-
Notifications
You must be signed in to change notification settings - Fork 0
/
sentiment_analysis.py
276 lines (236 loc) · 9.21 KB
/
sentiment_analysis.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
import pandas as pd
from torch.utils.data import DataLoader
from torch.optim import AdamW
from transformers import BertConfig, BertTokenizer, get_constant_schedule_with_warmup
from torch.utils.data import Dataset
from torch.nn.utils.rnn import pad_sequence
import numpy as np
from torch import nn
import json
import random
import os
from apex import amp
from tqdm import tqdm as tqdm
import torch
from transformers import BertTokenizer, BertModel, BertConfig
from sklearn.model_selection import KFold
def seed_everything(seed=1234):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
seed_everything()
usual_train = pd.read_excel('2020_SMP_raw_data/usual_train.xlsx')
virus_train = pd.read_excel('2020_SMP_raw_data/virus_train.xlsx')
usual_test = pd.read_excel('2020_SMP_raw_data/usual_eval.xlsx')
virus_test = pd.read_excel('2020_SMP_raw_data/virus_eval.xlsx')
usual_train['type'] = 'usual'
usual_test['type'] = 'usual'
virus_train['type'] = 'virus'
virus_test['type'] = 'virus'
print(usual_train.head())
print(virus_train.head())
data = usual_train.append(virus_train)
data['文本'] = data['文本'].astype('str')
# 打乱数据
data = data.reset_index(drop=True)
print(data.shape)
label_dict = {}
for label in data['情绪标签'].unique():
label_dict[label] = len(label_dict)
print(label_dict)
data['情绪标签'] = data['情绪标签'].map(label_dict)
# 定义基本组件
tokenizer = BertTokenizer.from_pretrained("ernie")
config = BertConfig.from_json_file('ernie/config.json')
bert_path = 'ernie'
config.num_labels = len(label_dict)
class SentimentDataset(Dataset):
def __init__(self, df, valid=False):
self.raw_sentence = df['文本'].tolist()
self.label = df['情绪标签'].tolist()
self.type = df['type'].tolist()
self._tokenizer = tokenizer
self.max_len = 600
self.sentence = self.sen_tokenize()
def sen_tokenize(self):
result = []
for index, item in enumerate(self.raw_sentence):
# if self.type[index] == 'virus':
# item = '[疫情]'+item
# else:
# item = '[普通]'+item
vector = self._tokenizer.encode(item[0:self.max_len])
# # 添加标志位区分普通数据和疫情数据
# if self.type[index]=='usual':
# vector.insert(1,1)
# else:
# vector.insert(1,2)
result.append(vector)
# print(item)
# print(self._tokenizer.decode(vector))
return result
def __len__(self):
return len(self.sentence)
def __getitem__(self, idx):
return torch.tensor(self.sentence[idx]), self.label[idx]
def collate_fn(batch):
token, label = zip(*batch)
label = torch.tensor(label)
token = pad_sequence(token, batch_first=True)
return token,label
class SentimentModel(nn.Module):
def __init__(self, config, bert_model):
super().__init__()
self.num_labels = config.num_labels
self.bert = BertModel.from_pretrained(bert_model)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
# self.lstm = nn.LSTM(768,config.hidden_size,1,batch_first=True)
# self.classify = nn.Sequential(
# # nn.BatchNorm1d(config.hidden_size),
# nn.Dropout(p=0.5),
# nn.Linear(in_features=config.hidden_size, out_features=config.num_labels)
# )
def forward(
self,
input_ids=None,
attention_mask=None
):
outputs = self.bert(
input_ids,
attention_mask=attention_mask
)
sequence_output = outputs[1]
logits = self.classifier(sequence_output)
return logits
def get_learning_rate(optimizer):
lr = []
for param_group in optimizer.param_groups:
lr += [param_group['lr']]
lr = lr[0]
return lr
batch_size = 4
lr = 3e-5
weight_decay = 0
adam_epsilon = 1e-8
n_epochs = 2
step = 1
warmup = 0.05
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
torch.cuda.set_device(0)
kfold = KFold(n_splits=5, shuffle=True, random_state=2019)
r = 0
for train_index, test_index in kfold.split(np.zeros(len(data))):
train = data.loc[train_index,:].reset_index()
val = data.loc[test_index,:].reset_index()
model = SentimentModel(config, bert_path)
model.to(device)
# 准确训练模型
# Prepare optimizer and schedule (linear warmup and decay)
optimizer = AdamW([
{'params': model.bert.parameters(), 'lr': 2e-5}
], lr=1e-3)
t_total = int(len(train)*n_epochs/batch_size)
warmup_steps = int(t_total*warmup)
scheduler = get_constant_schedule_with_warmup(optimizer, num_warmup_steps=warmup_steps)
# model, optimizer = amp.initialize(model, optimizer, opt_level='O2', verbosity=0)
report_each = 100
loss_fn = nn.CrossEntropyLoss()
for e in range(n_epochs):
model.train()
train_losses = []
train_set = SentimentDataset(train)
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
valid_set = SentimentDataset(val, valid=True)
valid_loader = DataLoader(valid_set, batch_size=batch_size, shuffle=False, collate_fn=collate_fn)
for i, (token, label) in tqdm(enumerate(train_loader)):
input_mask = (token > 0).to(device)
token, label = token.to(device), label.to(device)
outputs = model(input_ids=token, attention_mask=input_mask)
loss = loss_fn(outputs, label)
# if (i + 1) % step == 0:
# with amp.scale_loss(loss, optimizer) as scaled_loss:
loss.backward()
optimizer.step()
optimizer.zero_grad()
scheduler.step()
# else:
# loss.backward()
train_losses.append(loss.item())
mean_loss = np.mean(train_losses[-report_each:])
if i % 1500 == 0:
print('loss: ', mean_loss)
lr = get_learning_rate(optimizer)
# validate
model.eval()
valid_losses = 0
pred_set = []
for i, (token, label) in tqdm(enumerate(valid_loader)):
input_mask = (token > 0).to(device)
token, label = token.to(device), label.to(device)
with torch.no_grad():
outputs = model(input_ids=token, attention_mask=input_mask)
loss = loss_fn(outputs,label)
valid_losses += loss.item()
pred_set.append(outputs.cpu().numpy())
# token = token.cpu()
# valid_loss = valid_loss / len(dev_X)
pred_set = np.concatenate(pred_set, axis=0)
label_set = val['情绪标签']
top_class = np.argmax(pred_set, axis=1)
equals = top_class == label_set
accuracy = np.mean(equals)
usual_acc = np.mean(equals[val['type']=='usual'])
virus_acc = np.mean(equals[val['type']=='virus'])
print('acc %f, usual acc %f, virus acc %f' % (accuracy,usual_acc,virus_acc))
print('epoch %d, train loss %f, val loss %f' % (r, sum(train_losses)/len(train_loader), valid_losses/len(valid_loader)))
torch.save(model.state_dict(), 'model/model_%d.pth' % r)
r += 1
# submit
usual_test['情绪标签'] = -1
virus_test['情绪标签'] = -1
usual_test_set = SentimentDataset(usual_test, valid=True)
usual_test_loader = DataLoader(usual_test_set, batch_size=batch_size,shuffle=False, collate_fn=collate_fn)
virus_test_set = SentimentDataset(virus_test, valid=True)
virus_test_loader = DataLoader(virus_test_set, batch_size=batch_size,shuffle=False, collate_fn=collate_fn)
id_label = {}
for key,value in label_dict.items():
id_label[value] = key
def make_prediction(test_data, df):
preds = None
for r in range(5):
model.load_state_dict(torch.load('model/model_%d.pth'%r))
model.eval()
pred_set = []
for i, (token, label) in enumerate(test_data):
input_mask = (token > 0).to(device)
token, label = token.to(device), label.to(device)
with torch.no_grad():
outputs = model(input_ids=token, attention_mask=input_mask)
pred_set.append(nn.Softmax(dim=1)(outputs).cpu().numpy())
pred_set = np.concatenate(pred_set, axis=0)
if preds is not None:
preds += pred_set
else:
preds = pred_set
top_class = np.argmax(preds, axis=1)
result = []
for index, id in enumerate(df['数据编号']):
line = {}
line['id'] = id
line['label'] = id_label[top_class[index]]
result.append(line)
return result
usual_result = make_prediction(usual_test_loader, usual_test)
virus_result = make_prediction(virus_test_loader, virus_test)
with open('usual_result.txt', 'w', encoding='utf-8') as f:
json.dump(usual_result, f)
with open('virus_result.txt', 'w', encoding='utf-8') as f:
json.dump(virus_result, f)
# acc 0.710072, usual acc 0.700700, virus acc 0.740106
# acc 0.705052, usual acc 0.691643, virus acc 0.748021
# acc 0.703640, usual acc 0.697200, virus acc 0.724274
# train loss 0.709274, val loss 0.839431
# acc 0.701031, usual acc 0.692712, virus acc 0.727118