-
Notifications
You must be signed in to change notification settings - Fork 161
/
Copy pathtask_question_answer_generation_by_seq2seq.py
180 lines (146 loc) · 6.61 KB
/
task_question_answer_generation_by_seq2seq.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
#! -*- coding: utf-8- -*-
# 用Seq2Seq做阅读理解构建
# 根据篇章先采样生成答案，然后采样生成问题
# 数据集同 https://github.com/bojone/dgcnn_for_reading_comprehension
import json, os
from bert4torch.models import build_transformer_model
from bert4torch.tokenizers import Tokenizer, load_vocab
from bert4torch.snippets import sequence_padding, text_segmentate, ListDataset
from bert4torch.generation import AutoRegressiveDecoder
from bert4torch.callbacks import Callback
from bert4torch.losses import CausalLMLoss
from tqdm import tqdm
import torch
from torchinfo import summary
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, Dataset
import numpy as np
# 基本参数
max_p_len = 128
max_q_len = 64
max_a_len = 16
batch_size = 24
epochs = 100
# bert配置
config_path = 'E:/data/pretrain_ckpt/bert/google@chinese_L-12_H-768_A-12/bert4torch_config.json'
checkpoint_path = 'E:/data/pretrain_ckpt/bert/google@chinese_L-12_H-768_A-12/pytorch_model.bin'
dict_path = 'E:/data/pretrain_ckpt/bert/google@chinese_L-12_H-768_A-12/vocab.txt'
device = 'cuda' if torch.cuda.is_available() else 'cpu'
def process_data():
if os.path.exists('F:/data/corpus/qa/CIPS-SOGOU/train_data_list_format.json'):
return
# 标注数据
webqa_data = json.load(open('F:/data/corpus/qa/WebQA.json', encoding='utf-8'))
sogou_data = json.load(open('F:/data/corpus/qa/SogouQA.json', encoding='utf-8'))
# 筛选数据
seps, strips = u'\n。！？!?；;，, ', u'；;，, '
data = []
for d in webqa_data + sogou_data:
for p in d['passages']:
if p['answer']:
for t in text_segmentate(p['passage'], max_p_len - 2, seps, strips):
if p['answer'] in t:
data.append((t, d['question'], p['answer']))
del webqa_data
del sogou_data
# 保存一个随机序（供划分valid用）
random_order = list(range(len(data)))
np.random.seed(2022)
np.random.shuffle(random_order)
# 划分valid
train_data = [data[j] for i, j in enumerate(random_order) if i % 10 != 0]
valid_data = [data[j] for i, j in enumerate(random_order) if i % 10 == 0]
json.dump(train_data, open('F:/data/corpus/qa/CIPS-SOGOU/train_data_list_format.json', 'w'), indent=4)
json.dump(valid_data, open('F:/data/corpus/qa/CIPS-SOGOU/valid_data_list_format.json', 'w'), indent=4)
process_data()
class MyDataset(ListDataset):
@staticmethod
def load_data(file_path):
return json.load(open(file_path))
# 加载并精简词表，建立分词器
token_dict, keep_tokens = load_vocab(
dict_path=dict_path,
simplified=True,
startswith=['[PAD]', '[UNK]', '[CLS]', '[SEP]'],
)
tokenizer = Tokenizer(token_dict, do_lower_case=True)
def collate_fn(batch):
"""单条样本格式：[CLS]篇章[SEP]答案[SEP]问题[SEP]
"""
batch_token_ids, batch_segment_ids = [], []
for (p, q, a) in batch:
p_token_ids, _ = tokenizer.encode(p, maxlen=max_p_len + 1)
a_token_ids, _ = tokenizer.encode(a, maxlen=max_a_len)
q_token_ids, _ = tokenizer.encode(q, maxlen=max_q_len)
token_ids = p_token_ids + a_token_ids[1:] + q_token_ids[1:] # 去掉answer和question的cls位
segment_ids = [0] * len(p_token_ids)
segment_ids += [1] * (len(token_ids) - len(p_token_ids))
batch_token_ids.append(token_ids)
batch_segment_ids.append(segment_ids)
batch_token_ids = torch.tensor(sequence_padding(batch_token_ids), dtype=torch.long, device=device)
batch_segment_ids = torch.tensor(sequence_padding(batch_segment_ids), dtype=torch.long, device=device)
return [batch_token_ids, batch_segment_ids], [batch_token_ids, batch_segment_ids]
train_dataloader = DataLoader(MyDataset('F:/data/corpus/qa/CIPS-SOGOU/train_data_list_format.json'),
batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
valid_dataset = MyDataset('F:/data/corpus/qa/CIPS-SOGOU/valid_data_list_format.json')
valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, collate_fn=collate_fn)
model = build_transformer_model(
config_path,
checkpoint_path,
with_mlm=True,
application='unilm',
keep_tokens=keep_tokens, # 只保留keep_tokens中的字，精简原字表
add_trainer=True
).to(device)
summary(model, input_data=[next(iter(train_dataloader))[0]])
model.compile(loss=CausalLMLoss(offset=True, logits_index=1, ignore_index=0), optimizer=optim.Adam(model.parameters(), 1e-5))
class QuestionAnswerGeneration(AutoRegressiveDecoder):
"""随机生成答案，并且通过beam search来生成问题
"""
@AutoRegressiveDecoder.wraps(default_rtype='logits')
def predict(self, inputs, output_ids, states):
token_ids, segment_ids = inputs
token_ids = torch.cat([token_ids, output_ids], 1)
segment_ids = torch.cat([segment_ids, torch.ones_like(output_ids, device=device)], 1)
_, y_pred = model.predict([token_ids, segment_ids])
return y_pred[:, -1, :]
def generate(self, passage, top_k=1, top_p=0.95):
token_ids, segment_ids = tokenizer.encode(passage, maxlen=max_p_len)
a_ids = self.random_sample([token_ids, segment_ids], n=1, top_p=top_p)[0] # 基于随机采样
token_ids += list(a_ids.cpu().numpy())
segment_ids += [1] * len(a_ids)
q_ids = self.beam_search([token_ids, segment_ids], top_k=top_k)[0] # 基于beam search
return (tokenizer.decode(q_ids.cpu().numpy()), tokenizer.decode(a_ids.cpu().numpy()))
qag = QuestionAnswerGeneration(bos_token_id=None, eos_token_id=tokenizer._token_end_id, max_new_tokens=max_q_len, device=device)
def predict_to_file(data, filename, top_k=1):
"""将预测结果输出到文件，方便评估
"""
with open(filename, 'w', encoding='utf-8') as f:
for d in tqdm(iter(data), desc=u'正在预测(共%s条样本)' % len(data)):
q, a = qag.generate(d[0])
s = '%s\t%s\t%s\n' % (q, a, d[0])
f.write(s)
f.flush()
class Evaluator(Callback):
"""评估与保存
"""
def __init__(self):
self.lowest = 1e10
def on_epoch_end(self, steps, epoch, logs=None):
# 保存最优
if logs['loss'] <= self.lowest:
self.lowest = logs['loss']
# model.save_weights('./best_model.pt')
predict_to_file(valid_dataset.data[:100], 'qa.csv')
if __name__ == '__main__':
evaluator = Evaluator()
model.fit(
train_dataloader,
steps_per_epoch=None,
epochs=epochs,
callbacks=[evaluator]
)
else:
model.load_weights('./best_model.pt')
# predict_to_file(valid_data, 'qa.csv')