GATConv

Graph Attention Network layer (Velickovic et al., 2018). Multi-head attention mechanism over graph edges. αij=softmaxj(LeakyReLU(aT[WxiWxj]))\alpha_{ij} = \text{softmax}_j(\text{LeakyReLU}(\mathbf{a}^T [\mathbf{W}\mathbf{x}_i \| \mathbf{W}\mathbf{x}_j]))

Record fields

NameDescription
AttSrc
AttTgt
Bias
Concat
Heads
NegativeSlope
OutChannels
Weight

Instance members

NameDescription
this.forward (x, edgeIndex)