Is your feature request related to a problem? Please describe.
I am fine-tuning a PET model on a very big training set (>1M structures). At the end of the training, mtt train takes 1 hour 20 minutes to compute the RMSE and MAE on the whole training set.
However, as far as I know mtt train cannot save the labels predicted in this calculation to a file; in order to do this, I need to invoke mtt eval and repeat the expensive evaluation of all labels in the training set.
Describe the solution you'd like
I think a nice improvement might be the following: an optional flag in mtt train that collects data when the final error metrics are evaluated at the end of the training and saves the predicted labels from the training, validation, and test sets in .xyz or memmap format. In this way, it is not necessary to recompute them with an additional, possibly expensive, mtt eval calculation.
Describe alternatives you've considered
No response
Additional context
No response
Is your feature request related to a problem? Please describe.
I am fine-tuning a PET model on a very big training set (>1M structures). At the end of the training, mtt train takes 1 hour 20 minutes to compute the RMSE and MAE on the whole training set.
However, as far as I know mtt train cannot save the labels predicted in this calculation to a file; in order to do this, I need to invoke mtt eval and repeat the expensive evaluation of all labels in the training set.
Describe the solution you'd like
I think a nice improvement might be the following: an optional flag in mtt train that collects data when the final error metrics are evaluated at the end of the training and saves the predicted labels from the training, validation, and test sets in .xyz or memmap format. In this way, it is not necessary to recompute them with an additional, possibly expensive, mtt eval calculation.
Describe alternatives you've considered
No response
Additional context
No response