-
Notifications
You must be signed in to change notification settings - Fork 13
/
utils.py
171 lines (134 loc) · 7.13 KB
/
utils.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
import os
import torch
import json
import re
from torch.utils.data import DataLoader, Dataset
from itertools import cycle
import numpy as np
from transformers import BertTokenizer, GPT2Tokenizer
def save_hparams(args, path):
with open(path, 'w', encoding='utf-8') as f:
for attr, value in sorted(vars(args).items()):
f.writelines("{}={}\n".format(attr.upper(), value))
class KGDataset(Dataset):
def __init__(self, data_path, max_knowledge=32):
# load data
self._data = []
with open(data_path, 'r', encoding='utf-8') as f:
for line in f.readlines():
self._data.append(json.loads(line))
self._n_data = len(self._data)
self._max_knowledge = max_knowledge
def __len__(self):
return self._n_data
def __getitem__(self, i):
knowledge = self._data[i]['knowledge']
history = self._data[i]['history']
user = self._data[i]['user'] # response: 1, another: 0
response = self._data[i]['response']
if len(knowledge) > self._max_knowledge:
# wizard
keepers = 1 + np.random.choice(len(knowledge) - 1, self._max_knowledge, False)
keepers[0] = 0
knowledge = [knowledge[id] for id in keepers]
return ('\n\n'.join(knowledge), '\n\n'.join(history), np.array(user), response)
def collate_fn(batch):
knowledges = [item[0] for item in batch]
histories = [item[1] for item in batch]
users = [item[2] for item in batch]
responses = [item[3] for item in batch]
knowledge_lens = [len(knowledge.strip().split('\n\n')) for knowledge in knowledges]
max_user = max([u.shape[0] for u in users])
users = [np.pad(u, (0, max_user - u.shape[0]), 'constant', constant_values=-1) for u in users]
return knowledges, histories, users, responses, knowledge_lens
def get_batch_loader(dataset, collate_fn, batch_size=2, num_workers=0, is_test=True):
loader = DataLoader(
dataset, batch_size=batch_size,
shuffle=(not is_test), num_workers=num_workers, collate_fn=collate_fn
)
return loader if is_test else cycle(loader)
class DisBatcher:
def __init__(self, block_size, bert_config, cuda=True):
self.block_size = block_size
self.tokenizer = BertTokenizer.from_pretrained(bert_config, do_lower_case=True)
self.pad_id = self.tokenizer.pad_token_id
self.device = torch.device('cuda' if cuda else 'cpu')
def tokenize(self, text, text_pair=None, max_length=128):
return self.tokenizer.encode(
text, text_pair=text_pair, add_special_tokens=True,
max_length=max_length, pad_to_max_length=True
)
def __call__(self, knowledges, histories, knowledge_lens, n_sent):
knowledge_ids = []
max_knowledge = 0
for know, his in zip(knowledges, histories):
his = ' '.join(his)
ids = [self.tokenize(his, k, self.block_size) for k in know]
knowledge_ids.append(ids)
max_knowledge = max(max_knowledge, len(ids))
padding = [self.pad_id] * self.block_size
knowledge_ids = [ids + [padding] * (max_knowledge - len(ids)) for ids in knowledge_ids]
knowledge_ids = torch.tensor(knowledge_ids, device=self.device, dtype=torch.long)
n_sent = min(n_sent, min(knowledge_lens))
return (knowledge_ids, knowledge_lens, n_sent)
class GenBatcher:
def __init__(self, knowledge_truncate, text_truncate, block_size, gpt2_config, cuda=True):
self.knowledge_truncate = knowledge_truncate
self.text_truncate = text_truncate
self.block_size = block_size
self.tokenizer = GPT2Tokenizer.from_pretrained(gpt2_config)
SPECIAL_TOKENS_DICT = {'additional_special_tokens': ["<user1>", "<user2>", "<knowledge>"]}
self.tokenizer.add_special_tokens(SPECIAL_TOKENS_DICT)
self.eos_id = self.tokenizer.eos_token_id
self.device = torch.device('cuda' if cuda else 'cpu')
# todo
self.user_id = [self.tokenizer.convert_tokens_to_ids('<user1>'), self.tokenizer.convert_tokens_to_ids('<user2>')]
self.know_id = self.tokenizer.convert_tokens_to_ids('<knowledge>')
def tokenize(self, text, text_pair=None):
return self.tokenizer.encode(text, text_pair=text_pair, add_special_tokens=True)
def __call__(self, knowledges, histories, users, responses=None, segment=True, training=True):
if training:
assert responses is not None
input_ids, targets, token_type_ids = [], [], []
for know, his, user, resp in zip(knowledges, histories, users, responses):
knowledge_input = [self.tokenize(k)[:self.knowledge_truncate] for k in know]
knowledge_input = [w for k in knowledge_input for w in k + [self.know_id]][:-1]
knowledge_type = len(knowledge_input) * [self.know_id]
user = [u for u in user.tolist() if u >= 0]
history_input, history_type = [], []
for h, u in zip(his, user):
tmp = [self.user_id[u]] + self.tokenize(h)[:self.text_truncate]
history_input += tmp
history_type += len(tmp) * [self.user_id[u]]
response_input = [self.user_id[1]] + self.tokenize(resp)
response_type = len(response_input) * [self.user_id[1]]
ids = knowledge_input + history_input + response_input
type_ids = knowledge_type + history_type + response_type
tgt = [-1] * (len(knowledge_input) + len(history_input)) + response_input[1:] + [self.eos_id]
ids = ids[-self.block_size:]
type_ids = type_ids[-self.block_size:]
tgt = tgt[-self.block_size:]
ids = ids + [0] * (self.block_size - len(ids))
type_ids = type_ids + [0] * (self.block_size - len(type_ids))
tgt = tgt + [-1] * (self.block_size - len(tgt))
input_ids.append(ids)
token_type_ids.append(type_ids)
targets.append(tgt)
input_ids = torch.tensor(input_ids, device=self.device, dtype=torch.long)
token_type_ids = torch.tensor(token_type_ids, device=self.device, dtype=torch.long)
targets = torch.tensor(targets, device=self.device, dtype=torch.long)
if segment:
return input_ids, token_type_ids, targets
else:
return input_ids, None, targets
else:
assert len(knowledges) == 1 # batch_size == 1
knowledge_input = [self.tokenize(k)[:self.knowledge_truncate] for k in knowledges[0]]
knowledge_input = [w for k in knowledge_input for w in k + [self.know_id]][:-1]
user = [u for u in users[0].tolist() if u >= 0]
history_input = []
for h, u in zip(histories[0], user):
history_input += [self.user_id[u]] + self.tokenize(h)[:self.text_truncate]
input_ids = knowledge_input + history_input + [self.user_id[1]]
input_ids = torch.tensor(input_ids, device=self.device, dtype=torch.long).unsqueeze(0)
return input_ids