def eval(sentence, predict, tags):
entityres, entityall = [], []
entityres = calculate(sentence, predict.permute(1, 0), id2word, id2tag,
entityres)
entityall = calculate(sentence, tags, id2word, id2tag, entityall)
intersection = []
acc, recall, f1 = [], [], []
for i in range(len(entityres)):
intersection.append([j for j in entityres[i] if j in entityall[i]])
for i in range(len(intersection)):
if len(entityres[i]) == 0:
acc_ = 1.
else:
acc_ = float(len(intersection[i]) / len(entityres[i]))
if len(entityall[i]) == 0:
recall_ = 1.
else:
recall_ = float(len(intersection[i]) / len(entityall[i]))
if acc_ + recall_ == 0.:
f1_ = 0.
else:
f1_ = (2 * acc_ * recall_) / (acc_ + recall_)
acc.append(acc_)
recall.append(recall_)
f1.append(f1_)
return np.mean(acc), np.mean(recall), np.mean(f1)
def train(model,sess):
for epoch in range(training_epochs):
for batch in range(batch_num):
x_batch, y_batch = data_train.next_batch(batch_size)
feed_dict = {model.input_data:x_batch, model.labels:y_batch}
pre,_ = sess.run([model.viterbi_sequence,model.train_op], feed_dict)
acc = 0
if batch%100==0:
for i in range(len(y_batch)):
for j in range(len(y_batch[0])):
if y_batch[i][j]==pre[i][j]:
acc+=1
print float(acc)/(len(y_batch)*len(y_batch[0]))
path_name = "./model/model"+str(epoch)+".ckpt"
print path_name
if epoch%2==0:
saver.save(sess, path_name)
print "model has been saved"
entityres=[]
entityall=[]
for batch in range(batch_num):
x_batch, y_batch = data_train.next_batch(batch_size)
feed_dict = {model.input_data:x_batch, model.labels:y_batch}
pre = sess.run([model.viterbi_sequence], feed_dict)
pre = pre[0]
entityres = calculate(x_batch,pre,id2word,id2tag,entityres)
entityall = calculate(x_batch,y_batch,id2word,id2tag,entityall)
jiaoji = [i for i in entityres if i in entityall]
if len(jiaoji)!=0:
zhun = float(len(jiaoji))/len(entityres)
zhao = float(len(jiaoji))/len(entityall)
print "train"
print "zhun:", zhun
print "zhao:", zhao
print "f:", (2*zhun*zhao)/(zhun+zhao)
else:
print "zhun:",0
entityres=[]
entityall=[]
for batch in range(batch_num_test):
x_batch, y_batch = data_test.next_batch(batch_size)
feed_dict = {model.input_data:x_batch, model.labels:y_batch}
pre = sess.run([model.viterbi_sequence], feed_dict)
pre = pre[0]
entityres = calculate(x_batch,pre,id2word,id2tag,entityres)
entityall = calculate(x_batch,y_batch,id2word,id2tag,entityall)
jiaoji = [i for i in entityres if i in entityall]
if len(jiaoji)!=0:
zhun = float(len(jiaoji))/len(entityres)
zhao = float(len(jiaoji))/len(entityall)
print "test"
print "zhun:", zhun
print "zhao:", zhao
print "f:", (2*zhun*zhao)/(zhun+zhao)
else:
print "zhun:",0
def train():
model = BiLSTM_CRF(len(word2id) + 1, tag2id, EMBEDDING_DIM,
HIDDEN_DIM) # 实例化模型
optimizer = optim.SGD(model.parameters(), lr=0.005,
weight_decay=1e-4) # 随机梯度下降优化算法
# 训练
for epoch in range(EPOCHS):
index = 0
for sentence, tags in zip(x_train, y_train):
index += 1 # 计量每次迭代进行了多少步
model.zero_grad() # 清空梯度
sentence = torch.tensor(sentence, dtype=torch.long)
tags = torch.tensor([tag2id[t] for t in tags], dtype=torch.long)
loss = model.neg_log_likelihood(sentence, tags) # 模型定义的loss
loss.backward()
optimizer.step()
if index % 300 == 0:
print("epoch", epoch, "batch", index)
# 用来保存测试结果
entityres = []
entityall = []
# 每个epoch后测试一下
for sentence, tags in zip(x_test, y_test):
sentence = torch.tensor(sentence, dtype=torch.long)
tags = torch.tensor([tag2id[t] for t in tags], dtype=torch.long)
score, predict = model(sentence)
entityres = calculate(sentence, predict, id2word, id2tag,
entityres)
entityall = calculate(sentence, tags.numpy(), id2word, id2tag,
entityall)
jiaoji = [i for i in entityres if i in entityall]
if len(jiaoji) != 0:
zhun = float(len(jiaoji)) / len(entityres) # 准确率
zhao = float(len(jiaoji)) / len(entityall) # 召回率
print("test:")
print("precision:", zhun)
print("recall:", zhao)
print("F:", (2 * zhun * zhao) / (zhun + zhao))
else:
print("precision:", 0)
path_name = "./model/model.pkl"
print(path_name)
torch.save(model, path_name)
print("model has been saved")
def test():
path_name = "./model/model.pkl"
model = torch.load(path_name)
entityres = []
entityall = []
# 测试
for sentence, tags in zip(x_test, y_test):
sentence = torch.tensor(sentence, dtype=torch.long)
score, predict = model(sentence)
entityres = calculate(sentence, predict, id2word, id2tag, entityres)
entityall = calculate(sentence, tags, id2word, id2tag, entityall)
jiaoji = [i for i in entityres if i in entityall]
if len(jiaoji) != 0:
zhun = float(len(jiaoji)) / len(entityres) # 准确率
zhao = float(len(jiaoji)) / len(entityall) # 召回率
print("test:")
print("zhun:", zhun)
print("zhao:", zhao)
print("f:", (2 * zhun * zhao) / (zhun + zhao))
else:
print("zhun:", 0)
sentence = torch.tensor(sentence, dtype=torch.long)
tags = torch.tensor([tag2id[t] for t in tags], dtype=torch.long)
loss = model.neg_log_likelihood(sentence, tags)
loss.backward()
optimizer.step()
if index % 300 == 0:
print("epoch", epoch, "index", index)
entityres = []
entityall = []
for sentence, tags in zip(x_test, y_test):
sentence = torch.tensor(sentence, dtype=torch.long)
score, predict_tags = model(sentence)
entityres = calculate(sentence, predict_tags, id2word, id2tag,
entityres)
entityall = calculate(sentence, tags, id2word, id2tag, entityall)
common_entity = [i for i in entityres if i in entityall]
if len(common_entity) != 0:
zhun = float(len(common_entity)) / len(entityres)
zhao = float(len(common_entity)) / len(entityall)
print("test:")
print("Percision:", zhun)
print("Recall:", zhao)
print("F1:", (2 * zhun * zhao) / (zhun + zhao))
else:
print("Percision:", 0)
path_name = "C:\\Users\\Administrator\\PycharmProjects\\ChineseNER\\save_model\\save_model" + str(
epoch) + ".pkl"
for j in range(len(y_batch[0])):
if y_batch[i][j] == pre[i][j]:
acc += 1
print float(acc) / (len(y_batch) * len(y_batch[0]))
path_name = "./model/model" + str(epoch) + ".ckpt"
print path_name
saver.save(sess, path_name)
print "model has been saved"
entityres = []
entityall = []
for batch in range(batch_num):
x_batch, y_batch = data_train.next_batch(batch_size)
feed_dict = {input_data: x_batch, labels: y_batch}
pre = sess.run([viterbi_sequence], feed_dict)
pre = pre[0]
entityres = calculate(x_batch, pre, id2word, id2tag, entityres)
entityall = calculate(x_batch, y_batch, id2word, id2tag, entityall)
jiaoji = [i for i in entityres if i in entityall]
if len(jiaoji) != 0:
zhun = float(len(jiaoji)) / len(entityres)
zhao = float(len(jiaoji)) / len(entityall)
print "train"
print "zhun:", zhun
print "zhao:", zhao
print "f:", (2 * zhun * zhao) / (zhun + zhao)
else:
print "zhun:", 0
entityres = []
entityall = []
for batch in range(batch_num_test):
valLoader = DataLoader(dataset=validation, batch_size=batch_size, shuffle=True)
testLoader = DataLoader(dataset=test, batch_size=batch_size, shuffle=True)
for epoch in range(EPOCHS):
train_loss = 0.0
entityres = []
entityall = []
for i, data in enumerate(trainLoader):
x, y = data
x = x.permute(1, 0, 2)
#y = y.permute(1,0,2)
optimizer.zero_grad()
loss = model.neg_log_likelihood(x, y)
train_loss += loss
_, pred = model(x)
entityres = calculate(x, pred, id2word, id2tag, entityres)
entityall = calculate(x, y, id2word, id2tag, entityall)
train_acc, train_recall, train_f1 = matrix(entityres, entityall)
iter = epoch * len(trainLoader) + i
writer.add_scalar("train_loss", train_loss, iter)
writer.add_scalar("train_acc", train_acc, iter)
writer.add_scalar("train_recall", train_recall, iter)
writer.add_scalar("train_f1", train_f1, iter)
loss.backward()
optimizer.step()
val_loss = 0.0
val_res = []
val_all = []
for i, data in enumerate(valLoader):
x, y = data
x = x.permute(1, 0, 2)