EGATConv

class dgl.nn.pytorch.conv.EGATConv(in_node_feats, in_edge_feats, out_node_feats, out_edge_feats, num_heads, bias=True)[source]

Bases: torch.nn.modules.module.Module

Graph attention layer that handles edge features from Rossmann-Toolbox (see supplementary data)

The difference lies in how unnormalized attention scores \(e_{ij}\) are obtained:

\[ \begin{align}\begin{aligned}e_{ij} &= \vec{F} (f_{ij}^{\prime})\\f_{ij}^{\prime} &= \mathrm{LeakyReLU}\left(A [ h_{i} \| f_{ij} \| h_{j}]\right)\end{aligned}\end{align} \]

where \(f_{ij}^{\prime}\) are edge features, \(\mathrm{A}\) is weight matrix and

Math

vec{F} is weight vector. After that, resulting node features

\(h_{i}^{\prime}\) are updated in the same way as in regular GAT.

Parameters

  • in_node_feats (int) – Input node feature size \(h_{i}\).

  • in_edge_feats (int) – Input edge feature size \(f_{ij}\).

  • out_node_feats (int) – Output node feature size.

  • out_edge_feats (int) – Output edge feature size \(f_{ij}^{\prime}\).

  • num_heads (int) – Number of attention heads.

  • bias (bool, optional) – If True, add bias term to :math: f_{ij}^{prime}. Defaults: True.

Examples

>>> import dgl
>>> import torch as th
>>> from dgl.nn import EGATConv

>>> num_nodes, num_edges = 8, 30
>>> # generate a graph
>>> graph = dgl.rand_graph(num_nodes,num_edges)

>>> node_feats = th.rand((num_nodes, 20))
>>> edge_feats = th.rand((num_edges, 12))
>>> egat = EGATConv(in_node_feats=20,
...                 in_edge_feats=12,
...                 out_node_feats=15,
...                 out_edge_feats=10,
...                 num_heads=3)
>>> #forward pass
>>> new_node_feats, new_edge_feats = egat(graph, node_feats, edge_feats)
>>> new_node_feats.shape, new_edge_feats.shape
torch.Size([8, 3, 15]) torch.Size([30, 3, 10])
forward(graph, nfeats, efeats, get_attention=False)[source]

Compute new node and edge features.

Parameters

  • graph (DGLGraph) – The graph.

  • nfeats (torch.Tensor) –

    The input node feature of shape \((N, D_{in})\) where:

    \(D_{in}\) is size of input node feature, \(N\) is the number of nodes.

  • efeats (torch.Tensor) –

    The input edge feature of shape \((E, F_{in})\) where:

    \(F_{in}\) is size of input node feature, \(E\) is the number of edges.

  • get_attention (bool, optional) – Whether to return the attention values. Default to False.

Returns

  • pair of torch.Tensor – node output features followed by edge output features The node output feature of shape \((N, H, D_{out})\) The edge output feature of shape \((F, H, F_{out})\) where:

    \(H\) is the number of heads, \(D_{out}\) is size of output node feature, \(F_{out}\) is size of output edge feature.

  • torch.Tensor, optional – The attention values of shape \((E, H, 1)\). This is returned only when :attr: get_attention is True.

reset_parameters()[source]

Reinitialize learnable parameters.