Simplify arguments and reintroduce aggregation mode

[ktschuett]

Resolves #1:

• property argument is only needed during training call
• old "pool_mode" is reintroduced to cmd args as "aggregation_mode" (more fitting name)

[cod3licious]

I think at least for schnetpack_qm9.py this doesn't fix everything yet. When evaluate() is called, it is given args instead of train_args but uses args.property. However, it can't just be switched like that, as it also needs args.modelpath, which is evaluation specific (should it be? are folder names frequently changed between training and evaluation?).

Also I guess --logger and --log_every_n_epochs are also only relevant for training and not as general arguments.

Maybe in general at the start of the script in eval mode, train_args could be loaded and then only the "active" arguments of the current regular args could be used to replace the corresponding fields in train_args, and then everywhere only use train_args and not this mix between both, which makes it kind of hard to keep track of what goes where?

[ktschuett]

Ok, then we need to pass properties separately. I will take a look later and also move the logger and remove the model-parameters from eval since they are fixed with the trained model.

Of course the model path needs to be specific to the evaluation, otherwise, how should we know where the model is stored?

[cod3licious]

yes, of course you need to pass the modelpath when calling eval, but the path itself should still be the same as when the model was trained. And I think the modelpath and maybe flags like --cuda might be the only thing that really need to be given for eval? The data as well should still be at the same place as when training the model, no?

[mgastegger]

Looks fine.