Method forward

python/dgl/nn/pytorch/utils.py:492–556

r"""Compute the label propagation process. Parameters ---------- g : DGLGraph The input graph. labels : torch.Tensor The input node labels. There are three cases supported. * A LongTensor of shape :math:`(N, 1)` or :math:`(N,)` fo

(self, g, labels, mask=None)

        self.reset = reset

    def forward(self, g, labels, mask=None):
        r"""Compute the label propagation process.

        Parameters
        ----------
        g : DGLGraph
            The input graph.
        labels : torch.Tensor
            The input node labels. There are three cases supported.

            * A LongTensor of shape :math:`(N, 1)` or :math:`(N,)` for node class labels in
              multiclass classification, where :math:`N` is the number of nodes.
            * A LongTensor of shape :math:`(N, C)` for one-hot encoding of node class labels
              in multiclass classification, where :math:`C` is the number of classes.
            * A LongTensor of shape :math:`(N, L)` for node labels in multilabel binary
              classification, where :math:`L` is the number of labels.
        mask : torch.Tensor
            The bool indicators of shape :math:`(N,)` with True denoting labeled nodes.
            Default: None, indicating all nodes are labeled.

        Returns
        -------
        torch.Tensor
            The propagated node labels of shape :math:`(N, D)` with float type, where :math:`D`
            is the number of classes or labels.
        """
        with g.local_scope():
            # multi-label / multi-class
            if len(labels.size()) > 1 and labels.size(1) > 1:
                labels = labels.to(th.float32)
            # single-label multi-class
            else:
                labels = F.one_hot(labels.view(-1)).to(th.float32)

            y = labels
            if mask is not None:
                y = th.zeros_like(labels)
                y[mask] = labels[mask]

            init = (1 - self.alpha) * y
            in_degs = g.in_degrees().float().clamp(min=1)
            out_degs = g.out_degrees().float().clamp(min=1)
            if self.norm_type == "sym":
                norm_i = th.pow(in_degs, -0.5).to(labels.device).unsqueeze(1)
                norm_j = th.pow(out_degs, -0.5).to(labels.device).unsqueeze(1)
            elif self.norm_type == "row":
                norm_i = th.pow(in_degs, -1.0).to(labels.device).unsqueeze(1)
            else:
                raise ValueError(
                    f"Expect norm_type to be 'sym' or 'row', got {self.norm_type}"
                )

            for _ in range(self.k):
                g.ndata["h"] = y * norm_j if self.norm_type == "sym" else y
                g.update_all(fn.copy_u("h", "m"), fn.sum("m", "h"))
                y = init + self.alpha * g.ndata["h"] * norm_i

                if self.clamp: