test_y = test_data.targets
lstm = BiLSTM(INPUT_SIZE, HIDDEN_SIZE, NUM_LAYERS, 10).to(device)
optimizer = torch.optim.Adam(lstm.parameters(), lr=LR)
loss_func = nn.CrossEntropyLoss()
for epoch in range(EPOCH):
for step, (x, y) in enumerate(train_loader):
b_x = Variable(x.view(-1, TIME_STEP, INPUT_SIZE).to(device))
b_y = Variable(y.to(device))
output = lstm(b_x)
loss = loss_func(output, b_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step%50 == 0:
test_output = lstm(test_x.view(-1, TIME_STEP, INPUT_SIZE).to(device))
pred_y = torch.max(test_output, 1)[1].data.cpu().squeeze()
accuracy = sum(pred_y == test_y)/float(test_y.size(0))
print('Epoch: ',epoch, '| train loss:%.4f' %loss.data.item(),'| test accuracy:%.2f' %accuracy)
test_output = lstm(test_x[:10].view(-1, TIME_STEP, INPUT_SIZE).to(device))
pred_y = torch.max(test_output,1)[1].data.cpu().numpy().squeeze()
print(pred_y, 'prediction number')
print(test_y[:10].numpy(), 'real number')
torch.save(lstm.state_dict(), './model/model.ckpt')
def main(config):
trainDataPath = config['data']['trainDataPath']
validDataPath = config['data']['validDataPath']
testDataPath = config['data']['testDataPath']
modelName = config['modelName']
batchSize = config['model']['batchSize']
epochNum = config['model']['epochNum']
earlyStop = config['model']['earlyStop']
learningRate = config['model']['learningRate']
modelSavePath = config['model']['modelSavePath']
#GPU/CPU
DEVICE = config['DEVICE']
trianDataset = NERDataset(trainDataPath, config)
validDataset = NERDataset(validDataPath, config)
testDataset = NERDataset(testDataPath, config)
trainIter = data.DataLoader(dataset=trianDataset,
batch_size=batchSize,
shuffle=True,
num_workers=4,
collate_fn=pad)
validIter = data.DataLoader(dataset=validDataset,
batch_size=batchSize,
shuffle=False,
num_workers=4,
collate_fn=pad)
testIter = data.DataLoader(dataset=testDataset,
batch_size=batchSize,
shuffle=False,
num_workers=4,
collate_fn=pad)
if modelName == 'bilstm':
net = BiLSTM(config)
train = bilstmTrain
eval = bilstmEval
if torch.cuda.device_count() > 1:
net = nn.DataParallel(net)
if modelName == 'bilstm_crf':
net = BiLSTM_CRF(config)
train = bilstmCRFTrain
eval = bilstmCRFEval
if modelName == 'transformer_crf':
net = Transformer_CRF(config)
train = transformerCRFTrain
eval = transformerCRFEval
if modelName == 'cnn':
net = CNN(config)
train = cnnTrain
eval = cnnEval
net = net.to(DEVICE)
lossFunction = nn.NLLLoss()
optimizer = optim.Adam(net.parameters(),
lr=learningRate,
betas=(0.9, 0.999),
eps=1e-08)
earlyNumber, beforeLoss, maxScore = 0, sys.maxsize, -1
#开始训练
for epoch in range(epochNum):
print('第%d次迭代: ' % epoch)
totalLoss = train(net,
trainIter,
optimizer=optimizer,
criterion=lossFunction,
DEVICE=DEVICE)
print('训练损失为: %f' % totalLoss)
totalLoss, f1Score = eval(net,
validIter,
criterion=lossFunction,
DEVICE=DEVICE)
if f1Score > maxScore:
maxScore = f1Score
torch.save(net.state_dict(), modelSavePath)
print('验证损失为:%f f1Score:%f / %f' % (totalLoss, f1Score, maxScore))
if f1Score < maxScore:
earlyNumber += 1
print('earyStop: %d/%d' % (earlyNumber, earlyStop))
else:
earlyNumber = 0
if earlyNumber >= earlyStop: break
print('\n')
#加载最优模型
net.load_state_dict(torch.load(modelSavePath))
totalLoss, f1Score = eval(net,
testIter,
criterion=lossFunction,
DEVICE=DEVICE)
print('测试损失为: %f, f1Score: %f' % (totalLoss, f1Score))
ranks = len(values) - temp #Descending
#print(acc_labels, entry)
groundT = acc_labels.index(1)
MRR.append(1.0 / ranks[groundT])
test_mrr = sum(MRR) / len(MRR)
#print("mrr at epoch", epoch, test_mrr)
if best_MRR < test_mrr:
best_MRR = test_mrr
#import pdb; pdb.set_trace()
acc = get_accuracy(truth_res, pred_res)
if best_accuracy < acc:
best_accuracy = acc
torch.save(
model.state_dict(),
"/home/knarang2/UserCredibility/models/" +
dataset_NLP + "_LSTM.pt")
#print("Test Accuracy at epoch", epoch, acc)
print("Training and Test Accuracy and MRR at epoch", epoch,
'{0:.3f}'.format(train_acc), '{0:.3f}'.format(acc),
'{0:.3f}'.format(test_mrr))
if (train_acc - prev_train_acc) < 0.01:
print("Exiting training")
break
print("best_mrr and accuracy", best_MRR, best_accuracy)
result_list.append(best_MRR)
result_list_acc.append(best_accuracy)
except KeyboardInterrupt:
