class Labeler:
def __init__(self):
self.word_state = {}
self.label_state = {}
self.hyperParams = HyperParams()
def createAlphabet(self, trainInsts, devInsts, testInsts):
print("create alpha.................")
for inst in trainInsts:
for w in inst.words:
if w not in self.word_state:
self.word_state[w] = 1
else:
self.word_state[w] += 1
for l in inst.labels:
if l not in self.label_state:
self.label_state[l] = 1
else:
self.label_state[l] += 1
print("word state:", len(self.word_state))
self.addTestAlpha(devInsts)
print("word state:", len(self.word_state))
self.addTestAlpha(testInsts)
print("word state:", len(self.word_state))
self.word_state[self.hyperParams.unk] = self.hyperParams.wordCutOff + 1
self.hyperParams.wordAlpha.initial(self.word_state,
self.hyperParams.wordCutOff)
self.hyperParams.wordAlpha.set_fixed_flag(True)
self.hyperParams.wordNum = self.hyperParams.wordAlpha.m_size
self.hyperParams.unkWordID = self.hyperParams.wordAlpha.from_string(
self.hyperParams.unk)
self.hyperParams.labelAlpha.initial(self.label_state)
self.hyperParams.labelAlpha.set_fixed_flag(True)
self.hyperParams.labelSize = self.hyperParams.labelAlpha.m_size
print("Label num: ", self.hyperParams.labelSize)
print("Word num: ", self.hyperParams.wordNum)
def addTestAlpha(self, insts):
print("Add test alpha.............")
if self.hyperParams.wordFineTune == False:
for inst in insts:
for w in inst.words:
if (w not in self.word_state):
self.word_state[w] = 1
else:
self.word_state[w] += 1
def extractFeature(self, inst):
feat = Feature()
for w in inst.words:
wordId = self.hyperParams.wordAlpha.from_string(w)
if wordId == -1:
feat.wordIndexs.append(self.hyperParams.unkWordID)
else:
feat.wordIndexs.append(wordId)
feat.wordIndexs = torch.autograd.Variable(
torch.LongTensor(feat.wordIndexs))
return feat
def instance2Example(self, insts):
exams = []
for inst in insts:
example = Example()
example.feat = self.extractFeature(inst)
for l in inst.labels:
labelId = self.hyperParams.labelAlpha.from_string(l)
example.labelIndexs.append(labelId)
example.labelIndexs = torch.autograd.Variable(
torch.LongTensor(example.labelIndexs))
exams.append(example)
return exams
def train(self, train_file, dev_file, test_file):
self.hyperParams.show()
torch.set_num_threads(self.hyperParams.thread)
reader = Reader(self.hyperParams.maxInstance)
trainInsts = reader.readInstances(train_file)
devInsts = reader.readInstances(dev_file)
testInsts = reader.readInstances(test_file)
print("Training Instance: ", len(trainInsts))
print("Dev Instance: ", len(devInsts))
print("Test Instance: ", len(testInsts))
self.createAlphabet(trainInsts, devInsts, testInsts)
trainExamples = self.instance2Example(trainInsts)
devExamples = self.instance2Example(devInsts)
testExamples = self.instance2Example(testInsts)
self.model = RNNLabeler(self.hyperParams)
parameters = filter(lambda p: p.requires_grad, self.model.parameters())
optimizer = torch.optim.Adagrad(parameters,
lr=self.hyperParams.learningRate)
indexes = []
for idx in range(len(trainExamples)):
indexes.append(idx)
for iter in range(self.hyperParams.maxIter):
print('###Iteration' + str(iter) + "###")
random.shuffle(indexes)
for idx in range(len(trainExamples)):
self.model.zero_grad()
self.model.LSTMHidden = self.model.init_hidden()
exam = trainExamples[indexes[idx]]
lstm_feats = self.model(exam.feat)
loss = self.model.crf.neg_log_likelihood(
lstm_feats, exam.labelIndexs)
loss.backward()
optimizer.step()
if (idx + 1) % self.hyperParams.verboseIter == 0:
print('current: ', idx + 1, ", cost:", loss.data[0])
eval_dev = Eval()
for idx in range(len(devExamples)):
predictLabels = self.predict(devExamples[idx])
devInsts[idx].evalPRF(predictLabels, eval_dev)
print('Dev: ', end="")
eval_dev.getFscore()
eval_test = Eval()
for idx in range(len(testExamples)):
predictLabels = self.predict(testExamples[idx])
testInsts[idx].evalPRF(predictLabels, eval_test)
print('Test: ', end="")
eval_test.getFscore()
def predict(self, exam):
tag_hiddens = self.model(exam.feat)
_, best_path = self.model.crf._viterbi_decode(tag_hiddens)
predictLabels = []
for idx in range(len(best_path)):
predictLabels.append(
self.hyperParams.labelAlpha.from_id(best_path[idx]))
return predictLabels
def getMaxIndex(self, tag_score):
max = tag_score.data[0]
maxIndex = 0
for idx in range(1, self.hyperParams.labelSize):
if tag_score.data[idx] > max:
max = tag_score.data[idx]
maxIndex = idx
return maxIndex
class Labeler:
def __init__(self):
self.word_state = {}
self.label_state = {}
self.hyperParams = HyperParams()
self.wordAlpha = Alphabet()
self.labelAlpha = Alphabet()
def createAlphabet(self, trainInsts):
for inst in trainInsts:
for w in inst.words:
if w not in self.word_state:
self.word_state[w] = 1
else:
self.word_state[w] += 1
for l in inst.labels:
if l not in self.label_state:
self.label_state[l] = 1
else:
self.label_state[l] += 1
self.wordAlpha.initial(self.word_state, self.hyperParams.wordCutOff)
self.labelAlpha.initial(self.label_state)
self.labelAlpha.set_fixed_flag(True)
self.wordAlpha.set_fixed_flag(True)
self.hyperParams.wordNum = self.wordAlpha.m_size
self.hyperParams.labelSize = self.labelAlpha.m_size
print("word num: ", self.hyperParams.wordNum)
print("label num: ", self.hyperParams.labelSize)
def extractFeature(self, inst):
feat = Feature()
for w in inst.words:
wordId = self.wordAlpha.from_string(w)
feat.wordIndexs.append(wordId)
feat.wordIndexs = torch.autograd.Variable(
torch.LongTensor(feat.wordIndexs))
return feat
def instance2Example(self, insts):
exams = []
for inst in insts:
example = Example()
example.feat = self.extractFeature(inst)
for l in inst.labels:
labelId = self.labelAlpha.from_string(l)
example.labelIndexs.append(labelId)
example.labelIndexs = torch.autograd.Variable(
torch.LongTensor(example.labelIndexs))
exams.append(example)
return exams
def train(self, train_file, dev_file, test_file):
self.hyperParams.show()
torch.set_num_threads(self.hyperParams.thread)
reader = Reader(self.hyperParams.maxInstance)
trainInsts = reader.readInstances(train_file)
devInsts = reader.readInstances(dev_file)
trainExamples = self.instance2Example(trainInsts)
devExamples = self.instance2Example(devInsts)
print("Training Instance: ", len(trainInsts))
print("Dev Instance: ", len(devInsts))
self.createAlphabet(trainInsts)
self.model = RNNLabeler(self.hyperParams)
optimizer = torch.optim.Adagrad(self.model.parameters(),
lr=self.hyperParams.learningRate)
indexes = []
for idx in range(len(trainExamples)):
indexes.append(idx)
for iter in range(self.hyperParams.maxIter):
print('###Iteration' + str(iter) + "###")
random.shuffle(indexes)
for idx in range(len(trainExamples)):
self.model.zero_grad()
self.model.LSTMHidden = self.model.init_hidden()
exam = trainExamples[indexes[idx]]
tag_scores = self.model(exam.feat)
loss = torch.nn.functional.cross_entropy(
tag_scores, exam.labelIndexs)
loss.backward()
optimizer.step()
if (idx + 1) % self.hyperParams.verboseIter == 0:
print('current: ', idx + 1, ", cost:", loss.data[0])
eval_dev = Eval()
for idx in range(len(devExamples)):
predictLabels = self.predict(devExamples[idx])
devInsts[idx].evalPRF(predictLabels, eval_dev)
eval_dev.getFscore()
def predict(self, exam):
tag_scores = self.model(exam.feat)
if len(tag_scores) != len(exam.labelIndexs) or len(
tag_scores.data[0]) != self.hyperParams.labelSize:
print("error")
predictIndexs = []
for idx in range(len(tag_scores)):
pred_idx = self.getMaxIndex(tag_scores[idx])
predictIndexs.append(pred_idx)
predictLabels = []
for idx in range(len(tag_scores)):
predictLabels.append(self.labelAlpha.from_id(predictIndexs[idx]))
return predictLabels
def getMaxIndex(self, tag_score):
max = tag_score.data[0]
maxIndex = 0
for idx in range(1, self.hyperParams.labelSize):
if tag_score.data[idx] > max:
max = tag_score.data[idx]
maxIndex = idx
return maxIndex
class Labeler:
def __init__(self):
self.word_state = {}
self.label_state = {}
self.hyperParams = HyperParams()
def createAlphabet(self, trainInsts, devInsts, testInsts):
print("create alpha.................")
for inst in trainInsts:
for w in inst.words:
if w not in self.word_state:
self.word_state[w] = 1
else:
self.word_state[w] += 1
l = inst.label
if l not in self.label_state:
self.label_state[l] = 1
else:
self.label_state[l] += 1
print("word state:", len(self.word_state))
self.addTestAlpha(devInsts)
print("word state:", len(self.word_state))
self.addTestAlpha(testInsts)
print("word state:", len(self.word_state))
self.word_state[self.hyperParams.unk] = self.hyperParams.wordCutOff + 1
self.word_state[self.hyperParams.padding] = self.hyperParams.wordCutOff + 1
self.hyperParams.wordAlpha.initial(self.word_state, self.hyperParams.wordCutOff)
self.hyperParams.wordAlpha.set_fixed_flag(True)
self.hyperParams.wordNum = self.hyperParams.wordAlpha.m_size
self.hyperParams.unkWordID = self.hyperParams.wordAlpha.from_string(self.hyperParams.unk)
self.hyperParams.paddingID = self.hyperParams.wordAlpha.from_string(self.hyperParams.padding)
self.hyperParams.labelAlpha.initial(self.label_state)
self.hyperParams.labelAlpha.set_fixed_flag(True)
self.hyperParams.labelSize = self.hyperParams.labelAlpha.m_size
print("Label num: ", self.hyperParams.labelSize)
print("Word num: ", self.hyperParams.wordNum)
print("Padding ID: ", self.hyperParams.paddingID)
print("UNK ID: ", self.hyperParams.unkWordID)
def addTestAlpha(self, insts):
print("Add test alpha.............")
if self.hyperParams.wordFineTune == False:
for inst in insts:
for w in inst.words:
if (w not in self.word_state):
self.word_state[w] = 1
else:
self.word_state[w] += 1
def extractFeature(self, inst):
feat = Feature()
feat.sentLen = len(inst.words)
feat.wordIndexs = torch.autograd.Variable(torch.LongTensor(1, feat.sentLen))
for idx in range(len(inst.words)):
w = inst.words[idx]
wordId = self.hyperParams.wordAlpha.from_string(w)
if wordId == -1:
wordId = self.hyperParams.unkWordID
feat.wordIndexs.data[0][idx] = wordId
return feat
def instance2Example(self, insts):
exams = []
for inst in insts:
example = Example()
example.labelIndex = torch.autograd.Variable(torch.LongTensor(1))
example.feat = self.extractFeature(inst)
l = inst.label
labelId = self.hyperParams.labelAlpha.from_string(l)
example.labelIndex.data[0] = labelId
exams.append(example)
return exams
def getBatchFeatLabel(self, exams):
maxSentSize = 0
for e in exams:
if maxSentSize < e.feat.sentLen:
maxSentSize = e.feat.sentLen
if maxSentSize > 40:
maxSentSize = 40
batch_feats = torch.autograd.Variable(torch.LongTensor(self.hyperParams.batch, maxSentSize))
batch_labels = torch.autograd.Variable(torch.LongTensor(self.hyperParams.batch))
for idx in range(len(batch_feats.data)):
e = exams[idx]
batch_labels.data[idx] = e.labelIndex.data[0]
for idy in range(maxSentSize):
if idy < e.feat.sentLen:
batch_feats.data[idx][idy] = e.feat.wordIndexs.data[0][idy]
else:
batch_feats.data[idx][idy] = self.hyperParams.paddingID
return batch_feats, batch_labels
def train(self, train_file, dev_file, test_file):
self.hyperParams.show()
torch.set_num_threads(self.hyperParams.thread)
reader = Reader()
trainInsts = reader.readInstances(train_file, self.hyperParams.maxInstance)
devInsts = reader.readInstances(dev_file, self.hyperParams.maxInstance)
testInsts = reader.readInstances(test_file, self.hyperParams.maxInstance)
print("Training Instance: ", len(trainInsts))
print("Dev Instance: ", len(devInsts))
print("Test Instance: ", len(testInsts))
self.createAlphabet(trainInsts, devInsts, testInsts)
trainExamples = self.instance2Example(trainInsts)
devExamples = self.instance2Example(devInsts)
testExamples = self.instance2Example(testInsts)
self.model = RNNLabeler(self.hyperParams)
parameters = filter(lambda p: p.requires_grad, self.model.parameters())
optimizer = torch.optim.Adam(parameters, lr=self.hyperParams.learningRate)
indexes = []
for idx in range(len(trainExamples)):
indexes.append(idx)
batchBlock = len(trainExamples) // self.hyperParams.batch
for iter in range(self.hyperParams.maxIter):
print('###Iteration' + str(iter) + "###")
random.shuffle(indexes)
self.model.train()
for updateIter in range(batchBlock):
#self.model.zero_grad()
optimizer.zero_grad()
exams = []
start_pos = updateIter * self.hyperParams.batch
end_pos = (updateIter + 1) * self.hyperParams.batch
for idx in range(start_pos, end_pos):
exams.append(trainExamples[indexes[idx]])
feats, labels = self.getBatchFeatLabel(exams)
output = self.model(feats, self.hyperParams.batch)
loss = torch.nn.functional.cross_entropy(output, labels)
loss.backward()
optimizer.step()
if (updateIter + 1) % self.hyperParams.verboseIter == 0:
print('current: ', idx + 1, ", cost:", loss.data[0])
self.model.eval()
eval_dev = Eval()
for idx in range(len(devExamples)):
predictLabel = self.predict(devExamples[idx])
devInsts[idx].evalACC(predictLabel, eval_dev)
print("dev: ", end='')
eval_dev.getACC()
eval_test = Eval()
for idx in range(len(testExamples)):
predictLabel = self.predict(testExamples[idx])
testInsts[idx].evalACC(predictLabel, eval_test)
print("test: ", end='')
eval_test.getACC()
def predict(self, exam):
output = self.model(exam.feat.wordIndexs)
labelID = self.getMaxIndex(output)
return self.hyperParams.labelAlpha.from_id(labelID)
def getMaxIndex(self, tag_score):
max = tag_score.data[0][0]
maxIndex = 0
for idx in range(1, self.hyperParams.labelSize):
if tag_score.data[0][idx] > max:
max = tag_score.data[0][idx]
maxIndex = idx
return maxIndex
class Labeler:
def __init__(self):
self.hyperParams=HyperParams()
pass
def train(self,trainFile,devFile,testFile):
self.hyperParams.show()
train=DataSet(trainFile)
dev=DataSet(devFile)
test=DataSet(testFile)
vocab=Alphabet(train)
self.hyperParams.vocabSize=vocab.size()
self.hyperParams.labelSize=vocab.label_size()
print('vocab_size:',self.hyperParams.vocabSize)
print('label_size:',self.hyperParams.labelSize)
train=IndexSet(train,vocab)
dev=IndexSet(dev,vocab)
test=IndexSet(test,vocab)
print('trainset_size: ',train.size())
print('devset_size: ',dev.size())
print('testset_size: ',test.size())
if self.hyperParams.embedFile!='':
pretrain=PretrainEmb(self.hyperParams.embedFile,vocab.word2id)
else:
pretrain=None
##############################
self.model = Encoder(self.hyperParams,pretrain) #encoder
self.crf = CRF(self.hyperParams.labelSize,vocab.label2id['<start>'],vocab.label2id['<padding>'],vocab)#decoder
parameters = filter(lambda p: p.requires_grad, self.model.parameters())
optimizer_rnn = torch.optim.Adam(parameters, lr = self.hyperParams.learningRate)
optimizer_crf = torch.optim.Adam(self.crf.parameters(), lr = self.hyperParams.learningRate)
##############################
indexes = []
for idx in range(train.size()):
indexes.append(idx)
batchBlock = len(train.word_mat) // self.hyperParams.batch
for iter in range(self.hyperParams.maxIter):#################
print('###Iteration' + str(iter) + "###")
random.shuffle(indexes)
self.model.train()###
for updateIter in range(batchBlock):
#self.model.zero_grad()
optimizer_rnn.zero_grad()
optimizer_crf.zero_grad()
start_pos = updateIter * self.hyperParams.batch
end_pos = (updateIter + 1) * self.hyperParams.batch
feats=[]
labels=[]
for idx in range(start_pos, end_pos):
feats.append(train.word_mat[indexes[idx]])
labels.append(train.label_mat[indexes[idx]])
batch=BatchBucket(len(feats),self.hyperParams.maxSentSize,feats,labels,vocab.word2id['<padding>'],vocab.label2id['<padding>'])
tag_scores = self.model(batch.batch_words, self.hyperParams.batch)
#print(tag_scores.size())
loss = self.crf.neg_log_likelihood(tag_scores, batch.batch_labels,batch.masks)
loss.backward()
optimizer_rnn.step()
optimizer_crf.step()
if (updateIter + 1) % self.hyperParams.verboseIter == 0:
print('current: ', idx + 1, ", cost:", loss.data[0])
self.model.eval()###
self.eval_predict(dev,vocab)
self.eval_predict(test,vocab)
def eval_predict(self,indexset,vocab) :
correct_num=0
total_num=0
batchBlock=len(indexset.label_mat)//self.hyperParams.batch
for updateIter in range(batchBlock):
start_pos = updateIter * self.hyperParams.batch
end_pos = (updateIter + 1) * self.hyperParams.batch
feats=[]
labels=[]
for idx in range(start_pos, end_pos):
feats.append(indexset.word_mat[idx])
labels.append(indexset.label_mat[idx])
batch=BatchBucket(len(feats),self.hyperParams.maxSentSize,feats,labels,vocab.word2id['<padding>'],vocab.label2id['<padding>'])
tag_scores = self.model(batch.batch_words, self.hyperParams.batch)
predict_labels=self.crf.viterbi_decode(tag_scores,batch.masks)
predict_labels=predict_labels.masked_select(batch.masks)
gold_labels=batch.batch_labels.masked_select(batch.masks)
correct_num+=torch.sum(torch.gt(predict_labels.float(),gold_labels.float())).data[0]
total_num+=torch.sum(batch.masks).data[0]
rate=correct_num/total_num
print('total_num: {} , correct_num: {}'.format(total_num,correct_num))
print('rate: ',rate)
