Good number of "sets" to train to.

[rohskopf]

I'm training a potential for a single system, with a few different sets inside. Does it make a difference whether or not I split this system into many different sets, or just a few small number of sets? What effect does this have on training?

[jameswind]

For each system, the last set is used as testing data. If there is only one set, it will purely be used for training. There are only small effects like how data is loaded and randomly picked in the sgd optimization process.

…

On Tue, Apr 21, 2020 at 3:05 PM Drew Rohskopf ***@***.***> wrote: I'm training a potential for a single system, with a few different sets inside. Does it make a difference whether or not I split this system into many different sets, or just a few small number of sets? What effect does this have on training? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/deepmodeling/deepmd-kit/issues/217>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEJ6DC32PRHE6FZJHZX445LRNXVAPANCNFSM4MNRF6QA> .

[dkussainova]

If there is only one set and it is purely used for training, what is used for testing and how is data selected? I have only one data set, and I can see learning curves both for training and testing sets.