def selfTrain(graphsage, labels, train, test, val, confList, adj_lists, filetime, features, temperature = 1, update_interval = 50, maxiter = 500, tol = 0.5, batch_size = 200 ):
opt = TrainOptions().parse()
#index = 0
#index_array = np.arange(test.shape[0])
closs = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(filter(lambda p : p.requires_grad, graphsage.parameters()), lr= opt.lr_slf, momentum = opt.momentum_slf )
loss_Data = []
y_pred_last = 0
encoder_scheduler = StepLR(optimizer,step_size=opt.step_size ,gamma= opt.gamma)
acNodes = []
nonActNod = []
for ite in range(int(maxiter)):
if ite % update_interval == 0:
labels_pse = np.empty((len(labels), 1), dtype=np.int64)
labels_pse[train] = labels[train]
test_scores = graphsage(test)
test_scores = test_scores #/ temperature.unsqueeze(1).expand(test_scores.size(0), test_scores.size(1))
test_output, _ = torch.max(F.softmax(test_scores,dim = 1), dim = 1)
y_pred = F.softmax(test_scores, dim = 1).data.numpy().argmax(axis=1)
#conf_test_nod, unconf_test_nod = vat_sel(graphsage, test, test_scores)
conf_test_nod = list((test_output > opt.thres).nonzero().data.numpy().squeeze())
unconf_test_nod = list((test_output <= opt.thres).nonzero().data.numpy().squeeze())
#print(len(list((test_output <= 0.4).nonzero().data.numpy().squeeze())))
labels_pse[ test[conf_test_nod]] = calPseLb(F.softmax(test_scores, dim = 1))[conf_test_nod]
if ite % (update_interval*2) == 0:
node, nonActN = activeUnL(graphsage, list(val), Variable(torch.LongTensor(labels[np.array(list(val))])).squeeze().data.numpy(), features, adj_lists, opt.num_features, opt.num_hidden, opt.num_cls, filetime, labels)
acNodes.extend(node)
nonActNod.extend(nonActN)
train_set = train +list(acNodes)
labels_pse[ acNodes ] = labels[ acNodes ]
#print(len(conf_test_nod), len(actNod))
#writetofile(len(actNod), opt.res_path, filetime)
#plotGraphStrc(train + acNodes + nonActNod,graphsage(train + acNodes + nonActNod).data, [0]*len(train)+[1]*len(acNodes)+[2]*len(nonActNod), adj_lists, time = filetime, name = 'act-slfT'+str(ite // update_interval))
else:
train_set = train + list(test[conf_test_nod])+list(acNodes)
labels_pse[ acNodes ] = labels[ acNodes ]
writetofile(len(conf_test_nod), opt.res_path, filetime)
plotGraphStrc(train + list(test[conf_test_nod]) + list(test[unconf_test_nod]),graphsage(train + list(test[conf_test_nod]) + list(test[unconf_test_nod])).data, [0]*len(train)+[1]*len(conf_test_nod)+[2]*len(unconf_test_nod), adj_lists, time = filetime, name = 'slfT'+str(ite // update_interval))
#plotGraphStrc(train + list(test) ,graphsage(train + list(test)).data, [0]*len(train)+ list(test_output) , adj_lists, time = filetime, name = 'slfT'+str(ite // update_interval))
print('\nIter {}: '.format(ite), end='')
writetofile('Iter:'+str(ite), opt.res_path, filetime)
print("Validation ACCU:", accuracy_score( labels[test], y_pred))
writetofile("Validation ACCU:" + str(accuracy_score( labels[test], y_pred)), opt.res_path, filetime)
print("confidence point ACCU:", accuracy_score( labels[test[conf_test_nod]], y_pred[conf_test_nod]))
writetofile("confidence point ACCU:" + str(accuracy_score( labels[test[conf_test_nod]], y_pred[conf_test_nod])), opt.res_path, filetime)
# check stop criterion
if ite >0:
delta_label = np.sum(y_pred != y_pred_last).astype(np.float) / y_pred.shape[0]
print('Fraction of documents with label changes: {} %'.format(np.round(delta_label*100, 3)))
writetofile("Fraction of documents with label changes: " + str(np.round(delta_label*100, 3)), opt.res_path, filetime)
y_pred_last = np.copy(y_pred)
if ite > 0 and delta_label <= tol/100:
print('\nFraction: {} % < tol: {} %'.format(np.round(delta_label*100, 3), tol))
print('Reached tolerance threshold. Stopping training.')
writetofile('Reached tolerance threshold. Stopping training.', opt.res_path, filetime)
break
scores = graphsage(train_set)
#scores = scores #/ temperature
#v_loss = vat_loss(graphsage, train_set,scores)
loss = closs(scores, Variable(torch.LongTensor(labels_pse[np.array(train_set)])).squeeze()) #+ activeL(graphsage, train_set, Variable(torch.LongTensor(labels_pse[np.array(train_set)])).squeeze(), features, adj_lists, opt.num_features, opt.num_hidden, opt.num_cls)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
encoder_scheduler.step(ite)
#index = index + 1 if (index + 1) * batch_size <= test.shape[0] else 0
loss_Data.append(loss.data)
return graphsage, loss_Data
loss = criterion(pre, batch_x4)
psnr = batch_PSNR(torch.clamp(pre, 0., 1.), batch_x4, 1.0)
loss_list.append(loss.item())
bar.set_description("Epoch: %d Loss: %.6f" % (ep, loss_list[-1]))
# Update the parameters
optimizer.zero_grad()
loss.backward()
optimizer.step()
print("[Epoch %d][G loss: %7f][PSNR : %7f]" %
(ep, loss_list[-1], psnr))
Loss_list.append(np.mean(loss_list))
# Save the training image
'''
if ep % opts.record_epoch == 0:
img = fusePostProcess(y_f, y_hat, patch1, patch2, single = False)
cv2.imwrite(os.path.join(opts.det, 'image', str(ep) + ".png"), img[0, :, :, :])
'''
# Save the training model
if (ep % 100 == 0):
torch.save(model.state_dict(), './Model/%d.pkl' % (ep))
if __name__ == '__main__':
opts = TrainOptions().parse()
train(opts)
#def test(opts):
#v_loss = vat_loss(graphsage, train_set,scores)
loss = closs(scores, Variable(torch.LongTensor(labels_pse[np.array(train_set)])).squeeze()) #+ activeL(graphsage, train_set, Variable(torch.LongTensor(labels_pse[np.array(train_set)])).squeeze(), features, adj_lists, opt.num_features, opt.num_hidden, opt.num_cls)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
encoder_scheduler.step(ite)
#index = index + 1 if (index + 1) * batch_size <= test.shape[0] else 0
loss_Data.append(loss.data)
return graphsage, loss_Data
if __name__ == "__main__":
if torch.cuda.is_available():
device = 'cuda'
else:
device = 'cpu'
opt = TrainOptions().parse()
if opt.dataset == 'cora':
opt.k = 80
opt.lr_pre = 1e-4
opt.epoch = 400
elif opt.dataset =='pubmed':
opt.k = 50
opt.lr_pre = 5e-4
elif opt.dataset == 'ppi':
opt.lr_pre = 5e-4
opt.k = 500
opt.num_hidden = 150
opt.epoch = 1000
elif opt.dataset == 'reddit':
opt.lr_pre = 5e-4
opt.k = 500
