default=0,
help='write the test result, set 1 for writing result to file')
parser.add_argument('--set', type=int, default=0, help='test set or valid set')
parser.add_argument('--flag',
default="PER",
help='entity type (PER/LOC/ORG/MISC)')
parser.add_argument('--lr_rate', type=int, default=1e-4, help='learning rate')
parser.add_argument('--dataset',
default="conll2003",
help='name of the dataset')
parser.add_argument('--type',
default="bnpu",
help='pu learning type (bnpu/bpu/upu)')
args = parser.parse_args()
dp = DataPrepare(args.dataset)
mutils = FeaturedDetectionModelUtils(dp)
trainSet, validSet, testSet, prior = mutils.load_dataset(
args.flag, args.dataset, 1.0)
setIter = [testSet, validSet]
detectionSet = setIter[args.set]
fileNameIter = [
"data/" + args.dataset + "/test.txt", "data/" + args.dataset + "/valid.txt"
]
fileName = fileNameIter[args.set]
charcnn = CharCNN(dp.char2Idx)
wordnet = WordNet(dp.wordEmbeddings, dp.word2Idx)
casenet = CaseNet(dp.caseEmbeddings, dp.case2Idx)
featurenet = FeatureNet()
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--beta', type=float, default=0.0)
parser.add_argument('--gamma', type=float, default=1.0)
parser.add_argument('--drop_out', type=float, default=0.5)
parser.add_argument('--m', type=float, default=0.3)
parser.add_argument('--p', type=float, default=0.046)
parser.add_argument('--flag', default="PER")
parser.add_argument('--dataset', default="conll2003")
parser.add_argument('--unlabel_size', type=int, default=10000)
parser.add_argument('--iter', type=int, default=1)
parser.add_argument('--model', default="")
args = parser.parse_args()
dp = DataPrepare(args.dataset)
mutils = AddUnlabeledModelUtils(dp)
# use for training
dataset = mutils.load_dataset_(args.flag, args.dataset, args.iter)
trainSet, validSet, testSet, prior = mutils.load_origin_dataset(
dataset, args.p)
train_sentences_X, train_sentences_Y, train_sentences_LF = zip(*trainSet)
trainSize = len(trainSet)
validSize = len(validSet)
testSize = len(testSet)
print(("train set size: {}, valid set size: {}, test set size: {}").format(
trainSize, validSize, testSize))
charcnn = CharCNN(dp.char2Idx)
os.makedirs(fdir)
fname = os.path.join(fdir, "train.txt")
allSentences = []
for sent in trainSentences:
allSentences.append(sent)
for sent in unlabeledSentences:
allSentences.append(sent)
shuffle_indices = np.random.permutation(np.arange(len(allSentences)))
allSentences = np.array(allSentences)[shuffle_indices]
with open(fname, "w") as fw:
for sentence in allSentences:
for word, label, tagIdxList in sentence:
labeled = 0
fw.write(word + " " + str(label) + " " + str(labeled) + "\n")
fw.write("\n")
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="PU NER")
# data
parser.add_argument('--dataset', default="conll2003")
args = parser.parse_args()
dp = DataPrepare(args.dataset)
make_unlabeled_data(dp, "data/" + args.dataset + "/train.txt",
"data/eng.raw", args.dataset)
print("Data Set Writing Successfully")
def main_feature():
parser = argparse.ArgumentParser(description="PU NER")
# data
parser.add_argument('--lr', type=float, default=1e-4, help='learning rate')
parser.add_argument('--beta',
type=float,
default=0.0,
help='beta of pu learning (default 0.0)')
parser.add_argument('--gamma',
type=float,
default=1.0,
help='gamma of pu learning (default 1.0)')
parser.add_argument('--drop_out',
type=float,
default=0.5,
help='dropout rate')
parser.add_argument('--m',
type=float,
default=0.3,
help='class balance rate')
parser.add_argument('--flag',
default="Disease",
help='entity type (PER/LOC/ORG/MISC)')
parser.add_argument('--dataset',
default="disease",
help='name of the dataset')
parser.add_argument('--batch_size',
type=int,
default=100,
help='batch size for training and testing')
parser.add_argument('--print_time',
type=int,
default=1,
help='epochs for printing result')
parser.add_argument('--pert',
type=float,
default=1.0,
help='percentage of data use for training')
parser.add_argument('--type',
type=str,
default='bnpu',
help='pu learning type (bnpu/bpu/upu)') # bpu upu
parser.add_argument('--num', type=str, default='0', help='file number')
args = parser.parse_args()
dp = DataPrepare(args.dataset)
mutils = FeaturedDetectionModelUtils(dp)
trainSet, validSet, testSet, prior = mutils.load_dataset(
args.flag, args.dataset, args.pert, args.num)
f1_best_train = 0
trainSize = len(trainSet)
validSize = len(validSet)
testSize = len(testSet)
print(("train set size: {}, valid set size: {}, test set size: {}").format(
trainSize, validSize, testSize))
charcnn = CharCNN(dp.char2Idx)
wordnet = WordNet(dp.wordEmbeddings, dp.word2Idx)
casenet = CaseNet(dp.caseEmbeddings, dp.case2Idx)
featurenet = FeatureNet()
pulstmcnn = PULSTMCNN(dp, charcnn, wordnet, casenet, featurenet, 150, 200,
1, args.drop_out)
if torch.cuda.is_available:
charcnn.cuda()
wordnet.cuda()
casenet.cuda()
featurenet.cuda()
pulstmcnn.cuda()
torch.cuda.manual_seed(1013)
trainer = Trainer(pulstmcnn, prior, args.beta, args.gamma, args.lr, args.m)
time = 0
bar = ProgressBar(maxval=int((len(trainSet) - 1) / args.batch_size))
train_sentences = dp.read_origin_file("data/" + args.dataset +
"/train.txt")
trainSize = int(len(train_sentences) * args.pert)
train_sentences = train_sentences[:trainSize]
train_words = []
train_efs = []
for s in train_sentences:
temp = []
temp2 = []
for word, ef, lf in s:
temp.append(word)
temp2.append(ef)
train_words.append(temp)
train_efs.append(temp2)
valid_sentences = dp.read_origin_file("data/" + args.dataset +
"/valid.txt")
valid_words = []
valid_efs = []
for s in valid_sentences:
temp = []
temp2 = []
for word, ef, lf in s:
temp.append(word)
temp2.append(ef)
valid_words.append(temp)
valid_efs.append(temp2)
test_sentences = dp.read_origin_file("data/" + args.dataset + "/test.txt")
test_words = []
test_efs = []
for s in test_sentences:
temp = []
temp2 = []
for word, ef, lf in s:
temp.append(word)
temp2.append(ef)
test_words.append(temp)
test_efs.append(temp2)
for e in range(1, 1000):
print("Epoch: {}".format(e))
bar.start()
risks = []
prisks = []
nrisks = []
for step, (x_word_batch, x_case_batch, x_char_batch, x_feature_batch,
y_batch, flag_batch) in enumerate(
mutils.iterateSet(trainSet,
batchSize=args.batch_size,
mode="TRAIN")):
bar.update(step)
batch = [
x_word_batch, x_case_batch, x_char_batch, x_feature_batch,
y_batch, flag_batch
]
acc, risk, prisk, nrisk = trainer.train_mini_batch(batch, args)
risks.append(risk)
prisks.append(prisk)
nrisks.append(nrisk)
meanRisk = np.mean(np.array(risks))
meanRisk2 = np.mean(np.array(prisks))
meanRisk3 = np.mean(np.array(nrisks))
print("risk: {}, prisk: {}, nrisk: {}".format(meanRisk, meanRisk2,
meanRisk3))
if e % 5 == 0:
trainer.decay_learning_rate(e, args.lr)
if e % args.print_time == 0:
# train set
pred_train = []
corr_train = []
for step, (x_word_train_batch, x_case_train_batch,
x_char_train_batch, x_feature_train_batch,
y_train_batch) in enumerate(
mutils.iterateSet(trainSet,
batchSize=100,
mode="TEST",
shuffle=False)):
trainBatch = [
x_word_train_batch, x_case_train_batch, x_char_train_batch,
x_feature_train_batch
]
correcLabels = []
for x in y_train_batch:
for xi in x:
correcLabels.append(xi)
lengths = [len(x) for x in x_word_train_batch]
predLabels, _ = trainer.test(trainBatch, lengths)
correcLabels = np.array(correcLabels)
# print(predLabels)
# print(correcLabels)
assert len(predLabels) == len(correcLabels)
start = 0
for i, l in enumerate(lengths):
end = start + l
p = predLabels[start:end]
c = correcLabels[start:end]
pred_train.append(p)
corr_train.append(c)
start = end
# prec, rec, f1 = dp.evaluate_ner_tagging("data/" + args.dataset + "/train.txt",pred_train)
newSentences = []
for i, s in enumerate(train_words):
sent = []
assert len(s) == len(train_efs[i]) == len(pred_train[i])
for j, item in enumerate(s):
sent.append([item, train_efs[i][j], pred_train[i][j]])
newSentences.append(sent)
newSentences_, newLabels, newPreds = dp.wordLevelGeneration(
newSentences)
print("newlabels : ", len(newLabels))
print("newpreds : ", len(newPreds))
print(newLabels[0])
p_train, r_train, f1_train = dp.compute_precision_recall_f1(
newLabels, newPreds, args.flag, 1)
print("Precision: {}, Recall: {}, F1: {}".format(
p_train, r_train, f1_train))
print(trainer.bestResult)
if f1_train > f1_best_train:
f1_best_train = f1_train
p_best_train = p_train
r_best_train = f1_train
p_train_partial, r_train_partial, f1_train_partial, partial_entity_train, partial_ratio_train = dp.compute_precision_recall_f1_partial(
newLabels, newPreds, args.flag, 1, True)
# valid set
pred_valid = []
corr_valid = []
for step, (x_word_test_batch, x_case_test_batch, x_char_test_batch,
x_feature_test_batch, y_test_batch) in enumerate(
mutils.iterateSet(validSet,
batchSize=100,
mode="TEST",
shuffle=False)):
validBatch = [
x_word_test_batch, x_case_test_batch, x_char_test_batch,
x_feature_test_batch
]
correcLabels = []
for x in y_test_batch:
for xi in x:
correcLabels.append(xi)
lengths = [len(x) for x in x_word_test_batch]
predLabels, _ = trainer.test(validBatch, lengths)
correcLabels = np.array(correcLabels)
assert len(predLabels) == len(correcLabels)
start = 0
for i, l in enumerate(lengths):
end = start + l
p = predLabels[start:end]
c = correcLabels[start:end]
pred_valid.append(p)
corr_valid.append(c)
start = end
newSentencesValid = []
for i, s in enumerate(valid_words):
sent = []
assert len(s) == len(valid_efs[i]) == len(pred_valid[i])
for j, item in enumerate(s):
sent.append([item, valid_efs[i][j], pred_valid[i][j]])
newSentencesValid.append(sent)
newSentencesValid_, newLabelsValid, newPredsValid = dp.wordLevelGeneration(
newSentencesValid)
p_valid, r_valid, f1_valid = dp.compute_precision_recall_f1(
newLabelsValid, newPredsValid, args.flag, 1)
print("Precision: {}, Recall: {}, F1: {}".format(
p_valid, r_valid, f1_valid))
if f1_valid <= trainer.bestResult:
time += 1
else:
trainer.bestResult = f1_valid
time = 0
trainer.bestResult = f1_valid
best_precision_valid = p_valid
best_rappel_valid = r_valid
p_valid_partial, r_valid_partial, f1_valid_partial, partial_entity_valid, partial_ratio_valid = dp.compute_precision_recall_f1_partial(
newLabelsValid, newPredsValid, args.flag, 1, True)
trainer.save((
"saved_model/{}_{}_{}_{}_lr_{}_prior_{}_beta_{}_gamma_{}_percent_{}"
).format(args.num, args.type, args.dataset, args.flag,
trainer.learningRate, trainer.m, trainer.beta,
trainer.gamma, args.pert))
if time > 5:
f_result_valid = open('result_val.txt', 'a')
print(("BEST RESULT ON VALIDATE DATA:{}").format(
trainer.bestResult),
file=f_result_valid)
break
pulstmcnn.load_state_dict(
torch.load(
"saved_model/{}_{}_{}_{}_lr_{}_prior_{}_beta_{}_gamma_{}_percent_{}"
.format(args.num, args.type, args.dataset, args.flag,
trainer.learningRate, trainer.m, trainer.beta,
trainer.gamma, args.pert)))
pred_test = []
corr_test = []
for step, (x_word_test_batch, x_case_test_batch, x_char_test_batch,
x_feature_test_batch, y_test_batch) in enumerate(
mutils.iterateSet(testSet,
batchSize=100,
mode="TEST",
shuffle=False)):
testBatch = [
x_word_test_batch, x_case_test_batch, x_char_test_batch,
x_feature_test_batch
]
correcLabels = []
for x in y_test_batch:
for xi in x:
correcLabels.append(xi)
lengths = [len(x) for x in x_word_test_batch]
predLabels, _ = trainer.test(testBatch, lengths)
correcLabels = np.array(correcLabels)
assert len(predLabels) == len(correcLabels)
start = 0
for i, l in enumerate(lengths):
end = start + l
p = predLabels[start:end]
c = correcLabels[start:end]
pred_test.append(p)
corr_test.append(c)
start = end
newSentencesTest = []
for i, s in enumerate(test_words):
sent = []
assert len(s) == len(test_efs[i]) == len(pred_test[i])
for j, item in enumerate(s):
sent.append([item, test_efs[i][j], pred_test[i][j]])
newSentencesTest.append(sent)
newSentencesValid_, newLabelsValid, newPredsValid = dp.wordLevelGeneration(
newSentencesTest)
p_valid, r_valid, f1_valid = dp.compute_precision_recall_f1(
newLabelsValid, newPredsValid, args.flag, 1)
p_test_partial, r_test_partial, f1_test_partial, partial_entity_test, partial_ratio_test = dp.compute_precision_recall_f1_partial(
newLabelsValid, newPredsValid, args.flag, 1, True)
dutils = utils.dict_utils.DictUtils()
inDico = get_dico(dp, dutils, "data/disease/test.txt", "fulldico.txt",
"Disease", "Train", "disease")
p_test_gene, r_test_gene, f1_test_gene, partial_test_gene, rpartial_test_gene, f1partial_test_gene, nb_notindico = dp.compute_metrics_generalize(
newLabelsValid, newPredsValid, args.flag, 1, inDico)
result_file = "results/" + args.num + "_final_results.txt"
with open(result_file, "a+") as fw:
print(args.lr, args.beta, args.gamma, args.drop_out, args.m, file=fw)
print(p_best_train, r_best_train, f1_best_train, file=fw)
print(p_train_partial,
r_train_partial,
f1_train_partial,
partial_entity_train,
partial_ratio_train,
file=fw)
print(best_precision_valid,
best_rappel_valid,
trainer.bestResult,
file=fw)
print(p_valid_partial,
r_valid_partial,
f1_valid_partial,
partial_entity_valid,
partial_ratio_valid,
file=fw)
print(p_valid, r_valid, f1_valid, file=fw)
print(p_test_partial,
r_test_partial,
f1_test_partial,
partial_entity_test,
partial_ratio_test,
file=fw)
print(p_test_gene, r_test_gene, f1_test_gene, file=fw)
print(partial_test_gene,
rpartial_test_gene,
f1partial_test_gene,
file=fw)
print(nb_notindico, file=fw)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="PU NER EVL")
parser.add_argument('--dataset', default="conll2003")
parser.add_argument('--type', default="bnpu")
args = parser.parse_args()
filenames = [
'result/' + args.type + '_feature_pu_' + args.dataset + '_PER_0.txt',
'result/' + args.type + '_feature_pu_' + args.dataset + '_LOC_0.txt',
'result/' + args.type + '_feature_pu_' + args.dataset + '_ORG_0.txt',
'result/' + args.type + '_feature_pu_' + args.dataset + '_MISC_0.txt'
]
origin_file = "data/" + args.dataset + "/test.txt"
dp = DataPrepare(args.dataset)
test_sentences = dp.read_origin_file(origin_file)
test_words = []
test_efs = []
lens = []
for s in test_sentences:
temp = []
temp2 = []
for word, ef, lf in s:
temp.append(word)
temp2.append(ef)
test_words.append(temp)
test_efs.append(temp2)
lens.append(len(s))
from sub_model import CharCNN, CaseNet, WordNet
import numpy as np
torch.manual_seed(10)
parser = argparse.ArgumentParser(description="PU NER")
parser.add_argument('--model', default="")
parser.add_argument('--output', default=0)
parser.add_argument('--set', type=int, default=0)
parser.add_argument('--flag', default="PER")
parser.add_argument('--lr_rate', type=int, default=1e-4)
parser.add_argument('--dataset', default="conll2003")
parser.add_argument('--pert', type=float, default=1.0)
args = parser.parse_args()
dp = DataPrepare(args.dataset)
mutils = SupervisedModelUtils(dp)
trainSet, validSet, testSet = mutils.load_dataset(args.flag, args.dataset,
args.pert)
setIter = [testSet, validSet]
detectionSet = setIter[args.set]
fileNameIter = [
"data/" + args.dataset + "/test.txt", "data/" + args.dataset + "/valid.txt"
]
fileName = fileNameIter[args.set]
charcnn = CharCNN(dp.char2Idx)
wordnet = WordNet(dp.wordEmbeddings, dp.word2Idx)
casenet = CaseNet(dp.caseEmbeddings, dp.case2Idx)
if __name__ == "__main__":
from utils.add_unlabeled_data_model_utils import AddUnlabeledModelUtils
from utils.data_utils import DataPrepare
dp = DataPrepare("conll2003")
temp = AddUnlabeledModelUtils(dp)
temp.load_dataset_("PER", "conll2003", 10000)
temp.load_dataset_("PER", "conll2003", 20000)
parser.add_argument('--gamma', type=float, default=1.0, help='gamma of pu learning (default 1.0)')
parser.add_argument('--drop_out', type=float, default=0.5, help='dropout rate')
parser.add_argument('--m', type=float, default=1.5, help='class balance rate')
parser.add_argument('--p', type=float, default=0.1, help='estimate value of prior')
parser.add_argument('--flag', default="Disease", help='entity type (PER/LOC/ORG/MISC)')
parser.add_argument('--dataset', default="disease", help='name of the dataset')
parser.add_argument('--lr', type=float, default=1e-4, help='learning rate')
parser.add_argument('--batch_size', type=int, default=100, help='batch size for training and testing')
parser.add_argument('--output', default=0, help='write the test result, set 1 for writing result to file')
parser.add_argument('--num', type=str, default='', help='file number')
parser.add_argument('--model', default="", help='saved model name')
parser.add_argument('--iter', type=int, default=1, help='iteration time')
args = parser.parse_args()
dp = DataPrepare(args.dataset)
mutils = AdaptivePUUtils(dp)
dutils = DictUtils()
trainSet, validSet, testSet, prior = mutils.load_new_dataset(args.flag, args.dataset, args.iter, args.p, args.num)
print(prior)
trainSize = len(trainSet)
validSize = len(validSet)
testSize = len(testSet)
print(("train set size: {}, valid set size: {}, test set size: {}").format(trainSize, validSize, testSize))
charcnn = CharCNN(dp.char2Idx)
wordnet = WordNet(dp.wordEmbeddings, dp.word2Idx)
casenet = CaseNet(dp.caseEmbeddings, dp.case2Idx)
featurenet = FeatureNet()
pulstmcnn = AdaPULSTMCNN2(dp, charcnn, wordnet, casenet, featurenet, 150, 200, 1, args.drop_out)
for word, ef, lf in sentence:
alltokens += 1
if ef == 1:
labeledtokens += 1
else:
unlabeledtokens += 1
prior = float(labeledtokens / alltokens)
assert alltokens == labeledtokens + unlabeledtokens
print(prior)
if __name__ == "__main__":
dataset = "muc"
flag = "PER"
dp = DataPrepare(dataset)
filename = "data/" + dataset + "/train." + flag + ".txt"
sentences = dp.read_processed_file(filename, flag)
compute_all_prior(sentences)
flag = "LOC"
filename = "data/" + dataset + "/train." + flag + ".txt"
sentences = dp.read_processed_file(filename, flag)
compute_all_prior(sentences)
flag = "ORG"
filename = "data/" + dataset + "/train." + flag + ".txt"
sentences = dp.read_processed_file(filename, flag)
compute_all_prior(sentences)
# flag = "MISC"
# filename = "data/" + dataset + "/train." + flag + ".txt"
# sentences = dp.read_processed_file(filename, flag)
# compute_all_prior(sentences)