def _create_eval_examples(self, lines):
examples = []
max_length = 0
for (i, line) in enumerate(lines):
try:
qas_id = int(line["query_id"])
query = tokenization.convert_to_unicode(line["query"])
query = clean(query)
answer_choice = tokenization.convert_to_unicode(line["alternatives"]).split("|")
answer_choice = [clean(ans) for ans in answer_choice]
answer_choice = list(set(answer_choice))
assert len(answer_choice) == 3
# random.shuffle(answer_choice)
context = tokenization.convert_to_unicode(line["passage"])
context = clean(context)
examples.append(data_feature_mrc.InputExample(
qas_id=qas_id,
question_text=query,
doc_tokens=context,
answer_choice=answer_choice,
choice=0
))
except:
continue
return examples
def _create_unsupervised_distillation_examples(self, lines,
distillation_prob):
examples = []
cnt = 0
for (i, line) in enumerate(lines):
content = line
guid = int(content["ID"])
text_a = content["sentence1"]
text_b = content["sentence2"]
label = content["gold_label"]
if isinstance(text_a, str) and isinstance(text_b, str):
text_a = tokenization.convert_to_unicode(text_a)
text_b = tokenization.convert_to_unicode(text_b)
input_labels = [label]
examples.append(
data_distillation_feature_classifier.InputExample(
guid=guid,
text_a=text_a,
text_b=text_b,
label=input_labels,
label_probs=distillation_prob[cnt],
label_ratio=0.0,
distillation_ratio=1.0))
cnt += 1
assert cnt == len(distillation_prob)
return examples
def _create_examples(self, lines, LABEL_SPLITTER="__label__"):
re_pattern = u"({}{})".format(LABEL_SPLITTER, "\d+")
examples = []
for (i, line) in enumerate(lines):
try:
guid = i
element_list = re.split(re_pattern, line)
text_a = clean("".join(element_list[-1].split()[1:]))
input_labels = clean(element_list[1]).split(LABEL_SPLITTER)[-1]
text_a = tokenization.convert_to_unicode(text_a)
input_labels = [
label.strip() for label in input_labels
if label.strip() in list(self.label2id.keys())
]
examples.append(
data_feature_classifier.InputExample(guid=guid,
text_a=text_a,
text_b=None,
label=input_labels))
except:
print(line, i)
return examples
def clean(text):
text = text.strip()
text = tokenization.convert_to_unicode(text)
text = HanziConv.toSimplified(text)
text = full2half(text)
text = re.sub(u"\\#.*?#|\\|.*?\\||\\[.*?]", "", text)
text = re.sub(u"\\s*", "", text)
return text
def _create_examples(self, lines):
examples = []
choice_cnt = {}
max_length = 0
for (i, line) in enumerate(lines):
try:
qas_id = int(line["query_id"])
query = tokenization.convert_to_unicode(line["query"])
query = clean(query)
answer = tokenization.convert_to_unicode(line["answer"])
answer = clean(answer)
answer_choice = tokenization.convert_to_unicode(line["alternatives"]).split("|")
answer_choice = [clean(ans) for ans in answer_choice]
answer_choice = list(set(answer_choice))
random.shuffle(answer_choice)
assert len(answer_choice) == 3
context = tokenization.convert_to_unicode(line["passage"])
context = clean(context)
for index, ans in enumerate(answer_choice):
if ans == answer:
choice = index
break
if choice in choice_cnt:
choice_cnt[choice] += 1
else:
choice_cnt[choice] = 1
examples.append(data_feature_mrc.InputExample(
qas_id=qas_id,
question_text=query,
doc_tokens=context,
answer_choice=answer_choice,
choice=choice
))
except:
continue
print(choice_cnt)
return examples
def _create_supervised_distillation_examples(self,
lines,
distillation_dict_lst,
LABEL_SPLITTER="__label__"):
re_pattern = u"({}{})".format(LABEL_SPLITTER, "\d+")
label_pattern = "(?<={})(\d+)".format(LABEL_SPLITTER)
examples = []
cnt = 0
for (i, line) in enumerate(lines):
try:
guid = i
element_list = re.split(re_pattern, line)
text_a = clean(element_list[-1])
input_labels = []
for l in re.finditer(label_pattern, line):
input_labels.append(l.group())
text_a = tokenization.convert_to_unicode(text_a)
input_labels = [
label.strip() for label in input_labels
if label.strip() in list(self.label2id.keys())
]
if len(input_labels) == 1:
assert int(input_labels[0]
) == distillation_dict_lst[cnt]["label_id"]
examples.append(
data_distillation_feature_classifier.InputExample(
guid=guid,
text_a=text_a,
text_b=None,
label=input_labels,
label_probs=distillation_dict_lst[cnt]["prob"],
label_ratio=1.0,
distillation_ratio=1.0,
feature=distillation_dict_lst[cnt]["feature"]))
cnt += 1
except:
print(line, i)
continue
assert cnt == len(distillation_dict_lst)
return examples
def _create_examples(self, lines, LABEL_SPLITTER="__label__"): examples = [] for (i, line) in enumerate(lines): guid = i element_list = line.split(LABEL_SPLITTER) text_a = tokenization.convert_to_unicode(element_list[0].strip()) text_a = clean(text_a) input_labels = element_list[1:] input_labels = [label.strip() for label in input_labels if label.strip() in list(self.label2id.keys())] examples.append(data_feature_classifier.InputExample( guid=guid, text_a=text_a, text_b=None, label=input_labels )) return examples
def _create_examples(self, frequent_phrases):
examples = []
for (i, line) in enumerate(frequent_phrases):
guid = i
text_a = clean(line[0])
input_labels = ["0"]
text_a = tokenization.convert_to_unicode(text_a)
input_labels = [
label.strip() for label in input_labels
if label.strip() in list(self.label2id.keys())
]
examples.append(
data_feature_classifier.InputExample(guid=guid,
text_a=text_a,
text_b=None,
label=input_labels))
return examples
def _create_supervised_distillation_examples(self,
lines,
distillation_prob,
LABEL_SPLITTER="__label__"):
re_pattern = u"({}{})".format(LABEL_SPLITTER, "\d+")
examples = []
# assert len(lines) == len(distillation_prob)
cnt = 0
for (i, line) in enumerate(lines):
try:
guid = i
element_list = re.split(re_pattern, line)
text_a = clean(element_list[-1])
input_labels = clean(element_list[1]).split(LABEL_SPLITTER)[-1]
except:
print(line, i)
continue
text_a = tokenization.convert_to_unicode(text_a)
input_labels = [
label.strip() for label in input_labels
if label.strip() in list(self.label2id.keys())
]
examples.append(
data_distillation_feature_classifier.InputExample(
guid=guid,
text_a=text_a,
text_b=None,
label=input_labels,
label_probs=distillation_prob[cnt],
label_ratio=1.0,
distillation_ratio=1.0))
cnt += 1
assert cnt == len(distillation_prob)
return examples
def _create_examples(self, lines, LABEL_SPLITTER="__label__"):
re_pattern = u"({}{})".format(LABEL_SPLITTER, "\d+")
label_pattern = "(?<={})(\d+)".format(LABEL_SPLITTER)
examples = []
for (i, line) in enumerate(lines):
try:
guid = i
element_list = re.split(re_pattern, line)
text_a = clean(element_list[-1])
input_labels = []
for l in re.finditer(label_pattern, line):
input_labels.append(l.group())
text_a = tokenization.convert_to_unicode(text_a)
input_labels = [
label.strip() for label in input_labels
if label.strip() in list(self.label2id.keys())
]
examples.append(
data_structure_distillation.InputExample(
guid=guid,
text_a=text_a,
text_b=None,
label=input_labels,
label_probs=[1.0 / len(self.label2id)] *
len(self.label2id),
label_ratio=1.0,
distillation_ratio=0.0,
feature=None))
except:
print(line, i)
return examples
print("End of dataset")
break
return pred_label, qas_id
print("===========begin to eval============")
[pred_label, qas_id] = eval_fn(result)
result = dict(zip(qas_id, pred_label))
print(len(result), "=====valid result======")
with tf.gfile.Open(FLAGS.eval_data_file, "r") as frobj:
qas_answer = {}
for line in frobj:
content = json.loads(line.strip())
qas_answer[int(
content["query_id"])] = tokenization.convert_to_unicode(
content["alternatives"]).split("|")
with tf.gfile.Open(FLAGS.result_file, "w") as fwobj:
cnt = 0
for index, key in enumerate(qas_answer):
if key in result:
cnt += 1
if index == 10:
print("==index=={}".format(index))
pred_ans = qas_answer[key][result[key]]
fwobj.write("\t".join([str(key), pred_ans]) + "\n")
else:
pred_ans = qas_answer[key][0]
fwobj.write("\t".join([str(key), pred_ans]) + "\n")
print(len(result), cnt, len(qas_answer),
