print('pretrain Autoencoder model (epoch %d, total_steps %d)' %
(epoch, pre_epoch_AE))
commonZ = []
if epoch > 0:
for i in range(len(train_dataset_a)):
tempimage_a = torch.from_numpy(train_dataset_a[i]).view(1, 1, dim1)
tempimage_b = torch.from_numpy(train_dataset_b[i]).view(1, 1, dim2)
model.set_input(tempimage_a, tempimage_b)
t_200 = np.array(model.test_commonZ().data.view(n_com).tolist())
commonZ.append(t_200)
##kmeans result
estimator = KMeans(n_clusters)
estimator.fit(commonZ)
centroids = estimator.cluster_centers_
label_pred = estimator.labels_
acc = metrics.acc(label_true, label_pred)
nmi = metrics.nmi(label_true, label_pred)
ACC_all.append(acc)
NMI_all.append(nmi)
# Z_path = 'commonZ' + str(epoch)
# sio.savemat(Z_path + '.mat', {'Z': commonZ})
print(' ' * 8 + '|==> acc: %.4f, nmi: %.4f <==|' % (acc, nmi))
if acc > 0.95:
break
centroids0 = estimator.cluster_centers_
########
center0 = torch.FloatTensor(centroids0).cuda()
model.clu.weights.data = center0
#########
sio.savemat('commonZAE.mat', {'Z': commonZ})
list(train_dataset_b) + list(fb_500) + list(untrain_dataset_b))
sio.savemat('fakea2.mat', {'fa': fa_500})
sio.savemat('fakeb2.mat', {'fb': fb_500})
commonZ_step2 = []
for i in range(len(test_dataset_A2000)):
tempimage_a = torch.from_numpy(test_dataset_A2000[i]).view(1, 1, dim1)
tempimage_b = torch.from_numpy(test_dataset_B2000[i]).view(1, 1, dim2)
model.set_input(tempimage_a, tempimage_b)
t_200 = np.array(model.test_commonZ().data.view(n_com).tolist())
commonZ_step2.append(t_200)
estimator = KMeans(n_clusters)
estimator.fit(commonZ_step2)
centroids_step2 = estimator.cluster_centers_
label_pred = estimator.labels_
acc = metrics.acc(label_true_all, label_pred)
nmi = metrics.nmi(label_true_all, label_pred)
ACC_all.append(acc)
NMI_all.append(nmi)
print(' ' * 8 + '|==> acc: %.4f, nmi: %.4f <==|' % (acc, nmi))
sio.savemat('commonZ.mat', {'commonZ': commonZ_step2})
# test
#groundtruth = []
#rmse_x = []
#rmse_y = []
#fa_600 = []
#fb_600 = []
#print(len(untrain_dataset_a))
#print(len(test_dataset_a))
if epoch % 1 == 0:
print("Epoch {}/{}".format(epoch, n_epochs2))
print("Loss is:{:.4f}".format(loss.item()))
print("Loss_cca is:{:.4f}".format(loss_cca.item()))
print("loss_selfc is:{:.4f}".format(loss_selfc.item()))
print("loss_selfc12 is:{:.4f}".format(loss_selfc12.item()))
print("loss_self is:{:.4f}".format(loss_self.item()))
print("loss_d is:{:.4f}".format(loss_d.item()))
coef = model.weight - torch.diag(torch.diag(model.weight))
commonC = coef.cpu().detach().numpy()
alpha = max(0.4 - (k_num - 1) / 10 * 0.1, 0.1)
All_C = thrC(commonC, alpha)
preds, _ = post_proC(All_C, k_num)
acc = metrics.acc(label_true, preds)
nmi = metrics.nmi(label_true, preds)
pur = metrics.purity_score(label_true, preds)
ACC.append(acc)
NMI.append(nmi)
PUR.append(pur)
print(' ' * 8 + '|==> acc: %.4f, nmi: %.4f， pur: %.4f <==|' %
(acc, nmi, pur))
if acc > 0.955:
C_path = 'All_C' + str(epoch)
sio.savemat(C_path + '.mat', {'C': All_C})
sio.savemat('Pred' + str(epoch) + '.mat', {'Pred': preds})
sio.savemat('ACC' + '.mat', {'acc': ACC})
sio.savemat('NMI' + '.mat', {'nmi': NMI})