def extract_triples(predicate_resume, disc_model_name=None):
date_time = strftime("%Y-%m-%d_%H_%M_%S", gmtime())
session = SnorkelSession()
if disc_model_name is None:
disc_model_name = "D" + predicate_resume["predicate_name"] + "Latest"
test_cands_query = get_test_cids_with_span(predicate_resume, session)
test_cands = test_cands_query.all()
lstm = reRNN()
logging.info("Loading marginals ")
lstm.load(disc_model_name)
predictions = lstm.predictions(test_cands)
dump_file_path3 = "./results/" + "triples_" + predicate_resume[
"predicate_name"] + date_time + ".csv"
subject_type = predicate_resume["subject_type"]
object_type = predicate_resume["object_type"]
subject_type_split = subject_type.split('/')
object_type_split = object_type.split('/')
subject_type_end = subject_type_split[len(subject_type_split) - 1]
object_type_end = object_type_split[len(object_type_split) - 1]
with open(dump_file_path3, 'w+b') as f:
writer = csv.writer(f, delimiter=',', quoting=csv.QUOTE_MINIMAL)
writer.writerow(["text", "marginal", "prediction"])
i = 0
for c in test_cands:
if predictions[i] == 1:
subject_span = getattr(c, "subject").get_span()
object_span = getattr(c, "object").get_span()
subject_uri = get_dbpedia_node(subject_span, subject_type_end)
object_uri = get_dbpedia_node(object_span, object_type_end)
predicate_uri = predicate_resume["predicate_URI"]
if subject_uri is not None and object_uri is not None:
row = [
str(subject_uri),
str(predicate_uri),
str(object_uri)
]
writer.writerow(row)
i = i + 1
def train_disc_model(predicate_resume, parallelism=8):
logging.info("Start training disc ")
session = SnorkelSession()
train_cids_query = get_train_cids_with_marginals_and_span(predicate_resume, session)
logging.info("Loading marginals ")
train_marginals = load_marginals(session, split=0, cids_query=train_cids_query)
train_kwargs = {
'lr': 0.01,
'dim': 50,
'n_epochs': 10,
'dropout': 0.25,
'print_freq': 1,
'max_sentence_length': 100
}
logging.info("Querying train cands")
candidate_subclass=predicate_resume["candidate_subclass"]
train_cands = session.query(candidate_subclass).filter(candidate_subclass.split == 0).order_by(candidate_subclass.id).all()#get_train_cands_with_marginals_and_span(predicate_resume, session).all()
logging.info("Querying dev cands")
dev_cands = get_dev_cands_with_span(predicate_resume, session).all()
logging.info("Querying gold labels")
L_gold_dev = get_gold_dev_matrix(predicate_resume, session)
logging.info("Training")
lstm = reRNN(seed=1701, n_threads=int(parallelism))
lstm.train(train_cands, train_marginals, **train_kwargs)
logging.info("Saving")
_save_model(predicate_resume, lstm)
#test model
candidate_subclass=predicate_resume["candidate_subclass"]
test_cands = session.query(candidate_subclass).filter(candidate_subclass.split == 2).order_by(candidate_subclass.id).all()
L_gold_test = get_gold_test_matrix(predicate_resume,session)
p, r, f1 = lstm.score(test_cands, L_gold_test)
print("Prec: {0:.3f}, Recall: {1:.3f}, F1 Score: {2:.3f}".format(p, r, f1))
logging.info("Prec: {0:.3f}, Recall: {1:.3f}, F1 Score: {2:.3f}".format(p, r, f1))
lstm.save_marginals(session, test_cands)
def score_disc_model(predicate_resume,
L_gold_test,
session,
date_time,
disc_model_name=None):
if disc_model_name is None:
disc_model_name = "D" + predicate_resume["predicate_name"] + "Latest"
candidate_subclass = predicate_resume["candidate_subclass"]
test_cands_query = session.query(candidate_subclass).filter(
candidate_subclass.split == 2).order_by(candidate_subclass.id)
test_cands = test_cands_query.all()
lstm = reRNN()
logging.info("Loading marginals ")
lstm.load(disc_model_name)
#lstm.save_marginals(session, test_cands)
p, r, f1 = lstm.score(test_cands, L_gold_test)
print("Prec: {0:.3f}, Recall: {1:.3f}, F1 Score: {2:.3f}".format(p, r, f1))
logging.info("Prec: {0:.3f}, Recall: {1:.3f}, F1 Score: {2:.3f}".format(
p, r, f1))
dump_file_path1 = "./results/" + "test_disc_1_" + predicate_resume[
"predicate_name"] + date_time + ".csv"
with open(dump_file_path1, 'w+b') as f:
writer = csv.writer(f, delimiter=',', quoting=csv.QUOTE_MINIMAL)
writer.writerow(["Precision", "Recall", "F1"])
writer.writerow(
["{0:.3f}".format(p), "{0:.3f}".format(r), "{0:.3f}".format(f1)])
tp, fp, tn, fn = lstm.error_analysis(session, test_cands, L_gold_test)
logging.info("TP: {}, FP: {}, TN: {}, FN: {}".format(
str(len(tp)), str(len(fp)), str(len(tn)), str(len(fn))))
dump_file_path2 = "./results/" + "test_disc_2_" + predicate_resume[
"predicate_name"] + date_time + ".csv"
with open(dump_file_path2, 'w+b') as f:
writer = csv.writer(f, delimiter=',', quoting=csv.QUOTE_MINIMAL)
writer.writerow(["TP", "FP", "TN", "FN"])
writer.writerow(
[str(len(tp)),
str(len(fp)),
str(len(tn)),
str(len(fn))])
predictions = lstm.predictions(test_cands)
marginals = lstm.marginals(test_cands)
dump_file_path3 = "./results/" + "test_disc_3_" + predicate_resume[
"predicate_name"] + date_time + ".csv"
with open(dump_file_path3, 'w+b') as f:
writer = csv.writer(f, delimiter=',', quoting=csv.QUOTE_MINIMAL)
writer.writerow(["text", "marginal", "prediction"])
i = 0
for candidate in test_cands:
start = candidate.subject.char_start
end = candidate.object.char_end + 1
if candidate.object.char_start < candidate.subject.char_start:
start = candidate.object.char_start
end = candidate.subject.char_end + 1
text = "\"" + candidate.get_parent().text[start:end].encode(
'ascii', 'ignore') + "\""
row = [text, str(marginals[i]), str(predictions[i])]
writer.writerow(row)
i = i + 1
dump_file_path4 = "./results/" + "triples_" + predicate_resume[
"predicate_name"] + date_time + ".csv"
subject_type = predicate_resume["subject_type"]
object_type = predicate_resume["object_type"]
subject_type_split = subject_type.split('/')
object_type_split = object_type.split('/')
subject_type_end = subject_type_split[len(subject_type_split) - 1]
object_type_end = object_type_split[len(object_type_split) - 1]
with open(dump_file_path4, 'w+b') as f:
writer = csv.writer(f, delimiter=',', quoting=csv.QUOTE_MINIMAL)
writer.writerow(["text", "marginal", "prediction"])
i = 0
for c in test_cands:
if predictions[i] == 1:
subject_span = getattr(c, "subject").get_span()
object_span = getattr(c, "object").get_span()
subject_uri = get_dbpedia_node(subject_span, subject_type_end)
object_uri = get_dbpedia_node(object_span, object_type_end)
predicate_uri = predicate_resume["predicate_URI"]
if subject_uri is not None and object_uri is not None:
row = [
str(subject_uri),
str(predicate_uri),
str(object_uri)
]
writer.writerow(row)
i = i + 1
print len(dev_cands)
# In[9]:
from snorkel.learning.disc_models.rnn import reRNN
train_kwargs = {
'lr': 0.01,
'dim': 50,
'n_epochs': 10,
'dropout': 0.25,
'print_freq': 1,
'max_sentence_length': 100
}
lstm = reRNN(seed=1701, n_threads=None)
lstm.train(train_cands,
train_marginals,
X_dev=dev_cands,
Y_dev=L_gold_dev,
**train_kwargs)
# In[10]:
p, r, f1 = lstm.score(test_cands, L_gold_test)
print("Prec: {0:.3f}, Recall: {1:.3f}, F1 Score: {2:.3f}".format(p, r, f1))
# In[11]:
tp, fp, tn, fn = lstm.error_analysis(session, test_cands, L_gold_test)
X_ends = np.array(map(lambda x: x[0], X_ends))
train_marginals = np.loadtxt(args["train_marginals"])
Y_dev = np.loadtxt(args["dev_labels"])
word_dict = read_word_dict(args["train_word_dict"])
train_marginals[train_marginals < 0] = 0
train_kwargs = {
'lr': 0.001,
'dim': 100,
'n_epochs': 10,
'dropout': 0.5,
'print_freq': 1,
'max_sentence_length': 2000,
}
lstm = reRNN(seed=100, n_threads=20)
lstm.word_dict = SymbolTable()
lstm.word_dict.d = word_dict
lstm.word_dict.s = max_val
np.random.seed(200)
training_size = int(len(X_train) * float(sys.argv[1]))
train_idx = np.random.randint(0, len(X_train), size=training_size)
lstm.train(X_train[train_idx],
X_ends[train_idx],
train_marginals[train_idx],
X_dev=X_dev,
Y_dev=Y_dev,
save_dir='{}'.format(args["save_dir"]),
**train_kwargs)
def run(candidate1, candidate2, pairing_name, cand1_ngrams, cand2_ngrams,
cand1Matcher, cand2Matcher, model_name, output_file_name,
corpus_parser):
print "Started"
session = SnorkelSession()
# The following line is for testing only. Feel free to ignore it.
candidate_pair = candidate_subclass(pairing_name, [candidate1, candidate2])
sentences = set()
docs = session.query(Document).order_by(Document.name).all()
for doc in docs:
for s in doc.sentences:
sentences.add(s)
cand_1_ngrams = Ngrams(n_max=cand1_ngrams)
# condition_ngrams = Ngrams(n_max=7)
cand_2_ngrams = Ngrams(n_max=cand2_ngrams)
# medium_ngrams = Ngrams(n_max=5)
# type_ngrams = Ngrams(n_max=5) # <--- Q: should we cut these down?
# # level_ngrams = Ngrams(n_max=1)
# unit_ngrams = Ngrams(n_max=1)
# Construct our Matchers
# cMatcher = matchers.getConditionMatcher()
# mMatcher = matchers.getMediumMatcher()
# tMatcher = matchers.getTypeMatcher()
# lMatcher = matchers.getLevelMatcher()
# uMatcher = matchers.getUnitMatcher()
# Building the CandidateExtractors
# candidate_extractor_BC = CandidateExtractor(BiomarkerCondition, [biomarker_ngrams, condition_ngrams], [bMatcher, cMatcher])
candidate_extractor = CandidateExtractor(candidate_pair,
[cand_1_ngrams, cand_2_ngrams],
[cand1Matcher, cand2Matcher])
# candidate_extractor_BM = CandidateExtractor(BiomarkerMedium, [biomarker_ngrams, medium_ngrams], [bMatcher, mMatcher])
# candidate_extractor_BT = CandidateExtractor(BiomarkerType, [biomarker_ngrams, type_ngrams], [bMatcher, tMatcher])
# candidate_extractor_BLU = CandidateExtractor(BiomarkerLevelUnit, [biomarker_ngrams, level_ngrams, unit_ngrams], [bMatcher, lMatcher, uMatcher])
# List of Candidate Sets for each relation type: [train, dev, test]
candidate_extractor.apply(sentences, split=4, clear=True)
cands = session.query(candidate_pair).filter(
candidate_pair.split == 4).order_by(candidate_pair.id).all()
session.commit()
# cands_BD = grabCandidates(candidate_extractor_BD, BiomarkerDrug)
# cands_BM = grabCandidates(candidate_extractor_BM, BiomarkerMedium)
# cands_BT = grabCandidates(candidate_extractor_BT, BiomarkerType)
# cands_BLU = grabCandidates(candidate_extractor_BLU, BiomarkerLevelUnit)
if (len(cands)) == 0:
print "No Candidates Found"
return
if (pairing_name == 'BiomarkerCondition'):
# session.rollback()
# print "Number of dev BC candidates without adj. boosting: ", len(cands_BC[1])
add_adj_candidate_BC(session, candidate_pair, cands, 4)
# fix_specificity(session, BiomarkerCondition, cands_BC[1])
# print "Number of dev BC candidates with adj. boosting: ", session.query(BiomarkerCondition).filter(BiomarkerCondition.split == 4).count()
session.commit()
lstm = reRNN(seed=1701, n_threads=None)
lstm.load(model_name)
predictions = lstm.predictions(cands)
output_file = open(output_file_name, 'wb')
import csv
csvWriter = csv.writer(output_file)
csvWriter.writerow(
['doc_id', 'sentence', candidate1, candidate2, 'prediction'])
for i in range(len(cands)):
doc_string = 'PMC' + str(cands[i].get_parent().get_parent())[9:]
sentence_string = cands[i].get_parent().text
cand_1_string = cands[i].get_contexts()[0].get_span()
cand_2_string = cands[i].get_contexts()[1].get_span()
prediction = predictions[i]
csvWriter.writerow([
unidecode(doc_string),
unidecode(sentence_string),
unidecode(cand_1_string),
unidecode(cand_2_string), prediction
])