torchmd-net
torchmd-net copied to clipboard
torchmd
Neural network potentials
My training runs always crash after exactly 50 epochs. Looking at the log, there are many repetitions of this error: ``` Exception in thread Thread-104 (_pin_memory_loop): Traceback (most recent call...
This PR adds overloads for float16 and bfloat16 input/output of the neighbor extension. Makes it possible to train in float16 or bfloat16 without casting in between.
DO NOT MERGE YET! In PyTorch Lightning, the LightningModule and LightningDataModule are two separate components with distinct responsibilities. Occasionally, conflicts in hyperparameters may arise between the model and the data...
The OP in this issue https://github.com/openmm/openmm-torch/issues/133 was led to believe the Graph Network is equivariant because TorchMD-Net provides the "equivariant_group" parameter (which is only used by TensorNet). This happens a...
Right now the TorchMD_Net module is tied to the idea that only one thing (and maybe its derivative) is returned: https://github.com/torchmd/torchmd-net/blob/a11684710d6ba22b29acbf709bfbc05b93f3402d/torchmdnet/models/model.py#L300 With TensorNet it might be interesting to have more...
At some point, the metrics.csv file seems to have changed to putting the columns in a random order. Literally. As in, every training run puts them in a different order!...
This PR changes the LNNP module so that the real time since the start is logged each epoch. Allows to track training time with the CSVLogger.
Current implementation of the cell list will fail if the provided box vectors do not represent a rectangular domain: https://github.com/torchmd/torchmd-net/blob/e20876ff79fe952db25e5fc55cc051a919ef218f/torchmdnet/neighbors/neighbors_cuda_cell.cuh#L351-L354 Supporting triclinic requires modifying the cell traversal code: https://github.com/torchmd/torchmd-net/blob/e20876ff79fe952db25e5fc55cc051a919ef218f/torchmdnet/neighbors/neighbors_cuda_cell.cuh#L331-L332 similarly...
Hi, I'm new using all the "torch" things and I wonder if it is possible to somehow use the Metal Performance Shaders (MPS) for Accelerated PyTorch training on Mac with...
