Doubt About Validation Parameter

[AnuragKr]

Hi all,

Q1. I have one trivial doubt regarding the validation parameter. As the validation parameter is optional If it is set to None then How validation data is being obtained?
Q2. As In most of the papers training size is being mentioned but not validation data. What should be standard criterial to keep size for validation data.
Ex - If total snapshots is 12000 then in what percentage we should split train | val | test with respect to atomistic modelling simulation. Genrally we keep train and test 70-30% then from 70% train data we further divide into train and val.
Q3. In paper if it is mentioned that 95000 molecules for training and remaining for testing - Does it imply that from 9500 molecules data deepmd itself will split idata nto train and validation data.

Thanks

[AnuragKr]

Answer based on my findings --
Answer 1. -- Few Training data are being considered as validation data.
Answer 2. -- It is suggested as 95 % Training data and 5 % Validation data.
Answer 3 -- No 9500 will be taken as training data and the remaining data for test and validation purposes. As per my understanding, DeePMD treats validation and test data the same.

Please feel free to comment and reopen the discussion if you have some other perspective.