-
Notifications
You must be signed in to change notification settings - Fork 9
/
Datasets.py
275 lines (261 loc) · 15.4 KB
/
Datasets.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
from torch.utils.data import Dataset
import pandas as pd
import json
import random
from datasets import load_dataset
class Pretrain(Dataset):
def __init__(self, tokenizer, type_path, num_samples, input_length, output_length, args, length=None):
self.args = args
print(f'split is {self.args.split}')
self.tokenizer = tokenizer
self.type_path = type_path
self.ssm = False
self.dataset_version = self.args.dataset_version
if 't5' in args.model_name_or_path:
self.model_type='T5'
elif 'gpt2' in args.model_name_or_path:
self.model_type='GPT2'
dataset_v = ['small', 'full']
ids_to_answers = None
if not self.dataset_version in dataset_v:
raise Exception(f'Provided the correct dataset version among {dataset_v}')
# dataset for continual training
if self.args.dataset == 'recentnews':
if type_path=='train':
if self.dataset_version=='small':
self.dataset = pd.read_csv('data/recent_news_small.csv')
elif self.dataset_version=='full':
self.dataset = pd.read_csv('data/recent_news_full.csv')
elif type_path =='split':
if self.args.split==1:
if self.dataset_version=='small':
self.dataset = pd.read_csv('data/split/recent_news_small1.csv')
else:
raise Exception('Not supporting split for full setting.')
elif self.args.split==2:
if self.dataset_version=='small':
self.dataset = pd.read_csv('data/split/recent_news_small2.csv')
else:
raise Exception('Not supporting split for full setting.')
else:
raise Exception('Currently only supporting two splits.')
# for mixreview pretraining corpus
elif type_path =='pretrain':
if self.dataset_version=='small':
total_line = 802776
skip = sorted(random.sample(range(1,total_line+1),total_line-length))
self.dataset = pd.read_csv('data/wikipedia_pretrain_small.csv', usecols=['input', 'output', 'original'], skiprows=skip)
elif self.dataset_version=='full':
total_line = 8021155
skip = sorted(random.sample(range(1,total_line+1),total_line-length))
self.dataset = pd.read_csv('data/wikipedia_pretrain_full.csv', usecols=['input', 'output'], skiprows=skip)
# GPT-2 was initially pretrained on WebText (Dec 2019), which consists of 8 million documents withWikipedia pages excluded.
# In order to measure the performance on INVARIANTLAMA constructed from Wikipedia, we continually pretrain GPT-2 on a subset of Wikipedia (May 2020) for 14k global training stepsbefore CKL.
elif self.args.dataset == 'wikitext103':
self.dataset = pd.read_csv('data/wikipedia_pretrain_1G_final.csv')
# dataset for evaluation
else:
if self.args.dataset == 'invariantlama':
# light tuning 5000 instances for GPT2 experiment
if type_path =='train':
self.dataset = pd.read_csv('data/trex_5000.csv')
else:
self.dataset = pd.read_csv('data/invariantLAMA.csv')
elif self.args.dataset == 'updatedlama':
if self.dataset_version == 'full':
rp_dir = 'data/updatedlama/updatedLAMA.csv'
else:
raise Exception('Not supporting small setting for updatedLAMA.')
self.dataset = pd.read_csv(rp_dir)
with open('data/updatedlama_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset == 'newlama':
if self.dataset_version == 'full':
rp_dir = 'data/newlama/newLAMA.csv'
else:
raise Exception('Not supporting small setting for newLAMA.')
self.dataset = pd.read_csv(rp_dir)
with open('data/recentlama_h_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset == 'newlama_easy' or self.args.dataset == 'newqa_easy':
if self.dataset_version == 'small':
if self.args.split:
if self.args.split==1:
rp_dir = 'data/newlama/newLAMA_easy_small_split1.csv'
else:
rp_dir = 'data/newlama/newLAMA_easy_small_split2.csv'
else:
rp_dir = 'data/newlama/newLAMA_easy_small.csv'
elif self.dataset_version == 'full':
rp_dir = 'data/newlama/newLAMA_easy.csv'
# light tuning 5000 instances for GPT2 experiment
if type_path =='train':
self.dataset = pd.read_csv('data/newlama/newLAMA_easy_5000.csv')
else:
self.dataset = pd.read_csv(rp_dir)
with open('data/recentlama_val_answers.json') as f:
ids_to_answers = json.load(f)
# kilt finetuning + evaluation
elif self.args.dataset== 'TriviaQA':
# Get the KILT task datasets
kilt_triviaqa = load_dataset("kilt_tasks", name="triviaqa_support_only")
# Most tasks in KILT already have all required data, but KILT-TriviaQA only provides the question IDs, not the questions themselves.
# Thankfully, we can get the original TriviaQA data with:
trivia_qa = load_dataset('trivia_qa', 'unfiltered.nocontext')
# The KILT IDs can then be mapped to the TriviaQA questions with:
triviaqa_map = {}
def add_missing_data(x, trivia_qa_subset, triviaqa_map):
i = triviaqa_map[x['id']]
x['input'] = trivia_qa_subset[i]['question']
#x['output']['original_answer'] = trivia_qa_subset[i]['answer']['value']
return x
for k in ['train', 'validation', 'test']:
triviaqa_map = dict([(q_id, i) for i, q_id in enumerate(trivia_qa[k]['question_id'])])
kilt_triviaqa[k] = kilt_triviaqa[k].filter(lambda x: x['id'] in triviaqa_map)
kilt_triviaqa[k] = kilt_triviaqa[k].map(add_missing_data, fn_kwargs=dict(trivia_qa_subset=trivia_qa[k], triviaqa_map=triviaqa_map))
self.dataset = kilt_triviaqa[type_path]
with open('data/tqa_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset== 'fever':
kilt_fever = load_dataset("kilt_tasks", name="fever")
self.dataset = kilt_fever[type_path]
elif self.args.dataset== 'AY2':
kilt_ay2 = load_dataset("kilt_tasks", name='aidayago2')
self.dataset = kilt_ay2[type_path]
elif self.args.dataset== 'WNED':
kilt_wned = load_dataset("kilt_tasks", name="wned")
self.dataset = kilt_wned[type_path]
elif self.args.dataset== 'CWEB':
kilt_cweb = load_dataset("kilt_tasks", name="cweb")
self.dataset = kilt_cweb[type_path]
elif self.args.dataset== 'TREX':
kilt_trex = load_dataset("kilt_tasks", name="trex")
self.dataset = kilt_trex[type_path]
with open('data/trex_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset== 'zsRE':
kilt_zsre = load_dataset("kilt_tasks", name="structured_zeroshot")
self.dataset = kilt_zsre[type_path]
with open('data/zsre_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset== 'NQ':
kilt_nq = load_dataset("kilt_tasks", name="nq")
self.dataset = kilt_nq[type_path]
with open('data/nq_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset== 'HotpotQA':
kilt_hotqa = load_dataset("kilt_tasks", name="hotpotqa")
self.dataset = kilt_hotqa[type_path]
with open('data/hotpotqa_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset== 'ELI5':
kilt_eli5 = load_dataset("kilt_tasks", name="eli5")
self.dataset = kilt_eli5[type_path]
with open('data/eli5_val_answers.json') as f:
ids_to_answers = json.load(f)
elif self.args.dataset== 'WOW':
kilt_wow = load_dataset("kilt_tasks", name="wow", ignore_verifications=True)
self.dataset = kilt_wow[type_path]
else:
raise NameError('Select the correct Dataset!')
print(f'Length of dataset retrieving is.. {len(self.dataset)}')
self.input_length = input_length
self.output_length = output_length
self.ids_to_answers = ids_to_answers
def __len__(self):
return len(self.dataset)
def convert_to_features(self, example_batch, index=None):
# continual pretraining
if self.args.dataset == 'recentnews':
if self.model_type == 'GPT2':
input_ = example_batch['original']
target_= example_batch['original']
elif self.model_type == 'T5':
input_ = example_batch['input']
target_ = example_batch['output']
if type(input_)!=str:
input_=''
if type(target_)!=str:
target_=''
elif self.args.dataset == 'wikitext103':
input_ = example_batch['original']
target_= example_batch['original']
# evaluation
else:
if self.args.dataset == 'invariantlama':
if self.model_type == 'GPT2':
input_pre = example_batch['input']
for index, word in enumerate(input_pre.split()):
if word == '<extra_id_0>':
input_pre = ' '.join(input_pre.split()[:index])
break
if self.type_path == 'train':
input_ = input_pre + ' ' + example_batch['output'] + '.'
target_= input_pre + ' ' + example_batch['output'] + '.'
else:
input_ = input_pre
ground_truth_ = example_batch['output']
target_ = input_pre + ' ' + example_batch['output'] + '.'
elif self.model_type == 'T5':
input_ = example_batch['input']
target_ = example_batch['output']
elif self.args.dataset == 'updatedlama':
input_ = example_batch['statement']
target_ = example_batch['new_answer']
elif self.args.dataset == 'newlama' or self.args.dataset == 'newlama_easy':
input_ = example_batch['statement']
target_ = example_batch['answer']
elif self.args.dataset == 'newqa_easy':
if self.model_type == 'GPT2':
if self.type_path == 'train':
input_ = example_batch['question'] + ' ' + example_batch['answer'].split(';')[0] + '.'
target_ = example_batch['question'] + ' ' + example_batch['answer'].split(';')[0] + '.'
else:
input_ = example_batch['question']
ground_truth_ = example_batch['answer'].split(';')[0] + '.'
target_ = str(example_batch['question']) + ' ' + str(example_batch['answer'])
elif self.model_type == 'T5':
input_ = example_batch['question']
target_ = example_batch['answer'].split(';')[0]
elif (self.args.dataset== 'TriviaQA' or self.args.dataset== 'fever' or self.args.dataset== 'AY2' or self.args.dataset== 'WNED' or self.args.dataset== 'CWEB'
or self.args.dataset== 'TREX' or self.args.dataset== 'zsRE' or self.args.dataset== 'NQ' or self.args.dataset== 'HotpotQA' or self.args.dataset== 'ELI5' or self.args.dataset== 'WOW'):
input_ = example_batch['input']
target_ = example_batch['output'][0]['answer']
else:
raise Exception('Select the correct dataset!')
source = self.tokenizer.batch_encode_plus([str(input_)], max_length=self.input_length,
padding='max_length', truncation=True, return_tensors="pt")
targets = self.tokenizer.batch_encode_plus([str(target_)], max_length=self.output_length,
padding='max_length', truncation=True, return_tensors="pt")
if self.type_path == 'validation' and self.model_type =='GPT2':
ground_truth = self.tokenizer.batch_encode_plus([str(ground_truth_)], max_length=self.output_length,
padding='max_length', truncation=True, return_tensors="pt")
else:
ground_truth = None
if (self.args.dataset == 'invariantlama' or self.args.dataset== 'TriviaQA' or self.args.dataset== 'fever' or self.args.dataset== 'AY2' or self.args.dataset== 'WNED' or self.args.dataset== 'CWEB'
or self.args.dataset== 'TREX' or self.args.dataset== 'zsRE' or self.args.dataset== 'NQ' or self.args.dataset== 'HotpotQA' or self.args.dataset== 'ELI5' or self.args.dataset== 'WOW'):
labels = example_batch['id']
elif (self.args.dataset == 'newlama' or self.args.dataset == 'updatedlama' or self.args.dataset == 'newlama_easy' or self.args.dataset == 'newqa_easy'):
labels = example_batch['unique_id']
else:
labels = None
return source, targets, labels, ground_truth
def __getitem__(self, index):
if (self.args.dataset== 'TriviaQA' or self.args.dataset== 'fever' or self.args.dataset== 'AY2' or self.args.dataset== 'WNED' or self.args.dataset== 'CWEB'
or self.args.dataset== 'TREX' or self.args.dataset== 'zsRE' or self.args.dataset== 'NQ' or self.args.dataset== 'HotpotQA' or self.args.dataset== 'ELI5' or self.args.dataset== 'WOW'):
source, targets, labels, ground_truth = self.convert_to_features(self.dataset[index])
else:
source, targets, labels, ground_truth = self.convert_to_features(self.dataset.iloc[index])
source_ids = source["input_ids"].squeeze()
target_ids = targets["input_ids"].squeeze()
src_mask = source["attention_mask"].squeeze()
target_mask = targets["attention_mask"].squeeze()
if labels is not None:
label_ids = labels
else:
label_ids = -1
if ground_truth is not None:
ground_truth_ids = ground_truth["input_ids"].squeeze()
else:
ground_truth_ids = -1
return {"source_ids": source_ids, "source_mask": src_mask, "target_ids": target_ids, "target_mask": target_mask, "label_ids": label_ids, "ground_truth_ids": ground_truth_ids}