hippynn Concepts

Layers/Networks

The Layers and Networks subpackages contain pure pytorch code. Each torch.nn.Module returns a list of tensors

Nodes

A node is basically a pytorch module that is dressed up with some extra metadata. More specifically, it is a container whose .torch_module implements the desired function.

The metadata assists in the creation of new nodes and graph structures.

• A node has parents that specify what tensors go into the node.

• As new nodes are created, the parents of that node get linked to their children.

Nodes have an IDXType that specify the form of the data output by the node. For example: Atom specifies data on atoms in the system. Molecule specifies data on an entire system. Pair specifies data that connects pairs of atoms.

The inclusion of the index types helps determine: 1. Whether it is possible for nodes to be connected. 2. If it is not directly possible, but the two data types are compatible, how to use index_transformers to transform the data from one index type to another.

A MultiNode is a node that itself produces several tensors. This covers cases where a calculation can produce several useful tensors in tandem, or when wrapping some modules already implemented.

SingleNode and MultiNode are base classes for subclassing.

Graphs

A GraphModule is a ‘compiled’ set of nodes; a torch.nn.Module that executes the graph.

GraphModules are used in a number of places within hippynn, such as the model, the loss, the evaluator, the predictor, the ASE interface, and the LAMMPS interface objects all use GraphModules.

Experiment

An experiment is a training run.