def get_multitags_uda(conf_list):
uda_list = []
labels_map = []
for conf in conf_list:
tokenizer = tokenization.FullTokenizer(vocab_file=conf['vocab_file'],
do_lower_case=True)
processor = raw_data_utils.CFFEXProcessor(conf["labels"])
uda = get_uda.Get_UDA(conf, tokenizer, processor)
uda_list.append(uda)
labels_map.append(conf["labels"])
return uda_list, labels_map
def main(_):
if FLAGS.max_seq_length > 512:
raise ValueError(
"Cannot use sequence length {:d} because the BERT model "
"was only trained up to sequence length {:d}".format(
FLAGS.max_seq_length, 512))
# processor = raw_data_utils.get_processor(FLAGS.task_name)
processor = raw_data_utils.CFFEXProcessor(FLAGS.labels, FLAGS.raw_data_dir,
FLAGS.sup_train_file,
FLAGS.sup_dev_file,
FLAGS.unsup_file)
# Create tokenizer
tokenizer = tokenization.FullTokenizer(vocab_file=FLAGS.vocab_file,
do_lower_case=FLAGS.do_lower_case)
if FLAGS.data_type == "sup":
sup_out_dir = FLAGS.output_base_dir
tf.logging.info("Create sup. data: subset {} => {}".format(
FLAGS.sub_set, sup_out_dir))
proc_and_save_sup_data(
processor,
FLAGS.sub_set,
FLAGS.raw_data_dir,
sup_out_dir,
tokenizer,
FLAGS.max_seq_length,
FLAGS.trunc_keep_right,
FLAGS.worker_id,
FLAGS.replicas,
FLAGS.sup_size,
)
elif FLAGS.data_type == "unsup":
assert FLAGS.aug_ops is not None, \
"aug_ops is required to preprocess unsupervised data."
unsup_out_dir = os.path.join(FLAGS.output_base_dir, FLAGS.aug_ops,
str(FLAGS.aug_copy_num))
data_stats_dir = os.path.join(FLAGS.raw_data_dir, "data_stats")
tf.logging.info("Create unsup. data: subset {} => {}".format(
FLAGS.sub_set, unsup_out_dir))
proc_and_save_unsup_data(processor, FLAGS.sub_set, FLAGS.raw_data_dir,
data_stats_dir, unsup_out_dir, tokenizer,
FLAGS.max_seq_length, FLAGS.trunc_keep_right,
FLAGS.aug_ops, FLAGS.aug_copy_num,
FLAGS.worker_id, FLAGS.replicas)
def illegal_eval_uda():
tokenizer = tokenization.FullTokenizer(vocab_file=conf['vocab_file'], do_lower_case=True)
processor = raw_data_utils.CFFEXProcessor(conf["labels"])
uda = get_uda.Get_UDA(conf, tokenizer, processor)
eval_data = read_csv('ssf_uda/data/CFFEX/illegal', 'eval_0.2.csv')
check_data = pd.DataFrame(columns=('content', 'label', 'predict'))
# label_dict = {7:"恶意抹黑监管机构", 3:"非法发行证券",1:"非法荐股",4:"非法配资",5:"负面言论风险言论",6:"煽动维权诈骗",
# 2:"诱导开户",0:"正常"}
label_dict = ["未涉嫌违规", "非法投资咨询", "非法诱导投资", "非法经营活动", "维权追损二次诈骗"]
labels = conf['labels']
TP = 0.0
FN = 0.0
FP = 0.0
false_examples = []
for i in range(len(eval_data)):
content = eval_data.iloc[i]['content']
label = label_dict[int(eval_data.iloc[i]['label'])]
probs = uda.predict(content)
label_p = label_dict[int(get_label(labels, probs))]
check_data = check_data.append([{'content': content, 'label': json.dumps(label, ensure_ascii=False),
'predict': json.dumps(label_p, ensure_ascii=False)}], ignore_index=True)
if label != "未涉嫌违规" and label_p != "未涉嫌违规":
TP += 1
elif label != "未涉嫌违规" and label_p == "未涉嫌违规":
# false_examples.append([content, label, label_p])
FN += 1
elif label == "未涉嫌违规" and label_p != "未涉嫌违规":
false_examples.append([content, label, label_p])
FP += 1
for w in false_examples:
print(w[0])
print("ori:{}\tpredict:{}".format(w[1], w[2]))
pcsion = TP / (TP + FP)
recall = TP / (TP + FN)
to_csv(check_data, 'ssf_uda/data/CFFEX/illegal', 'illegal_uda.csv')
print("误识别为违规言论条数：{}".format(len(false_examples)))
print("违规言论精确率：{}, 召回率：{}".format(pcsion, recall))
uda.kill()
def main(_):
# for i in range(5):
# data = read_csv('ssf_uda/data/CFFEX/multitags/data_tag_{}'.format(i), 'data_tag_{}.csv'.format(i))
# train, eval = split_set_by_ratio(data,0.8)
# to_csv(train,'ssf_uda/data/CFFEX/multitags/data_tag_{}'.format(i), 'train_0.8.csv')
# to_csv(eval, 'ssf_uda/data/CFFEX/multitags/data_tag_{}'.format(i), 'eval_0.2.csv')
# print(len(data))
# tag_list = []
# # print(str(tag_list).decode("string_escape"))
# # for tag in tag_list:
# # print(tag)
# tag_set = split_multi_tags(tag_list)
# print(str(tag_set).decode("string_escape"))
# train = read_csv('ssf_uda/data/CFFEX/illegal', 'train_0.8.csv')
# show_info(train)
# for t in range(0, 30):
# cnt = 0
# for i in range(len(eval)):
# content = eval.iloc[i]["content"]
# label = eval.iloc[i]['label']
# res = re_match(content, label, t)
# if res == 1:
# cnt +=1
# print(1.0*cnt/len(eval))
# cnt = 0
# print(1.0 * cnt / len(eval))
# eval = read_csv('ssf_uda/data/CFFEX/illegal', 'eval_0.2.csv')
# label = train.iloc[i]['label']
# if not label_keyword_map.has_key(label):
# label_keyword_map[label] = []
# key_words = summary(train.iloc[i]['content'])
# for word in key_words:
# if word not in label_keyword_map[label]:
# label_keyword_map[label].append(word)
# for i in range(1,5):
# print(str(label_keyword_map[i]).decode("unicode_escape"))
# train = dataframe_keyword_extraction(train)
# to_csv(train, 'ssf_uda/data/CFFEX/illegal', 'train_extraction_0.8.csv')
# unsup = read_csv('ssf_uda/data/CFFEX/illegal', 'unsup.csv')
# unsup['label'] = 'unsup'
# to_csv(unsup, 'ssf_uda/data/CFFEX/illegal', 'unsup.csv')
# eval = read_csv('ssf_uda/data/CFFEX/illegal', 'eval_0.2.csv')
# eval = dataframe_keyword_extraction(eval)
# to_csv(eval, 'ssf_uda/data/CFFEX/illegal', 'eval_extraction_0.2.csv')
# eval = dataframe_keyword_extraction(eval)
# unsup = dataframe_keyword_extraction(unsup)
# to_csv(unsup, 'ssf_uda/data/CFFEX/illegal', 'unsup_extration.csv')
# unsup.insert(1,'label','unsup')
# labels = ['unsup']*len(unsup)
# unsup['label'] = labels
# print(unsup.head())
# to_csv(unsup,'ssf_uda/data/CFFEX/illegal', 'unsup.csv')
# train_data = read_csv('ssf_uda/data/CFFEX/illegal', 'unsup.csv')
# # train_data.drop(['text_split','emb_list'],axis=1,inplace=True)
# show_info(train_data)
# label_index = train_data.columns.get_loc('content')
# for i in range(len(train_data)):
# content = train_data.iloc[i]['content']
# content = keyword_extraction(content)
#
# train_data.iloc[i,label_index] = content
# # print(train_data.iloc[i]['content'])
# show_info(train_data)
# to_csv(train_data,'ssf_uda/data/CFFEX/illegal','unsup.csv')
# to_csv(train_data,'ssf_uda/data/CFFEX/illegal',"train_extraction_0.8.csv")
# eval = read_csv('ssf_uda/data/CFFEX/theme3', 'eval_0.2.csv')
# unsup = read_csv('ssf_uda/data/CFFEX/illegal', 'unsup.csv')
# show_info(train_data)
# show_info(eval)
# show_info(unsup)
if FLAGS.process_for_split:
train_data = read_csv(FLAGS.raw_data_dir, FLAGS.sup_train_file)
dev_data = read_csv(FLAGS.raw_data_dir, FLAGS.sup_dev_file)
# unsup_data = read_csv(FLAGS.raw_data_dir, FLAGS.unsup_file)
train_data['text_split'] = train_data['content'].apply(
lambda x: get_split(x, FLAGS.segment_len, FLAGS.segment_overlap))
dev_data['text_split'] = dev_data['content'].apply(
lambda x: get_split(x, FLAGS.segment_len, FLAGS.segment_overlap))
# unsup_data['text_split'] = unsup_data['content'].apply(lambda x: get_split(x, FLAGS.segment_len, FLAGS.segment_overlap))
to_csv(train_data, FLAGS.raw_data_dir, FLAGS.sup_train_file)
to_csv(dev_data, FLAGS.raw_data_dir, FLAGS.sup_dev_file)
# to_csv(unsup_data, FLAGS.raw_data_dir, FLAGS.unsup_file)
if FLAGS.process_for_embedding:
tokenizer = tokenization.FullTokenizer(vocab_file=FLAGS.vocab_file,
do_lower_case=True)
processor = raw_data_utils.CFFEXProcessor(FLAGS.labels)
uda = get_uda.Get_UDA(conf, tokenizer, processor)
train_data = read_csv(FLAGS.raw_data_dir, FLAGS.sup_train_file)
dev_data = read_csv(FLAGS.raw_data_dir, FLAGS.sup_dev_file)
train_data = encode_emb_list(uda, train_data)
dev_data = encode_emb_list(uda, dev_data)
to_csv(train_data, FLAGS.raw_data_dir, FLAGS.sup_train_file)
to_csv(dev_data, FLAGS.raw_data_dir, FLAGS.sup_dev_file)
uda.kill()
)
def predict(self, str):
try:
counter = timeCounter('predict stage')
counter()
self.q.put(str)
predict_result = self.predictor.next()
counter()
except Exception as e:
print(repr(e))
return predict_result
if __name__ == "__main__":
tokenizer = tokenization.FullTokenizer(vocab_file=conf['vocab_file'],
do_lower_case=True)
processor = raw_data_utils.get_processor(conf['task_name'])
tokenizer = tokenization.FullTokenizer(vocab_file=conf['vocab_file'],
do_lower_case=True)
uda = Get_UDA(conf, tokenizer, processor)
tf.logging.info("***** Running prediction *****")
str = '股指期货就像孙悟空。人跟神，不能做亲密的小伙伴，实在不好愉快的玩耍。'
predict_result = uda.predict(str)
# for key in predict_result.keys():
# predict_result[key] = predict_result[key].item()
p = predict_result
print(type(p['probabilities']))
print(type(list(p['probabilities'])))
# for p in predict_result:
# print(p)
# break