weights_network.py

class GraphWeightsNetwork(nn.Module):
    def __init__ (self, continuous_dim, cat_dim, output_dim=1, hidden_dim=32, conv_depth=1):
        super(GraphMETNetwork, self).__init__()
        
        self.embed_charge = nn.Embedding(3, hidden_dim//4)
        self.embed_pdgid = nn.Embedding(7, hidden_dim//4)
        self.embed_pv = nn.Embedding(8, hidden_dim//4)
        
        self.embed_continuous = nn.Sequential(nn.Linear(continuous_dim,hidden_dim//2),
                                              nn.ELU(),                                             
                                              # nn.BatchNorm1d(hidden_dim) # uncomment if it starts overtraining
                                             )

        self.embed_categorical = nn.Sequential(nn.Linear(3*hidden_dim//4,hidden_dim//2),
                                               nn.ELU(),                                               
                                               # nn.BatchNorm1d(hidden_dim)
                                              )

        self.conv_continuous = nn.ModuleList()        
        for i in range(conv_depth):
            mesg = nn.Sequential(nn.Linear(2*hidden_dim, hidden_dim),
                                 nn.ELU(),                                 
                                 # nn.BatchNorm1d(hidden_dim)
                                )

            self.conv_continuous.append(
                EdgeConv(nn=mesg).jittable()
                #GATConv(hidden_dim, hidden_dim).jittable()
                #GCNConv(hidden_dim, hidden_dim).jittable()
                #SGConv(hidden_dim, hidden_dim).jittable()
            )            

        self.output = nn.Sequential(nn.Linear(hidden_dim, hidden_dim//2),
                                    nn.ELU(),
                                    nn.Linear(hidden_dim//2, output_dim)
                                   )

    def forward(self, x_cont, x_cat, edge_index, batch):
        emb_cont = self.embed_continuous(x_cont)        
        emb_chrg = self.embed_charge(x_cat[:, 1] + 1)
        emb_pdg = self.embed_pdgid(x_cat[:, 0])
        emb_pv = self.embed_pv(x_cat[:, 2])
                
        emb_cat = self.embed_categorical(torch.cat([emb_chrg, emb_pdg, emb_pv], dim=1))
        
        emb = torch.cat([emb_cat, emb_cont], dim=1)
        
        # graph convolution for continuous variables
        for co_conv in self.conv_continuous:
            #emb_cont = co_conv(emb_cont, edge_index)
            emb = emb + co_conv(emb, edge_index) # residual connections on the convolutional layer
                
        out = self.output(emb)
        
        return out.squeeze(-1)