Allow for specification of a train set size

[jwa7]

For demo purposes, I am trying to train a soap-bpnn on a small subset of qm7 (> 7000 structures) on my laptop.

One can specify in options.yaml the proportionate size of val_set and test_set, but cannot do so for the training set. As far as I understand, the train set size is inferred as the remaining proportion. In my case, I can make training faster by setting test_set: 0.999 for instance, but this of course makes post-training evaluation very slow.

In my case, if I want to train and test on a smaller subset it would require me to manually construct a smaller .xyz to pass as the input file. This is of course trivial, but having a way to specify a training size could be more convenient. For instance, allow setting train_set too, and allow train_set + val_set + test_set < 1.

Suppose I want to generate a learning curve, with randomly shuffled training and validation data of different sizes (i.e. different runs with different random seeds), but a fixed test set. Can I do this with the current setup? Is it possible to point to a different hold out .xyz file as the test set?