Method __call__

python/dgl/transforms/module.py:1377–1408

(self, g)

            return self.eps

    def __call__(self, g):
        # Step1: heat kernel diffusion
        # t A
        device = g.device
        eweight = self.t * g.edata.get(
            self.eweight_name, F.ones((g.num_edges(),), F.float32, device)
        )
        num_nodes = g.num_nodes()
        mat = F.zeros((num_nodes, num_nodes), F.float32, device)
        src, dst = g.edges()
        src, dst = F.astype(src, F.int64), F.astype(dst, F.int64)
        mat[dst, src] = eweight
        # t (A - I_n)
        nids = F.astype(g.nodes(), F.int64)
        mat[nids, nids] = mat[nids, nids] - self.t

        if is_bidirected(g):
            e, V = torch.linalg.eigh(mat, UPLO="U")
            diff_mat = V @ torch.diag(e.exp()) @ V.t()
        else:
            diff_mat_np = expm(mat.cpu().numpy())
            diff_mat = torch.Tensor(diff_mat_np).to(device)

        # Step2: sparsification
        num_nodes = g.num_nodes()
        eps = self.get_eps(num_nodes, diff_mat)
        dst, src = (diff_mat >= eps).nonzero(as_tuple=False).t()
        data_dict = {g.canonical_etypes[0]: (src, dst)}
        new_g = update_graph_structure(g, data_dict, copy_edata=False)
        new_g.edata[self.eweight_name] = diff_mat[dst, src]

        return new_g


class GDC(BaseTransform):