-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathcode_28_CDial.py
189 lines (146 loc) · 7.22 KB
/
code_28_CDial.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
# -*- coding: utf-8 -*-
"""
Created on Sat Dec 5 06:15:51 2020
@author: ljh
"""
from transformers import (GPT2LMHeadModel, GPT2Config,CONFIG_NAME, AdamW, BertTokenizer)
import json
import os
import torch
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data import DataLoader,Dataset
from itertools import chain
from torch.optim.lr_scheduler import LambdaLR
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class DialogDataset(Dataset):
def __init__(self, tokenizer, max_history=15, batch_first=True, lm_labels=True, *inputs, **kwargs):
super(DialogDataset, self).__init__()
self.tokenizer = tokenizer
self.max_history = max_history
self.pad = tokenizer.pad_token_id
self.batch_first = batch_first
self.lm_labels = lm_labels
self.target_file = kwargs['data_path']
SPECIAL_TOKENS = ["[CLS]", "[SEP]", "[speaker1]", "[speaker2]"]
self.bos, self.eos, self.speaker1, self.speaker2 = tokenizer.convert_tokens_to_ids(
SPECIAL_TOKENS)
def _get_line(self, index):
with open(self.target_file, "r", encoding="utf-8") as f:
f.seek(index)
line = f.readline()
return line
def __getitem__(self, index):
tokenizer = self.tokenizer
dialog = self._get_line(index)
dialog = dialog.strip().split("\t")
def tokenize(obj):
if isinstance(obj, str):
return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj))
if isinstance(obj, dict):
return dict((n, tokenize(o)) for n, o in obj.items())
return list(tokenize(o) for o in obj)
dialog = tokenize(dialog)
history = dialog[:-1]
candidates = dialog[-1]
return self.process(history, candidates)
def process(self, history, resposne, with_eos=True):
sequence = [[self.bos]] + history + [resposne + ([self.eos] if with_eos else [])]
sequence = [sequence[0]] + [[self.speaker2 if i % 2 else self.speaker1] + s
for i, s in enumerate(sequence[1:])]
instance = {}
instance["input_ids"] = list(chain(*sequence))
instance["token_type_ids"] = [self.bos] + [self.speaker2 if i % 2 else self.speaker1 for i, s in
enumerate(sequence[1:])
for _ in s]
instance["lm_labels"] = [-1] * len(instance["input_ids"])
if self.lm_labels:
instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:]
return instance
def collate(self, batch):
input_ids = pad_sequence(
[torch.tensor(instance["input_ids"], dtype=torch.long) for instance in batch],
batch_first=self.batch_first, padding_value=self.pad)
token_type_ids = pad_sequence(
[torch.tensor(instance["token_type_ids"], dtype=torch.long) for instance in batch],
batch_first=self.batch_first, padding_value=self.pad)
labels = pad_sequence(
[torch.tensor(instance["lm_labels"], dtype=torch.long) for instance in batch],
batch_first=self.batch_first, padding_value=-1)
return input_ids, token_type_ids, labels
print("Prepare datasets")
model_checkpoint = 'cgpt'
tokenizer = BertTokenizer(os.path.join(model_checkpoint, "vocab.txt"), do_lower_case=True)
valid_path="dataGPT/toy_valid.txt"
valid_dataset = DialogDataset(tokenizer, data_path=valid_path)
input_idssz = tokenizer.convert_ids_to_tokens(valid_dataset[0]['input_ids'])
token_type_idssz = tokenizer.convert_ids_to_tokens(valid_dataset[0]['token_type_ids'])
lm_labelssz = tokenizer.convert_ids_to_tokens(valid_dataset[0]['lm_labels'])
print('输入文本:',''.join(input_idssz))
print('段编码:',''.join(token_type_idssz))
print('标签:',''.join(lm_labelssz))
train_path="dataGPT/toy_train.txt"
train_dataset = DialogDataset(tokenizer, data_path=train_path)
config = GPT2Config.from_json_file(os.path.join(model_checkpoint, 'config.json'))
model = GPT2LMHeadModel(config)
model.to(device)
train_batch_size = 2
valid_batch_size = 2
num_workers = 0
train_loader = DataLoader(train_dataset,
collate_fn=train_dataset.collate,
num_workers=num_workers,
batch_size=train_batch_size,
shuffle=True
)
valid_loader = DataLoader(valid_dataset,
collate_fn=valid_dataset.collate,
num_workers=num_workers,
batch_size=valid_batch_size,
shuffle=False)
lr = 5e-5
optimizer = AdamW([{'params': model.parameters(), 'initial_lr': lr}], lr=lr, correct_bias=True)
lossfun = torch.nn.CrossEntropyLoss(ignore_index=-1)
def evaltest():
model.eval()
allloss = []
for batch in valid_loader:
with torch.no_grad():
input_ids, token_type_ids, lm_labels = tuple(input_tensor.to(device
) for input_tensor in batch)
lm_logits, *_ = model(input_ids, token_type_ids=token_type_ids)
lm_logits_flat_shifted = lm_logits[..., :-1, :].reshape(-1, lm_logits.size(-1))
lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1)
loss = lossfun(lm_logits_flat_shifted, lm_labels_flat_shifted)
allloss.append(lossfun)
return np.mean(allloss)
gradient_accumulation_steps = 16
model_size = 768
warmup_steps = 5000
from_step = -1
max_epochs =80
noam_lambda = lambda step: (
model_size ** (-0.5) * min((step + 1) ** (-0.5), (step + 1) * warmup_steps ** (-1.5)))
noam_scheduler = LambdaLR(optimizer, lr_lambda=noam_lambda, last_epoch=from_step)
logdir='./modelGPT/'
for epoch in range(max_epochs):
model.train()
allloss = []
for iteration,batch in enumerate( train_loader):
input_ids, token_type_ids, lm_labels = tuple(input_tensor.to(device) for input_tensor in batch)
lm_logits, *_ = model(input_ids, token_type_ids=token_type_ids)
lm_logits_flat_shifted = lm_logits[..., :-1, :].reshape(-1, lm_logits.size(-1))
lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1)
lm_loss = lossfun(lm_logits_flat_shifted, lm_labels_flat_shifted)
loss = lm_loss / gradient_accumulation_steps
loss.backward()
allloss.append(loss.item())
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
if iteration % gradient_accumulation_steps == 0:
optimizer.step()
noam_scheduler.step()
optimizer.zero_grad()
evalloss = evaltest()
print('train:',np.mean(allloss), "lr",optimizer.param_groups[0]['lr'],'eval',evalloss)
torch.save({'state_dict': model.cpu().state_dict()}, os.path.join(logdir, 'model_training.bin'))
model.config.to_json_file(os.path.join(logdir, CONFIG_NAME))
tokenizer.save_vocabulary(logdir)