-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtopic_ana_predict.py
executable file
·234 lines (212 loc) · 8.3 KB
/
topic_ana_predict.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
import re
import numpy as np
from tqdm import tqdm
# import math
# import MeCab
import json
import gensim
import pyLDAvis
import pyLDAvis.gensim
#from wordcloud import WordCloud
import matplotlib.pylab as plt
import os, sys
import kenlm
from datasets import load_dataset
import glob
from datetime import datetime
import fire
import utils
doc_num = 100000
def read_text(data, ppl):
if data["meta"]["ppl"] <= ppl:
return data["text"]
else:
return ""
num = re.compile('[0-9]+')
symbol = re.compile('[!"#$%&\'\\\\()*+,-./:;<=>?@[\\]^_`{|}~「」〔〕“”〈〉『』【】＆＊・（）＄＃＠。、？！｀＋￥％]')
alpha = p = re.compile('[a-z]|[A-Z]')
def get_topic_number(corpus, lda_model):
topics_number = []
for topics_per_document in lda_model[corpus]:
topic_number = 0
max_score = 0.0
for topics in topics_per_document:
if max_score < topics[1]:
topic_number = topics[0]
max_score = topics[1]
topics_number.append(topic_number+1)
return topics_number
#sudachi-kenlm
m = kenlm.LanguageModel("kenlm_merge-code_0.05_model.bin")
def cal_ppl(wakati_text):
ppl = m.perplexity(wakati_text)
return ppl
def remove_meta_from_cc(text):
paragraphs = text.split("\n\n")
pars = []
for par in paragraphs:
idx = par.find('\x1c')
if idx >= 0:
par = par[idx + 1:]
par = par.replace("\x02", "").replace("\x03", "")
pars.append(par)
return "\n\n".join(pars)
def read_cc(file):
# dataset = load_dataset("parquet", data_files=file)
print(file)
dataset = load_dataset("json", data_files=file)
train_ds = dataset["train"]
new_dataset = []
for data in train_ds:
new_data = {}
new_data["text"] = remove_meta_from_cc(data["text"])
if "docId" in data:
new_data["meta"] = {"docId":data["docId"], "url":data["docId"], "charset":data["charset"], "date":data["date"]}
new_dataset.append(new_data)
return new_dataset
# Textize datetime obj before JSONize
def serialize_dates(obj):
if isinstance(obj, dict):
return {k: serialize_dates(v) for k, v in obj.items()}
elif isinstance(obj, list):
return [serialize_dates(i) for i in obj]
elif isinstance(obj, datetime):
return obj.isoformat()
else:
return obj
def load_text_ja(calc_ppl=False):
# input_dir_path = "/scratch/ace14333cp/ja_llm/llm-jp-corpus/data/cc/ja_cc/raw"
#dataset = "/scratch/ace14333cp/ja_llm/classifier/data/evaluate/mc4-ja_token-num/dir1/0000_mc4-ja_token-num.jsonl"
target_ppls = [660, 4120, 57000, float('inf')]
count = 0
texts = []
new_dataset = []
folder_path = "/data/llm-jp-corpus-v2.1-CC/filtered_v3.0/segment=*/filter=null/"
paths = sorted(glob.glob(folder_path))
for i in range(0,100):
# file_index = str(i*10).zfill(5)
# input_file = os.path.join(input_dir_path, "part-" + file_index + "-55624510-b8c5-443c-8ca9-04d95450cbb6.c000.zstd.parquet")
path = paths[i%len(paths)]
gz_files = sorted([os.path.join(path, f) for f in os.listdir(path) if f.endswith('.gz')])
dataset = read_cc(gz_files[i%len(paths)])
for data in dataset:
#clean text
if count%1000 == 0:
print("noun extract", count, "clear!")
text = data["text"]
if text != "":
try:
texts.append(text)
if calc_ppl:
words, wakati_text = analyzer(text)
ppl = cal_ppl(wakati_text)
data["ppl"] = ppl
count += 1
new_dataset.append(data)
except:
pass
if count >= doc_num:
break
else:
continue
break
print("count:", count, "texts:", len(texts))
return texts, new_dataset
def load_text_en():
path = "corpus/en_pile_merge_1.200k.jsonl"
dataset = read_cc(path)
texts = [d['text'] for d in dataset]
# import ipdb; ipdb.set_trace()
print("texts:", len(texts))
return texts, dataset
# 最初からトピックモデル学習
def train_topic_model(work_dir, texts, new_dataset, lang="jp", num_topics=17):
tokenizer = utils.Tokenizer(lang)
texts = [tokenizer(p) for p in tqdm(texts, desc="Tokenizing")]
#save texts
with open(f"{work_dir}/topic_texts.jsonl", "w") as f:
for nouns, data in zip(texts, new_dataset):
data["nouns"] = nouns
json.dump(serialize_dates(data), f, ensure_ascii=False)
f.write("\n")
dictionary = gensim.corpora.Dictionary(texts)
dictionary.save_as_text(f"{work_dir}/dictionary.txt")
dictionary.filter_extremes(no_below=3, no_above=0.3)
dictionary.save(f"{work_dir}/dictionary.filtered.pkl")
corpus = [dictionary.doc2bow(t) for t in texts]
# import ipdb; ipdb.set_trace()
lda_model = gensim.models.ldamodel.LdaModel(corpus=corpus,
id2word=dictionary,
num_topics=num_topics,
random_state=0)
vis_pcoa = pyLDAvis.gensim.prepare(lda_model, corpus, dictionary, sort_topics=False)
lda_model.save(f'{work_dir}/lda_100k.model')
pyLDAvis.save_html(vis_pcoa, f'{work_dir}/pyldavis_lda_100k.html')
print("saved lda model")
# 辞書からトピックモデル学習
def middle(work_dir):
file = "/scratch/ace14333cp/ja_llm/bert_train/topic_texts.jsonl"
texts = []
with open(file, "r") as f:
for line in f:
data = json.loads(line)
texts.append(data["nouns"])
dictionary = gensim.corpora.Dictionary(texts)
dictionary.save_as_text(f"{work_dir}/dictionary.txt")
dictionary.filter_extremes(no_below=3, no_above=0.3)
corpus = [dictionary.doc2bow(t) for t in texts]
print("created corpus!")
num_topics = 16
lda_model = gensim.models.ldamodel.LdaModel(corpus=corpus,
id2word=dictionary,
num_topics=num_topics,
random_state=0)
vis_pcoa = pyLDAvis.gensim.prepare(lda_model, corpus, dictionary, sort_topics=False)
lda_model.save('lda_100k.model')
print("saved lda model")
pyLDAvis.save_html(vis_pcoa, 'pyldavis_lda_100k.html')
# 推論
def cal_topic_number(work_dir,file, lang="jp"):
tokenizer = utils.Tokenizer(lang)
# file = f"{work_dir}/topic_texts.jsonl"
texts = []
dataset = []
with open(file, "r") as f:
for line in tqdm(f):
data = json.loads(line)
dataset.append(data)
texts.append(tokenizer(data["text"]))
# dictionary = gensim.corpora.Dictionary(texts)
# dictionary.save_as_text(f"{work_dir}/dictionary.txt")
# dictionary.filter_extremes(no_below=3, no_above=0.3)
dictionary = gensim.corpora.Dictionary.load(f"{work_dir}/dictionary.filtered.pkl")
# import ipdb; ipdb.set_trace()
corpus = [dictionary.doc2bow(t) for t in texts]
topics_number = []
lda_model = gensim.models.ldamodel.LdaModel.load(f"{work_dir}/lda_100k.model")
for topics_per_document in lda_model[corpus]:
topic_number = 0
max_score = 0.0
for topics in topics_per_document:
if max_score < topics[1]:
topic_number = topics[0]
max_score = topics[1]
topics_number.append(topic_number+1)
with open(f"{work_dir}/topic_ana_100k.jsonl", "w") as f:
for data, topic in zip(dataset, topics_number):
data["topic"] = topic
json.dump(data, f, ensure_ascii=False)
f.write("\n")
def main(lang, num_topics=17):
work_dir=f'output.{lang}.{num_topics}'
os.makedirs(work_dir, exist_ok=True)
if lang=="jp":
texts, new_dataset = load_text_ja()
train_topic_model(work_dir, texts, new_dataset, lang='jp', num_topics=num_topics)
cal_topic_number (work_dir, "corpus/CC-MAIN.txt", lang="jp")
elif lang=="en":
texts, new_dataset = load_text_en()
train_topic_model(work_dir, texts, new_dataset, lang='en', num_topics=num_topics)
cal_topic_number (work_dir, "corpus/en_pile_merge_2.200k.jsonl", lang="en")
if __name__ == "__main__":
fire.Fire(main)