Source code for torch_geometric.utils.to_dense_adj

import torch
from torch_scatter import scatter


[docs]def to_dense_adj(edge_index, batch=None, edge_attr=None, max_num_nodes=None): r"""Converts batched sparse adjacency matrices given by edge indices and edge attributes to a single dense batched adjacency matrix. Args: edge_index (LongTensor): The edge indices. batch (LongTensor, optional): Batch vector :math:`\mathbf{b} \in {\{ 0, \ldots, B-1\}}^N`, which assigns each node to a specific example. (default: :obj:`None`) edge_attr (Tensor, optional): Edge weights or multi-dimensional edge features. (default: :obj:`None`) max_num_nodes (int, optional): The size of the output node dimension. (default: :obj:`None`) :rtype: :class:`Tensor` """ if batch is None: batch = edge_index.new_zeros(edge_index.max().item() + 1) batch_size = batch.max().item() + 1 one = batch.new_ones(batch.size(0)) num_nodes = scatter(one, batch, dim=0, dim_size=batch_size, reduce='add') cum_nodes = torch.cat([batch.new_zeros(1), num_nodes.cumsum(dim=0)]) if edge_attr is None: edge_attr = torch.ones(edge_index.size(1), device=edge_index.device) if max_num_nodes is None: max_num_nodes = num_nodes.max().item() size = [batch_size, max_num_nodes, max_num_nodes] size += list(edge_attr.size())[1:] adj = torch.zeros(size, dtype=edge_attr.dtype, device=edge_index.device) idx0 = batch[edge_index[0]] idx1 = edge_index[0] - cum_nodes[batch][edge_index[0]] idx2 = edge_index[1] - cum_nodes[batch][edge_index[1]] flattened_size = batch_size * max_num_nodes * max_num_nodes adj = adj.view([flattened_size] + list(adj.size())[3:]) idx = idx0 * max_num_nodes * max_num_nodes + idx1 * max_num_nodes + idx2 scatter(edge_attr, idx, dim=0, out=adj, reduce='add') adj = adj.view(size) return adj