GATv2Conv

class dgl.nn.pytorch.conv.GATv2Conv(in_feats, out_feats, num_heads, feat_drop=0.0, attn_drop=0.0, negative_slope=0.2, residual=False, activation=None, allow_zero_in_degree=False, bias=True, share_weights=False)[source]

Bases: Module

GATv2 from How Attentive are Graph Attention Networks?

\[h_i^{(l+1)} = \sum_{j\in \mathcal{N}(i)} \alpha_{ij}^{(l)} W^{(l)}_{right} h_j^{(l)}\]

where \(\alpha_{ij}\) is the attention score bewteen node \(i\) and node \(j\):

\[ \begin{align}\begin{aligned}\alpha_{ij}^{(l)} &= \mathrm{softmax_i} (e_{ij}^{(l)})\\e_{ij}^{(l)} &= {\vec{a}^T}^{(l)}\mathrm{LeakyReLU}\left( W^{(l)}_{left} h_{i} + W^{(l)}_{right} h_{j}\right)\end{aligned}\end{align} \]
Parameters:
  • in_feats (int, or pair of ints) – Input feature size; i.e, the number of dimensions of \(h_i^{(l)}\). If the layer is to be applied to a unidirectional bipartite graph, in_feats specifies the input feature size on both the source and destination nodes. If a scalar is given, the source and destination node feature size would take the same value.

  • out_feats (int) – Output feature size; i.e, the number of dimensions of \(h_i^{(l+1)}\).

  • num_heads (int) – Number of heads in Multi-Head Attention.

  • feat_drop (float, optional) – Dropout rate on feature. Defaults: 0.

  • attn_drop (float, optional) – Dropout rate on attention weight. Defaults: 0.

  • negative_slope (float, optional) – LeakyReLU angle of negative slope. Defaults: 0.2.

  • residual (bool, optional) – If True, use residual connection. Defaults: False.

  • activation (callable activation function/layer or None, optional.) – If not None, applies an activation function to the updated node features. Default: None.

  • allow_zero_in_degree (bool, optional) – If there are 0-in-degree nodes in the graph, output for those nodes will be invalid since no message will be passed to those nodes. This is harmful for some applications causing silent performance regression. This module will raise a DGLError if it detects 0-in-degree nodes in input graph. By setting True, it will suppress the check and let the users handle it by themselves. Defaults: False.

  • bias (bool, optional) – If set to False, the layer will not learn an additive bias. (default: True)

  • share_weights (bool, optional) – If set to True, the same matrix for \(W_{left}\) and \(W_{right}\) in the above equations, will be applied to the source and the target node of every edge. (default: False)

Note

Zero in-degree nodes will lead to invalid output value. This is because no message will be passed to those nodes, the aggregation function will be applied on empty input. A common practice to avoid this is to add a self-loop for each node in the graph if it is homogeneous, which can be achieved by:

>>> g = ... # a DGLGraph
>>> g = dgl.add_self_loop(g)

Calling add_self_loop will not work for some graphs, for example, heterogeneous graph since the edge type can not be decided for self_loop edges. Set allow_zero_in_degree to True for those cases to unblock the code and handle zero-in-degree nodes manually. A common practise to handle this is to filter out the nodes with zero-in-degree when use after conv.

Examples

>>> import dgl
>>> import numpy as np
>>> import torch as th
>>> from dgl.nn import GATv2Conv
>>> # Case 1: Homogeneous graph
>>> g = dgl.graph(([0,1,2,3,2,5], [1,2,3,4,0,3]))
>>> g = dgl.add_self_loop(g)
>>> feat = th.ones(6, 10)
>>> gatv2conv = GATv2Conv(10, 2, num_heads=3)
>>> res = gatv2conv(g, feat)
>>> res
tensor([[[ 1.9599,  1.0239],
        [ 3.2015, -0.5512],
        [ 2.3700, -2.2182]],
        [[ 1.9599,  1.0239],
        [ 3.2015, -0.5512],
        [ 2.3700, -2.2182]],
        [[ 1.9599,  1.0239],
        [ 3.2015, -0.5512],
        [ 2.3700, -2.2182]],
        [[ 1.9599,  1.0239],
        [ 3.2015, -0.5512],
        [ 2.3700, -2.2182]],
        [[ 1.9599,  1.0239],
        [ 3.2015, -0.5512],
        [ 2.3700, -2.2182]],
        [[ 1.9599,  1.0239],
        [ 3.2015, -0.5512],
        [ 2.3700, -2.2182]]], grad_fn=<GSpMMBackward>)
>>> # Case 2: Unidirectional bipartite graph
>>> u = [0, 1, 0, 0, 1]
>>> v = [0, 1, 2, 3, 2]
>>> g = dgl.heterograph({('A', 'r', 'B'): (u, v)})
>>> u_feat = th.tensor(np.random.rand(2, 5).astype(np.float32))
>>> v_feat = th.tensor(np.random.rand(4, 10).astype(np.float32))
>>> gatv2conv = GATv2Conv((5,10), 2, 3)
>>> res = gatv2conv(g, (u_feat, v_feat))
>>> res
tensor([[[-0.0935, -0.4273],
        [-1.1850,  0.1123],
        [-0.2002,  0.1155]],
        [[ 0.1908, -1.2095],
        [-0.0129,  0.6408],
        [-0.8135,  0.1157]],
        [[ 0.0596, -0.8487],
        [-0.5421,  0.4022],
        [-0.4805,  0.1156]],
        [[-0.0935, -0.4273],
        [-1.1850,  0.1123],
        [-0.2002,  0.1155]]], grad_fn=<GSpMMBackward>)
forward(graph, feat, get_attention=False)[source]

Description

Compute graph attention network layer.

param graph:

The graph.

type graph:

DGLGraph

param feat:

If a torch.Tensor is given, the input feature of shape \((N, D_{in})\) where \(D_{in}\) is size of input feature, \(N\) is the number of nodes. If a pair of torch.Tensor is given, the pair must contain two tensors of shape \((N_{in}, D_{in_{src}})\) and \((N_{out}, D_{in_{dst}})\).

type feat:

torch.Tensor or pair of torch.Tensor

param get_attention:

Whether to return the attention values. Default to False.

type get_attention:

bool, optional

returns:
  • torch.Tensor – The output feature of shape \((N, H, D_{out})\) where \(H\) is the number of heads, and \(D_{out}\) is size of output feature.

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

raises DGLError:

If there are 0-in-degree nodes in the input graph, it will raise DGLError since no message will be passed to those nodes. This will cause invalid output. The error can be ignored by setting allow_zero_in_degree parameter to True.

reset_parameters()[source]

Description

Reinitialize learnable parameters.

Note

The fc weights \(W^{(l)}\) are initialized using Glorot uniform initialization. The attention weights are using xavier initialization method.