Philipp Thölke

Loss is computed in `step`, no matter if during train, val or test. The function `validation_step` is called by pytorch-lightning to evaluate the model on all validation dataloaders. `test_step` is...

@peastman maybe this helps with your idea: https://pytorch-lightning.readthedocs.io/en/latest/guides/speed.html#set-validation-check-frequency-within-1-training-epoch

The model outputs force predictions. https://github.com/torchmd/torchmd-net/blob/main/torchmdnet/models/model.py#L207 I agree that the naming is inconsistent but as far as I remember, when I implemented QM9, MD17 and ANI-1, they all loaded forces...

16 bit floats?

> Batch size seems to have very little effect on memory use. With 5 layers I can use batch size 100. Add a sixth layer and it runs out of...

It applies always, not only during training. The argument determines the maximum number of neighbors that are collected in the neighbor list algorithm. You can overwrite it when you load...

I think that should work but better make sure it actually overwrites it. I'd recommend using the `torchmdnet.models.model.load_model` function to load the model for inference, which strips away the pytorch...

Have you tried the new CUDA Graph API released in PyTorch 1.10 yet? This might already reduce the CPU overhead. Otherwise I saw a decent inference speedup when using PyTorch's...

> To keep the GPU code simpler, I'd like to minimize the number of options that affect the code. TorchMD_ET has a lot of options. Is it reasonable to assume...

Where did you find the discontinuity? This is the cutoff function with lower cutoff 1 and upper cutoff 5.  In case you were looking at the exponential normals with...