TorchMD-NET provides state-of-the-art neural networks potentials (NNPs) and a mechanism to train them. It offers efficient and fast implementations if several NNPs and it is integrated in GPU-accelerated molecular dynamics code like ACEMD, OpenMM and TorchMD. TorchMD-NET exposes its NNPs as PyTorch modules.
Documentation is available at https://torchmd-net.readthedocs.io
TorchMD-Net is available in conda-forge and can be installed with:
mamba install torchmd-netWe recommend using Mamba instead of conda.
TorchMD-Net is installed using pip, but you will need to install some dependencies before. Check this documentation page.
Specifying training arguments can either be done via a configuration yaml file or through command line arguments directly. Several examples of architectural and training specifications for some models and datasets can be found in examples/. Note that if a parameter is present both in the yaml file and the command line, the command line version takes precedence.
GPUs can be selected by setting the CUDA_VISIBLE_DEVICES environment variable. Otherwise, the argument --ngpus can be used to select the number of GPUs to train on (-1, the default, uses all available GPUs or the ones specified in CUDA_VISIBLE_DEVICES). Keep in mind that the GPU ID reported by nvidia-smi might not be the same as the one CUDA_VISIBLE_DEVICES uses.
For example, to train the Equivariant Transformer on the QM9 dataset with the architectural and training hyperparameters described in the paper, one can run:
mkdir output
CUDA_VISIBLE_DEVICES=0 torchmd-train --conf torchmd-net/examples/ET-QM9.yaml --log-dir output/Run torchmd-train --help to see all available options and their descriptions.
See here for instructions on how to load pretrained models.
If you want to train on custom data, first have a look at torchmdnet.datasets.Custom, which provides functionalities for
loading a NumPy dataset consisting of atom types and coordinates, as well as energies, forces or both as the labels.
Alternatively, you can implement a custom class according to the torch-geometric way of implementing a dataset. That is,
derive the Dataset or InMemoryDataset class and implement the necessary functions (more info here). The dataset must return torch-geometric Data
objects, containing at least the keys z (atom types) and pos (atomic coordinates), as well as y (label), neg_dy (negative derivative of the label w.r.t atom coordinates) or both.
In addition to implementing a custom dataset class, it is also possible to add a custom prior model to the model. This can be
done by implementing a new prior model class in torchmdnet.priors and adding the argument --prior-model <PriorModelName>.
As an example, have a look at torchmdnet.priors.Atomref.
In order to train models on multiple nodes some environment variables have to be set, which provide all necessary information to PyTorch Lightning. In the following we provide an example bash script to start training on two machines with two GPUs each. The script has to be started once on each node. Once torchmd-train is started on all nodes, a network connection between the nodes will be established using NCCL.
In addition to the environment variables the argument --num-nodes has to be specified with the number of nodes involved during training.
export NODE_RANK=0
export MASTER_ADDR=hostname1
export MASTER_PORT=12910
mkdir -p output
CUDA_VISIBLE_DEVICES=0,1 torchmd-train --conf torchmd-net/examples/ET-QM9.yaml.yaml --num-nodes 2 --log-dir output/NODE_RANK: Integer indicating the node index. Must be0for the main node and incremented by one for each additional node.MASTER_ADDR: Hostname or IP address of the main node. The same for all involved nodes.MASTER_PORT: A free network port for communication between nodes. PyTorch Lightning suggests port12910as a default.
- Due to the way PyTorch Lightning calculates the number of required DDP processes, all nodes must use the same number of GPUs. Otherwise training will not start or crash.
- We observe a 50x decrease in performance when mixing nodes with different GPU architectures (tested with RTX 2080 Ti and RTX 3090).
- Some CUDA systems might hang during a multi-GPU parallel training. Try
export NCCL_P2P_DISABLE=1, which disables direct peer to peer GPU communication.
If you use TorchMD-NET in your research, please cite the following papers:
@misc{pelaez2024torchmdnet,
title={TorchMD-Net 2.0: Fast Neural Network Potentials for Molecular Simulations},
author={Raul P. Pelaez and Guillem Simeon and Raimondas Galvelis and Antonio Mirarchi and Peter Eastman and Stefan Doerr and Philipp Thölke and Thomas E. Markland and Gianni De Fabritiis},
year={2024},
eprint={2402.17660},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
@inproceedings{simeon2023tensornet,
title={TensorNet: Cartesian Tensor Representations for Efficient Learning of Molecular Potentials},
author={Guillem Simeon and Gianni De Fabritiis},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=BEHlPdBZ2e}
}
@inproceedings{
tholke2021equivariant,
title={Equivariant Transformers for Neural Network based Molecular Potentials},
author={Philipp Th{\"o}lke and Gianni De Fabritiis},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=zNHzqZ9wrRB}
}
@article{Majewski2023,
title = {Machine learning coarse-grained potentials of protein thermodynamics},
volume = {14},
ISSN = {2041-1723},
url = {http://dx.doi.org/10.1038/s41467-023-41343-1},
DOI = {10.1038/s41467-023-41343-1},
number = {1},
journal = {Nature Communications},
publisher = {Springer Science and Business Media LLC},
author = {Majewski, Maciej and Pérez, Adrià and Th\"{o}lke, Philipp and Doerr, Stefan and Charron, Nicholas E. and Giorgino, Toni and Husic, Brooke E. and Clementi, Cecilia and Noé, Frank and De Fabritiis, Gianni},
year = {2023},
month = sep
}
To implement a new architecture, you need to follow these steps:
1. Create a new class in torchmdnet.models that inherits from torch.nn.Model. Follow TorchMD_ET as a template. This is a minimum implementation of a model:
class MyModule(nn.Module):
def __init__(self, parameter1, parameter2):
super(MyModule, self).__init__()
# Define your model here
self.layer1 = nn.Linear(10, 10)
...
# Initialize your model parameters here
self.reset_parameters()
def reset_parameters(self):
# Initialize your model parameters here
nn.init.xavier_uniform_(self.layer1.weight)
...
def forward(self,
z: Tensor, # Atomic numbers, shape (n_atoms, 1)
pos: Tensor, # Atomic positions, shape (n_atoms, 3)
batch: Tensor, # Batch vector, shape (n_atoms, 1). All atoms in the same molecule have the same value and are contiguous.
q: Optional[Tensor] = None, # Atomic charges, shape (n_atoms, 1)
s: Optional[Tensor] = None, # Atomic spins, shape (n_atoms, 1)
) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]:
# Define your forward pass here
scalar_features = ...
vector_features = ...
# Return the scalar and vector features, as well as the atomic numbers, positions and batch vector
return scalar_features, vector_features, z, pos, batch2. Add the model to the __all__ list in torchmdnet.models.__init__.py. This will make the tests pick your model up.
3. Tell models.model.create_model how to initialize your module by adding a new entry, for instance:
elif args["model"] == "mymodule":
from torchmdnet.models.torchmd_mymodule import MyModule
is_equivariant = False # Set to True if your model is equivariant
representation_model = MyModule(
parameter1=args["parameter1"],
parameter2=args["parameter2"],
**shared_args, # Arguments typically shared by all models
)4. Add any new parameters required to initialize your module to scripts.train.get_args. For instance:
parser.add_argument('--parameter1', type=int, default=32, help='Parameter1 required by MyModule')
...5. Add an example configuration file to torchmd-net/examples that uses your model.
6. Make tests use your configuration file by adding a case to tests.utils.load_example_args. For instance:
if model_name == "mymodule":
config_file = join(dirname(dirname(__file__)), "examples", "MyModule-QM9.yaml")At this point, if your module is missing some feature the tests will let you know, and you can add it. If you add a new feature to the package, please add a test for it.
We use black. Please run black on your modified files before committing.
To run the tests, install the package and run pytest in the root directory of the repository. Tests are a good source of knowledge on how to use the different components of the package.