The energy and force values won't update

[mozhde]

Hello,
I was playing around with the example to run a Qbc for a system of mine. My energy and force files are in CP2K format, and when I run the run-Qbc.py it generates the RuNNer format file properly. However, when I look at the "eval_train" and "predict" folders, I can see that the values of energy and forces in the "input.data" file are zero. Would you please help me find out what could go wrong?

Thank you

[cschran]

Dear mozhde,

this behavior is actually expected as the eval_train and predict folders are essentially intermediate directories within the QbC process of AML. The actual reference energies are saved internally for every structure and could be accessed within python, e.g. by a loop like this:

for s in structures:
    print(s.properties['reference'].energy)

Similarly, there will also be properties for the various individual NNP models under the key predict-00X.

If you want to get the force error of your model, you can make use of the functionality in scores/forces.py or have a look at the score-all.py script under from-trajectroy-to-model/05-validation.

[mozhde]

Dear Christoph Thanks for your reply; it was constructive. You are right; we don't need the reference values of the forces and energies to evaluate the prediction of the committee members with respect to each other. If it is fine with you, I have another question to ask you; Suppose I want to get the score given by function " evaluate_cda_forces" for an already trained committee. I can see you did it in class "Qbc," but you've defined a new blank committee and don't use an already trained one. More precisely, I want to query a committee outside the "Qbc" iteration to pick the highest disagreement for my data manually. Would it be possible? Regards

…

On Tue, Oct 5, 2021 at 1:39 PM Christoph Schran ***@***.***> wrote: Dear mozhde, this behavior is actually expected as the eval_train and predict folders are essentially intermediate directories within the QbC process of AML. The actual reference energies are saved internally for every structure and could be accessed within python, e.g. by a loop like this: for s in structures: print(s.properties['reference'].energy) Similarly, there will also be properties for the various individual NNP models under the key predict-00X. If you want to get the force error of your model, you can make use of the functionality in scores/forces.py or have a look at the score-all.py script under from-trajectroy-to-model/05-validation. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADLHGS72GNDUG4MS4YVFJ53UFLWWLANCNFSM5FLSWGPA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[cschran]

If you have obtained the prediction of the structures, you can use the function select_highest_error, e.g. as done in the QbC procedure:

idx_selected = structures_candidate.select_highest_error(
    n=n_add, label_prop='pred*', f_reduce=moment_std, name_prop='forces')

This also allows you to change the function f_reduce that reduces the force disagreement into a single score per structure.