ModelTemplate (Python)

The ModelTemplate Cython class wraps the C++ model_template. User GNN model classes must subclass both ModelTemplate and torch.nn.Module, then override forward(data: graph_t*).

Properties

Property	Type	Description
o_graph	dict[str, str]	Output graph-level feature map (feature name → output key).
o_node	dict[str, str]	Output node-level feature map.
o_edge	dict[str, str]	Output edge-level feature map.
i_graph	list[str]	Input graph-level feature names expected by the model.
i_node	list[str]	Input node-level feature names expected by the model.
i_edge	list[str]	Input edge-level feature names expected by the model.
device	str	Target compute device (e.g. "cpu", "cuda:0").
checkpoint_path	str	Directory where model checkpoints are written/read.
weight_name	str	File stem used when saving model weights.
tree_name	str	ROOT tree name used when saving predictions.
name	str	Model name string.

C++ Interface (called from forward)

These methods are available on the graph_t* pointer passed to forward:

Method	Description
data.get_data_graph(name: bytes, model: self) → torch.Tensor	Retrieve a graph-level data tensor by name.
data.get_data_node(name: bytes, model: self) → torch.Tensor	Retrieve a node-level data tensor by name.
data.get_data_edge(name: bytes, model: self) → torch.Tensor	Retrieve an edge-level data tensor by name.
data.get_truth_graph(name: bytes, model: self) → torch.Tensor	Retrieve a graph-level truth tensor by name.
data.get_truth_node(name: bytes, model: self) → torch.Tensor	Retrieve a node-level truth tensor by name.
data.get_truth_edge(name: bytes, model: self) → torch.Tensor	Retrieve an edge-level truth tensor by name.
data.get_edge_index(model: self) → torch.Tensor	Retrieve the [2, E] edge-index tensor.
prediction_graph_feature(name: bytes, tensor)	Store a graph-level prediction tensor.
prediction_node_feature(name: bytes, tensor)	Store a node-level prediction tensor.
prediction_edge_feature(name: bytes, tensor)	Store an edge-level prediction tensor.