metatensor / metatrain

Training and evaluating machine learning models for atomistic systems.
https://metatensor.github.io/metatrain/
BSD 3-Clause "New" or "Revised" License
16 stars 4 forks source link

Option for dumping training/test/val split #378

Open HannaTuerk opened 2 hours ago

HannaTuerk commented 2 hours ago

Hi, for the validation of my model, I wanted to do a parity plot. Currently there is no option for dumping the train/test/validation split the models do internally. I have been discussing this with @abmazitov , and we came up with 2 possible solutions:

1) add a tutorial on that if the splits are necessary, the user should do them before training and feed the split in the model

2) Provide the option to dump the indices of the split as e.g. numpy arrays (not the files directly as they can be large for large datasets), maybe even just in the model evaluation with a separate keyword

Thoughts welcome

frostedoyster commented 2 hours ago

I would add the tutorial, but saving the npy in the outputs directory is also an option