import torch import copy import numpy as np def to_target_index(target, x_dict): cnt = 0 x = torch.concat([x.values()], dim=0) for node_type, x in x_dict.items(): if node_type == target: target_index = cnt + torch.arange(len(x.size(0))) cnt += x.size(0) return target_index def dict_to_matrix(x_dict, edge_index_dict, return_x=False): all_x = torch.concat(list(x_dict.values()), dim=0) if return_x: return all_x # collect node types x_types = [] cnt, cum_cnt = 0, {} for idx, (node_type, x) in enumerate(x_dict.items()): cum_cnt[node_type] = cnt cnt += x.size(0) x_types.append(torch.ones(x.size(0)).long() * idx) x_types = torch.concat(x_types).to(all_x.device) # construct edge_index & edge types edge_list, edge_types = [], [] _edge_index_dict = copy.deepcopy(edge_index_dict) for idx, (edge_type, edges) in enumerate(_edge_index_dict.items()): edges[0, :] = edges[0, :] + cum_cnt[edge_type[0]] edges[1, :] = edges[1, :] + cum_cnt[edge_type[-1]] edge_list.append(edges) edge_types.append(torch.ones(edges.size(-1)).long() * idx) edge_index = torch.concat(edge_list, dim=-1) edge_types = torch.concat(edge_types).to(all_x.device) return all_x, x_types, edge_index, edge_types def matrix_to_dict(x, x_dict): values = [_x for node_type, _x in x_dict.items()] n_types = np.array([0] + [x_type_i.size(0) for x_type_i in values]) cum_n_types = n_types.cumsum() keys = list(x_dict.keys()) for i, key in enumerate(keys): x_dict[key] = x[cum_n_types[i] : cum_n_types[i + 1]] return x_dict