Encoders (GNN Backbones)

Three interchangeable backbones — GCN, GIN, Graph Transformer — swappable via one config line for any model. All expose the same signature:

forward(x, edge_index, batch, edge_attr=None) -> Tensor[N, out_dim]

and are registered via @ENCODERS.register("name") (see Registry Pattern).

Shared config (EncoderConfig)

Field	Type	Default	Notes
name	str	—	'gin', 'gcn', 'transformer'
hidden_dim	int	—	hidden/output dimension
num_layers	int	—	number of layers
norm_type	str	"batch"	'batch', 'layer', 'none' — uniform across all encoders
weight_standardization	bool	False	GCN only: standardize the convolution weights of every layer after the first
pool	bool	True	graph-level task: the model pools the node embeddings per graph
readout	str	"mean"	'mean', 'sum' or 'max': the pooling used when pool=True
drop	float | None	None	dropout rate; None keeps the encoder's own default (GIN 0.2, GCN 0, Transformer 0.1)
mlp_ratio	float	2.0	hidden-dim multiplier for the internal MLP
edge_dim	int | None	None	enables edge-feature-aware convolution (GINEConv / TransformerConv)
node_emb_num_classes	int | None	None	replaces Linear with nn.Embedding for categorical node features (e.g. ZINC: 28 atom types)
edge_emb_num_classes	int | None	None	nn.Embedding for categorical edge features. Requires edge_dim

EncoderConfig.build(in_channels) uses inspect.signature to forward only the kwargs the specific encoder actually accepts, so the same config schema works across encoders that don't share every feature. A field left at its default is simply not forwarded to an encoder that has no use for it. A field that you set and that the encoder does not take raises a ValueError instead of being dropped:

EncoderConfig(name="gcn", hidden_dim=64, num_layers=2, edge_dim=16).build(in_channels)
# ValueError: encoder 'gcn' does not take edge_dim: the value would be discarded. ...

The same holds for the required fields (hidden_dim, num_layers), which are always set. pool and readout are exempt (the models read them, not the encoder), and an encoder whose constructor has **kwargs receives every field. Before 0.2.0 such fields were ignored without notice — which is how a drop given to the GCN or Transformer encoder, and num_layers given to the GCN encoder, had no effect.

GCN

• num_layers × (GCNConv → norm → PReLU).
• drop applies dropout to the input of every layer, as in Kipf & Welling (2017). No dropout by default.
• Optional weight standardization (weight_standardization) of the convolution weights of every layer after the first — BGRL's setting on ogbn-arxiv, together with layer norm.
• Backward compatible with the older batchnorm=True/False / layernorm=True/False flags.
• Exposes reset_parameters() — used by BGRL to give its target encoder different initial weights (see BGRL).

GIN (Graph Isomorphism Network)

• A stack of GINLayers with pre-norm + residual connections: norm(x) → GINConv(h) → ReLU → Dropout → x + h.
• edge_dim != None switches to GINEConv (edge-feature aware). If edge_dim is set but edge_attr=None at runtime, the layer builds a correctly-shaped zero tensor instead of crashing — the same encoder works with or without edge features.
• node_emb_num_classes / edge_emb_num_classes swap in nn.Embedding for categorical node/edge features (see the ZINC example).

Graph Transformer

• A stack of TransformerBlocks (PyG's TransformerConv), pre-norm + FFN + residual.
• Supports edge_attr via TransformerConv(edge_dim=...).
• drop sets both of its dropout rates (feed-forward and attention); left unset, both are 0.1.
• Same norm_type API as GIN (batch/layer/none), and the same node_emb_num_classes / edge_emb_num_classes support.

Reference

See the Encoders API for full signatures.