torchmd-net
torchmd-net copied to clipboard
torchmd
Neural network potentials
Hello, I am running into a problem with using torchscript to integrate a trained tensornet model with openmm for dynamics. This is in the newest version of the code as...
This PR proposes several enhancements to enable the usage of extra arguments in both training and inference. The key modifications include - [x] standardizing the naming conventions for all extra...
I could not make `Union[ Tuple[Tensor,Tensor], Tuple[Tensor,Tensor,Tensor,Tensor]]` work. It will jit.scipt but then if I try and use the model as energy,_ = model(..) it will complain: ``` RuntimeError: Union[Tuple[Tensor,...
Allowing vector output from TensorNet. This will allow the use of `EquivariantScalar` or `EquivariantVectorOutput` from `TensorNet`. Resolves #297.
I was trying to use [gradient checkpointing](https://pytorch.org/docs/stable/checkpoint.html) with TorchMD model. For some reason, I get this gradient mismatch whenever the warning pops up https://github.com/torchmd/torchmd-net/blob/8a1be712a397d88eed82ae12066f1533f12152ee/torchmdnet/models/utils.py#L492-L499
Pytorch introduced a new API to handle extensions, it is "documented" here: https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit It makes it possible to write meta registrations for C++ extensions, which I could not make before....
Currently it is not possible to run backwards twice with torch.compile. For instance, this code fails: ```python from torch import nn, Tensor class Model(nn.Module): def forward(self, input: Tensor) -> Tensor:...
##### Feature Request Currently the vector output is set to `None`. This could be easily calculated based on skew symmetric matrix based on the paper. Some papers suggest using force...
Check version when loading a model. Print a warning if the checkpoint is loaded with a version different to the one used to create it. Closes #291
This feature lets you inject extra per-sample or per-atom values into the model. They are appended to the embedding vector for each atom, then a linear layer mixes them in...
