# if nothing can be loaded, generate the initial matching results and save them
m, n = data.num_graphs, data.num_nodes
Kt = torch.tensor(data.K).reshape(-1, n * n, n * n).cuda()
ns_src = torch.ones(m * m).int().cuda() * n
ns_tgt = torch.ones(m * m).int().cuda() * n
X_continue = rrwm(Kt, n, ns_src,
ns_tgt).reshape(m * m, n,
n).transpose(1, 2).contiguous()
X = hungarian(X_continue, ns_src,
ns_tgt).reshape(m, m, n, n).cpu().detach().numpy()
with open(init_mat_path, "wb") as f:
pickle.dump(X, f)
init_mat = X
X = mgm_floyd(init_mat[:15, :15], data.K[:15, :15], 15, data.num_nodes)
online_acc[i_iter][15] = evaluate_matching_accuracy(
X, data.gt[:15, :15], 15, data.num_inlier)
# apply MGM-SPFA to get better matching results
for n_graph in range(16, data.num_graphs):
tic = time.time()
X_new = init_mat[:n_graph, :n_graph, :, :]
X_new[:-1, :-1, :, :] = X
if test_mode == "fast-spfa":
if (n_graph == 16):
pc_ma = get_pairwise_socre(X_new, n_graph)
x = 1 - pc_ma
# X, pc_ma = fast_spfa(data.K[:n_graph, :n_graph, :, :], X_new, n_graph, pc_ma)
# cluster = DBSCAN(eps =np.min(X), min_samples= 1, metric= "precomputed" )
# cluster.fit(X)
# print(cluster.labels_)
# print(X)
# if nothing can be loaded, generate the initial matching results and save them
m, n = data.num_graphs, data.num_nodes
Kt = torch.tensor(data.K).reshape(-1, n * n, n * n).cuda()
ns_src = torch.ones(m * m).int().cuda() * n
ns_tgt = torch.ones(m * m).int().cuda() * n
X_continue = rrwm(Kt, n, ns_src,
ns_tgt).reshape(m * m, n,
n).transpose(1, 2).contiguous()
X = hungarian(X_continue, ns_src,
ns_tgt).reshape(m, m, n, n).cpu().detach().numpy()
with open(init_mat_path, "wb") as f:
pickle.dump(X, f)
toc = time.time()
# evaluate the initial matching results
rrwm_acc = evaluate_matching_accuracy(X, data.gt, data.num_graphs,
data.num_inlier)
rrwm_time = toc - tic
# apply MGM-Floyd to get better matching results
tic = time.time()
mat = mgm_floyd(X, data.K, data.num_graphs, data.num_nodes)
toc = time.time()
# evaluate the final matching results
time_cost.append(toc - tic)
acc = evaluate_matching_accuracy(mat, data.gt, data.num_graphs,
data.num_inlier)
accuracy.append(acc)
print(
"iter: {}/{}, init_acc: {:.4f}, floyd_acc: {:.4f}, init_time: {:.4f}, floyd_time: {:.4f}"