I learned ResGatedGCN: Residual Gated Graph ConvNets . But i have a miss understand some concept. What is src and dst in message_func . Example in below code what is different g.ndata[‘Dh’] = self.D(h) and edges.src[‘Dh’] . Thank so much
class GatedGCNLayer(nn.Module):
“”"
Param: []
“”"
def init(self, input_dim, output_dim, dropout, graph_norm, batch_norm, residual=False):
super().init()
self.in_channels = input_dim
self.out_channels = output_dim
self.dropout = dropout
self.graph_norm = graph_norm
self.batch_norm = batch_norm
self.residual = residual
if input_dim != output_dim:
self.residual = False
self.A = nn.Linear(input_dim, output_dim, bias=True)
self.B = nn.Linear(input_dim, output_dim, bias=True)
self.C = nn.Linear(input_dim, output_dim, bias=True)
self.D = nn.Linear(input_dim, output_dim, bias=True)
self.E = nn.Linear(input_dim, output_dim, bias=True)
self.bn_node_h = GraphNorm(output_dim)
self.bn_node_e = GraphNorm(output_dim)
def message_func(self, edges):
Bh_j = edges.src['Bh']
e_ij = edges.data['Ce'] + edges.src['Dh'] + edges.dst['Eh'] # e_ij = Ce_ij + Dhi + Ehj
edges.data['e'] = e_ij
return {'Bh_j' : Bh_j, 'e_ij' : e_ij}
def reduce_func(self, nodes):
Ah_i = nodes.data['Ah']
Bh_j = nodes.mailbox['Bh_j']
e = nodes.mailbox['e_ij']
sigma_ij = torch.sigmoid(e) # sigma_ij = sigmoid(e_ij)
#h = Ah_i + torch.mean( sigma_ij * Bh_j, dim=1 ) # hi = Ahi + mean_j alpha_ij * Bhj
h = Ah_i + torch.sum(sigma_ij * Bh_j, dim=1 ) / ( torch.sum( sigma_ij, dim=1 ) + 1e-6 ) # hi = Ahi + sum_j eta_ij/sum_j' eta_ij' * Bhj <= dense attention
return {'h' : h}
def forward(self, g, h, e, snorm_n, snorm_e, graph_node_size, graph_edge_size):
h_in = h # for residual connection
e_in = e # for residual connection
# g.ndata['h'] = h
g.ndata['h'] = h
g.ndata['Ah'] = self.A(h)
g.ndata['Bh'] = self.B(h)
g.ndata['Dh'] = self.D(h)
g.ndata['Eh'] = self.E(h)
g.edata['e'] = e
g.edata['Ce'] = self.C(e)
g.update_all(self.message_func,self.reduce_func)
h = g.ndata['h'] # result of graph convolution
e = g.edata['e'] # result of graph convolution
if self.graph_norm:
h = h * snorm_n # normalize activation w.r.t. graph size
e = e * snorm_e # normalize activation w.r.t. graph size
if self.batch_norm:
h = self.bn_node_h(h, graph_node_size) # graph normalization
e = self.bn_node_e(e, graph_edge_size) # graph normalization
h = F.relu(h) # non-linear activation
e = F.relu(e) # non-linear activation
if self.residual:
h = h_in + h # residual connection
e = e_in + e # residual connection
h = F.dropout(h, self.dropout, training=self.training)
e = F.dropout(e, self.dropout, training=self.training)
return h, e