17aafdea24122bbb777f71a8ea7c2505e3fee84c,torch_geometric/nn/models/autoencoder.py,GAE,reconstruction_loss,#GAE#Any#Any#Any#,71

Before Change

        y = torch.cat([pos_y, neg_y], dim=0)
        pred = torch.cat([adj[row, col], adj[neg_adj_mask]], dim=0)

        pos_weight = torch.tensor([neg_y.size(0) / pos_y.size(0)])
        pos_weight = pos_weight.to(adj.device)
        self.loss_op = torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight)

        return self.loss_op(pred, y)

After Change

    def reconstruction_loss(self, adj, edge_index, neg_adj_mask):
        row, col = edge_index
        loss = -torch.log(torch.sigmoid(adj[row, col])).mean()
        loss = loss - torch.log(1 - torch.sigmoid(adj[neg_adj_mask])).mean()
        return loss

    def eval(self, adj, edge_index, neg_edge_index):
        pos_y = adj.new_ones(edge_index.size(1))

Instances