doesn't contain virials, the total loss for this data will not have the virial part.

We note that there is a known bug that virial loss will still exist in lcurve.out even if the total loss doesn't contain the virial loss.

[alikhamze]

Hi @njzjz, thank you for the reply.

So it sounds like training should proceed without an issue with the mixed dataset, and that the due to the bug the rmse_v_train column of the lcurve.out will always be printed, even when the data points in the batch don't have virial data?

[njzjz]

Yes

[alikhamze]

great, thank you!