Philipp Thölke

The cutoff function with lower cutoff > 0 is implemented as  as you can see here https://github.com/compsciencelab/torchmd-net/blob/28fdcb544af651d95c7e0e1ba1ac8babc21e9d65/torchmdnet/models/utils.py#L170-L184 It is not relevant for the paper as we never use a...

> We also need the graph network, which is effectively the same architecture as SchNet. Depending on some arguments, yes. The following corresponds to the original SchNet architecture: `TorchMD_GN(rbf_type="gauss", trainable_rbf=False,...

We actually start distributing the RBFs only beyond the lower cutoff in `GaussianSmearing` and `ExpNormalSmearing`. This does not require changing the functional form and in fact we don't even use...

We do not want to throw away self interactions. I believe the locations where this is currently happening with lower cutoffs larger 0 are your 1st, 5th and 6th points....

Initially I thought torch geometric was requiring setting `y` when constructing a `Data` object but now I'm not sure anymore. I agree that `energy` and `forces` makes more sense than...

Do you mean renaming both `y` and `dy` are confusing or would you want `dy` to be `forces` and `y` to stay as it is?

Changing the model to return the positive gradient now would silently break a lot of applications, I don't think that's a good idea. I'd be okay with calling it `y`...

I updated (hopefully) all places to consistently use `y` as the label/prediction and `neg_dy` as the negative derivative of the label/prediction.

I haven't tried using TorchMD-Net in C++ so I don't know. You could try breaking the model down and exporting submodules to narrow down the issue. Maybe also trying just...

I think both #134 and this PR will be useful and as far as I understand it they're not mutually exclusive? @peastman or is the conflict more substantial? I like...