Philipp Thölke

I would say the API in this PR is more general and therefore preferable but I think it's still useful to keep e.g. the distinction between `prior_args` and `prior_init_args` from...

`BasePrior.forward()` returns the updated atomwise energy predictions, not only the prior contribution (see the `Atomref` prior as an example). It would be possible to add the prior divided by the...

### Multi-head implementation Yes, currently models can only have a single output head. I think adding multi-head support could be useful, however, requires changes to the current training interface and...

And yes, through the config file and model creation/loading you are currently limited to just one prior model, however, technically it would be possible to wrap the model in multiple...

You could simply pass `args` to the prior model in the same step as passing the dataset. If needed you could add an `initial_prior_args` of some sorts that can maybe...

Sounds good! I think version two is way more readable, which I think is what we should be going for in the config files. I reckon both versions will be...

Ah, sorry for the duplicate, moving to #2013.

Alright, since I can't reopen the other issues I'll stay here :D Any plans on addressing this issue? If you are sure that `mvCleanupViewport` provides the intended functionality I could...

The model's `forward` already has arguments for charge and spin which could be passed to the prior models without a problem. https://github.com/torchmd/torchmd-net/blob/7ea833d8ab1e1c6f072b1d83ebb423b83b8ae909/torchmdnet/models/model.py#L192-L199 Currently that requires setting `--charge` to true in...

The solution with the reference node works for me by simply initializing it far away where it is unlikely to be found, since hiding causes a Segmentation Fault (#2151). Then...