TorchScript API reference

We provide a PyTorch C++ extension to make rascaline compatible with Torch and TorchScript in three ways:

  • registering rascaline calculators as special nodes in Torch’s computational graph, allowing to use backward propagation of derivatives to compute gradients of arbitrary quantities with respect to atomic positions and cell (e.g. forces and stress when the quantity is the energy of a system);

  • saving and loading calculators inside a torch Model (the calculators are exposed as special case of torch.nn.Module)

  • exporting a model trained in Python and loading it back without needing the Python interpreter, for example inside a pure C++ or Fortran molecular simulation engine.

Please refer to the installation instructions to know how to install the Python and C++ sides of this library. The core classes of rascaline are documented below for an usage from Python:

If you want to use rascaline’s TorchScript API from C++, you might be interested in the following documentation: