"""
construct_DGLgraph
"""
import collections
import dgl
import numpy as np
import torch
from sklearn.metrics import pairwise_distances as pair
[docs]def construct_DGLgraph_for_non_graph(x, labels, k=3, method="euclidean"):
if method == "heat":
return construct_DGLgraph_for_non_graph_by_heat(x, labels, k)
if torch.is_tensor(x) is False:
x = torch.Tensor(x)
if torch.is_tensor(labels) is False:
labels = torch.Tensor(labels)
knn_g = dgl.knn_graph(x, k, dist=method)
edges = knn_g.edges()
edges = np.array(process_edges_info(edges))
return build_graph(x, edges, x.shape[0], labels)
[docs]def construct_DGLgraph_for_non_graph_by_heat(x, labels, k=3):
if torch.is_tensor(x) is False:
x = torch.Tensor(x)
if torch.is_tensor(labels) is False:
labels = torch.Tensor(labels)
dist = None
dist = -0.5 * pair(x)**2 # 时间参数t默认为2？？？？？,计算任意两个节点之间的距离，并返回一个矩阵
dist = np.exp(dist) # 得到相似度矩阵
inds = []
for i in range(dist.shape[0]):
ind = np.argpartition(dist[i, :], -(k + 1))[-(k + 1):] # 获取k个最相似的节点的下标
inds.append(ind)
src = []
dst = []
counter = 0
for i, v in enumerate(inds):
for vv in v:
if vv == i:
pass
else:
if labels[vv] != labels[i]:
counter += 1
src.append(i)
dst.append(vv)
edges = torch.stack((torch.Tensor(src), torch.Tensor(dst)), dim=1)
edges = np.array(process_edges_info(edges))
return build_graph(x, edges, x.shape[0], labels)
[docs]def construct_DGLgraph_for_graph(x, labels, edges):
if torch.is_tensor(x) is False:
x = torch.Tensor(x)
if torch.is_tensor(labels) is False:
labels = torch.Tensor(labels)
edges = np.array(process_edges_info(edges))
return build_graph(x, edges, x.shape[0], labels)
[docs]def process_edges_info(
edges): # 由于DGL图会将重复的边也算进边的总数内，所有要去除重复的边和自环，到后面统一添加自环,并且DGL
edge_dict = collections.defaultdict(bool)
pair_list = []
u = edges[0]
v = edges[1]
length = len(edges[0])
for i in range(length):
if (edge_dict[(u[i].item(), v[i].item())] is False
and u[i].item() != v[i].item()):
edge_dict[(u[i].item(), v[i].item())] = True
edge_dict[(v[i].item(), u[i].item())] = True
pair_list.append((u[i].item(), v[i].item()))
pair_list.append((v[i].item(), u[i].item()))
return pair_list
[docs]def build_graph(features, edges, num_nodes, labels): # 构图，减少代码重复
graph = dgl.graph((edges[0], edges[1]), num_nodes=num_nodes)
graph.ndata["feat"] = features
graph.ndata["label"] = labels
graph = dgl.add_self_loop(graph)
return graph