-
Notifications
You must be signed in to change notification settings - Fork 15
/
train.py
382 lines (321 loc) · 13.6 KB
/
train.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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
#! -*- coding:utf-8 -*-
import json
import numpy as np
from tqdm import tqdm
import time
import logging
from sklearn.model_selection import StratifiedKFold
from keras_bert import load_trained_model_from_checkpoint, Tokenizer
from keras.optimizers import Adam
import keras.backend.tensorflow_backend as KTF
from keras.layers import *
from keras.models import Model
import keras.backend as K
from keras.callbacks import Callback
import tensorflow as tf
import os
import pandas as pd
from keras.utils.np_utils import to_categorical
from sklearn.metrics import mean_absolute_error, accuracy_score, f1_score
import random
import argparse
# 超参数
parser = argparse.ArgumentParser()
parser.add_argument("--counter", default=0, type=int, required=False)
parser.add_argument("--name", default='', type=str, required=False)
parser.add_argument("--model", default=0, type=int, required=False)
parser.add_argument("--model1", default=1, type=int, required=False)
parser.add_argument("--title_len", default=128, type=int, required=False)
parser.add_argument("--content_len", default=512, type=int, required=False)
parser.add_argument("--learning_rate", default=5e-5, type=float, required=False)
parser.add_argument("--min_learning_rate", default=1e-5, type=float, required=False)
parser.add_argument("--random_seed", default=123, type=int, required=False)
parser.add_argument("--batch_size", default=16, type=int, required=False)
parser.add_argument("--epoch", default=8, type=int, required=False)
parser.add_argument("--fold", default=7, type=int, required=False)
args = parser.parse_args()
counter = args.counter
name = args.name
model = args.model
model1 = args.model1
MAX_LENT = args.title_len
MAX_LENC = args.content_len
learning_rate = args.learning_rate
min_learning_rate = args.min_learning_rate
random_seed = args.random_seed
bs = args.batch_size
epoch = args.epoch
fold = args.fold
# cpu运行
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
# 预训练所在文件夹
bert_path = ['chinese_L-12_H-768_A-12',
'chinese_wwm_ext_L-12_H-768_A-12',
'chinese_roberta_wwm_ext_L-12_H-768_A-12',
'roeberta_zh_L-24_H-1024_A-16']
# 不全部占满显存, 按需分配
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
KTF.set_session(sess)
# 加载对应预训练
config_path = './ckpt/' + bert_path[model] + '/bert_config.json'
checkpoint_path = './ckpt/' + bert_path[model] + '/bert_model.ckpt'
dict_path = './ckpt/' + bert_path[model] + '/vocab.txt'
config_path1 = './ckpt/' + bert_path[model1] + '/bert_config.json'
checkpoint_path1 = './ckpt/' + bert_path[model1] + '/bert_model.ckpt'
dict_path1 = './ckpt/' + bert_path[model1] + '/vocab.txt'
# 加载词汇表
token_dict = {}
with open(dict_path, 'r', encoding='utf-8') as reader:
for line in reader:
token = line.strip()
token_dict[token] = len(token_dict)
tokenizer = Tokenizer(token_dict)
token_dict1 = {}
with open(dict_path1, 'r', encoding='utf-8') as reader:
for line in reader:
token = line.strip()
token_dict1[token] = len(token_dict1)
tokenizer1 = Tokenizer(token_dict1)
file_path = './log/'
# 创建一个logger
logger = logging.getLogger('mylogger')
logger.setLevel(logging.DEBUG)
# 创建一个handler,
timestamp = time.strftime("%Y.%m.%d_%H.%M.%S", time.localtime())
fh = logging.FileHandler(file_path + 'log_' + timestamp + '.txt')
fh.setLevel(logging.DEBUG)
# 再创建一个handler,用于输出到控制台
ch = logging.StreamHandler()
ch.setLevel(logging.DEBUG)
# 定义handler的输出格式
formatter = logging.Formatter('[%(asctime)s][%(levelname)s] ## %(message)s')
fh.setFormatter(formatter)
ch.setFormatter(formatter)
# 给logger添加handler
logger.addHandler(fh)
logger.addHandler(ch)
# 数据读入与预处理
def read_data(file_path, id, name):
train_id = []
train_title = []
train_text = []
with open(file_path, 'r', encoding='utf-8-sig') as f:
for idx, line in enumerate(f):
line = line.strip().split(',')
train_id.append(line[0].replace('\'', '').replace(' ', ''))
train_title.append(line[1])
train_text.append(','.join(line[2:]))
output = pd.DataFrame(dtype=str)
output[id] = train_id
output[name + '_title'] = train_title
output[name + '_content'] = train_text
return output
# 读取数据
train_interrelation = pd.read_csv('./input/Train_Interrelation.csv', dtype=str)
Train_Achievements = read_data('./input/Train_Achievements.csv', 'Aid', 'Achievements')
Requirements = read_data('./input/Requirements.csv', 'Rid', 'Requirements')
TestPrediction = pd.read_csv('./input/TestPrediction.csv', dtype=str)
Test_Achievements = read_data('./input/Test_Achievements.csv', 'Aid', 'Achievements')
# 将train和test数据表连接成大表并选出有用信息
train = pd.merge(train_interrelation, Train_Achievements, on='Aid', how='left')
train = pd.merge(train, Requirements, on='Rid', how='left')
test = pd.merge(TestPrediction, Test_Achievements, on='Aid', how='left')
test = pd.merge(test, Requirements, on='Rid', how='left')
# 对数据进行预处理。
# 将内容为空白或如“图片”之类的无用信息替换为对应标题
for i in range(len(train)):
if len(train['Achievements_content'][i]) < 14:
train['Achievements_content'][i] = train['Achievements_title'][i]
if len(train['Requirements_content'][i]) < 10:
train['Requirements_content'][i] = train['Requirements_title'][i]
print("train预处理完毕")
for i in range(len(test)):
if len(test['Achievements_content'][i]) < 14:
test['Achievements_content'][i] = test['Achievements_title'][i]
if len(test['Requirements_content'][i]) < 10:
test['Requirements_content'][i] = test['Requirements_title'][i]
print("test预处理完毕")
train_achievements = train['Achievements_title'].values
train_requirements = train['Requirements_title'].values
train_achievementsc = train['Achievements_content'].values
train_requirementsc = train['Requirements_content'].values
test_achievements = test['Achievements_title'].values
test_requirements = test['Requirements_title'].values
test_achievementsc = test['Achievements_content'].values
test_requirementsc = test['Requirements_content'].values
labels = train['Level'].astype(int).values - 1
labels_cat = to_categorical(labels)
labels_cat = labels_cat.astype(np.int32)
# 数据生成
class data_generator:
def __init__(self, data, batch_size=bs):
self.data = data
self.batch_size = batch_size
self.steps = len(self.data[0]) // self.batch_size
if len(self.data[0]) % self.batch_size != 0:
self.steps += 1
def __len__(self):
return self.steps
def __iter__(self):
while True:
X1, X2, X3, X4, y = self.data
idxs = list(range(len(self.data[0])))
np.random.shuffle(idxs)
T, T_, Y = [], [], []
X, X_, Z = [], [], []
for c, i in enumerate(idxs):
achievements = X1[i]
requirements = X2[i]
achievementsc = X3[i]
requirementsc = X4[i]
t, t_ = tokenizer.encode(first=achievements, second=requirements, max_len=MAX_LENT)
x, x_ = tokenizer1.encode(first=achievementsc, second=requirementsc, max_len=MAX_LENC)
T.append(t)
T_.append(t_)
X.append(x)
X_.append(x_)
Y.append(y[i])
if len(T) == self.batch_size or i == idxs[-1]:
T = np.array(T)
T_ = np.array(T_)
X = np.array(X)
X_ = np.array(X_)
Y = np.array(Y)
yield [T, T_, X, X_], Y
T, T_, Y = [], [], []
X, X_, Z = [], [], []
# 模型构建
# 在第一个bert中对标题进行相似度判别
# 在第二个bert中对内容进行相似度判别
# 分别取出[CLS]进行拼接后进行分类
def get_model():
bert_model = load_trained_model_from_checkpoint(config_path, checkpoint_path)
bert_model1 = load_trained_model_from_checkpoint(config_path1, checkpoint_path1)
for l in bert_model.layers:
l.trainable = True
T1 = Input(shape=(None,))
T2 = Input(shape=(None,))
X1 = Input(shape=(None,))
X2 = Input(shape=(None,))
T = bert_model([T1, T2])
X = bert_model1([X1, X2])
T = Lambda(lambda x: x[:, 0])(T)
X = Lambda(lambda x: x[:, 0])(X)
T = Concatenate(axis=-1)([T, X])
T = Dense(384)(T)
# T = Dropout(0.1)(T)
output = Dense(4, activation='softmax')(T)
model = Model([T1, T2, X1, X2], output)
model.compile(
loss='categorical_crossentropy',
optimizer=Adam(1e-5), # 用足够小的学习率
metrics=['MAE']
)
model.summary()
return model
class Evaluate(Callback):
def __init__(self, val_data, val_index):
self.score = []
self.best = 0.
self.early_stopping = 0
self.val_data = val_data
self.val_index = val_index
self.predict = []
self.lr = 0
self.passed = 0
# 第一个epoch用来warmup,第二个epoch把学习率降到最低
def on_batch_begin(self, batch, logs=None):
if self.passed < self.params['steps']:
self.lr = (self.passed + 1.) / self.params['steps'] * learning_rate
K.set_value(self.model.optimizer.lr, self.lr)
self.passed += 1
elif self.params['steps'] <= self.passed < self.params['steps'] * 2:
self.lr = (2 - (self.passed + 1.) / self.params['steps']) * (learning_rate - min_learning_rate)
self.lr += min_learning_rate
K.set_value(self.model.optimizer.lr, self.lr)
self.passed += 1
def on_epoch_end(self, epoch, logs=None):
score, acc, f1 = self.evaluate()
if score > self.best:
self.best = score
self.early_stopping = 0
model.save_weights('./model_save/bert{}.w'.format(fold))
else:
self.early_stopping += 1
logger.info('fold: %d, lr: %.6f, score: %.4f, acc: %.4f, f1: %.4f,best: %.4f\n' % (
fold, self.lr, score, acc, f1, self.best))
def evaluate(self):
self.predict = []
prob = []
val_x1, val_x2, val_x3, val_x4, val_y, val_cat = self.val_data
for i in tqdm(range(len(val_x1))):
achievements = val_x1[i]
requirements = val_x2[i]
achievementsc = val_x3[i]
requirementsc = val_x4[i]
t1, t1_ = tokenizer.encode(first=achievements, second=requirements, max_len=MAX_LENT)
x1, x1_ = tokenizer1.encode(first=achievementsc, second=requirementsc, max_len=MAX_LENC)
T1, T1_ = np.array([t1]), np.array([t1_])
X1, X1_ = np.array([x1]), np.array([x1_])
_prob = model.predict([T1, T1_, X1, X1_])
oof_train[self.val_index[i]] = _prob[0]
self.predict.append(np.argmax(_prob, axis=1)[0] + 1)
prob.append(_prob[0])
score = 1.0 / (1 + mean_absolute_error(val_y + 1, self.predict))
acc = accuracy_score(val_y + 1, self.predict)
f1 = f1_score(val_y + 1, self.predict, average='macro')
return score, acc, f1
skf = StratifiedKFold(n_splits=fold, shuffle=True, random_state=random_seed)
def predict(data):
prob = []
val_x1, val_x2, val_x3, val_x4 = data
for i in tqdm(range(len(val_x1))):
achievements = val_x1[i]
requirements = val_x2[i]
achievementsc = val_x3[i]
requirementsc = val_x4[i]
t1, t1_ = tokenizer.encode(first=achievements, second=requirements, max_len=MAX_LENT)
x1, x1_ = tokenizer1.encode(first=achievementsc, second=requirementsc, max_len=MAX_LENC)
T1, T1_ = np.array([t1]), np.array([t1_])
X1, X1_ = np.array([x1]), np.array([x1_])
_prob = model.predict([T1, T1_, X1, X1_])
prob.append(_prob[0])
return prob
oof_train = np.zeros((len(train), 4), dtype=np.float32)
oof_test = np.zeros((len(test), 4), dtype=np.float32)
logger.info("加载{}和{}".format(bert_path[model], bert_path[model1]))
timestamp = time.time()
for fold, (train_index, valid_index) in enumerate(skf.split(train_achievements, labels)):
logger.info('------------ %d fold take: %.1f minute ------------' % (fold, (time.time() - timestamp) / 60))
timestamp = time.time()
x1 = train_achievements[train_index]
x2 = train_requirements[train_index]
x3 = train_achievementsc[train_index]
x4 = train_requirementsc[train_index]
y = labels_cat[train_index]
val_x1 = train_achievements[valid_index]
val_x2 = train_requirements[valid_index]
val_x3 = train_achievementsc[valid_index]
val_x4 = train_requirementsc[valid_index]
val_y = labels[valid_index]
val_cat = labels_cat[valid_index]
train_D = data_generator([x1, x2, x3, x4, y])
evaluator = Evaluate([val_x1, val_x2, val_x3, val_x4, val_y, val_cat], valid_index)
model = get_model()
model.fit_generator(train_D.__iter__(),
steps_per_epoch=len(train_D),
epochs=epoch,
callbacks=[evaluator]
)
model.load_weights('./model_save/bert{}.w'.format(fold))
oof_test += predict([test_achievements, test_requirements, test_achievementsc, test_requirementsc])
K.clear_session()
oof_test /= epoch
cv_score = 1.0 / (1 + mean_absolute_error(labels + 1, np.argmax(oof_train, axis=1) + 1))
logger.info(cv_score)
np.savetxt('./submit/w_{}.txt'.format(counter), oof_test)
test['Level'] = np.argmax(oof_test, axis=1) + 1
test[['Guid', 'Level']].to_csv('./submit/{}.csv'.format(counter), index=False)