About the input training dataset

[0WellyWang0]

Hi. I have a question about the dataset for training. For example, if I would like to find a "universal" potential of Fe and O system, could I combine and train the trajectory of Fe3O4.xyz, FeO.xyz, and Fe2O3.xyz together and finally produce a single result? I got problems when doing this kind of work: 1000 data points from Fe3O4.xyz + 1000 from FeO.xyz + 1000 from Fe2O3.xyz led to a combined.xyz file, but allegro only took the first 1000 data points from Fe3O4.xyz to train. Thanks.

[Linux-cpp-lisp]

Hi @0WellyWang0 ,

The files you provide determine what data is available to pick from, but you still need to set n_train and n_val to control how many data points are taken for the training and validation sets respectively.

[0WellyWang0]

Thanks. Just a more specific question on the code: since FeOtraj.xyz and Fe2O3traj.xyz contained different compositions of Fe and O, shall I load them separately with 2 command sets and set n_train and n_val to train together like the following?

# -- data --
dataset: ase                                                                   
dataset_file_name:     FeOtraj.xyz                #1000 data points
ase_args:
  format: extxyz

dataset: ase                                                                   
dataset_file_name:     Fe2O3traj.xyz                #1000 data points
ase_args:
  format: extxyz

# training
n_train: 1600
n_val: 200

[Linux-cpp-lisp]

No, you should merge all your training data into one XYZ file.

It is possible to use separate dataset definitions for the training dataset and validation dataset:

dataset: ase                                                                   
dataset_file_name:     FeOtraj.xyz                #1000 data points
ase_args:
  format: extxyz
                                                                  
validation_dataset_file_name:     somethingelse.xyz                #1000 data points

[0WellyWang0]

I got the point and thanks a lot. Good luck with your work.