-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathdata_utils.py
More file actions
executable file
·285 lines (242 loc) · 11.1 KB
/
data_utils.py
File metadata and controls
executable file
·285 lines (242 loc) · 11.1 KB
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
277
278
279
280
281
282
283
284
285
# Script from https://github.com/mlwu22/RED
import os
import numpy as np
from torch.utils.data import Subset
from datasets import load_from_disk
from datasets import disable_caching
disable_caching()
glue_data_keys_map = {
"cola": ("sentence", None),
"sst2": ("sentence", None),
"mrpc": ("sentence1", "sentence2"),
"stsb": ("sentence1", "sentence2"),
"qqp": ("question1", "question2"),
"mnli": ("premise", "hypothesis"),
"qnli": ("question", "sentence"),
"rte": ("sentence1", "sentence2")
}
glue_data_num_labels_map = {
"cola": 2,
"sst2": 2,
"mrpc": 2,
"stsb": 1,
"qqp": 2,
"mnli": 3,
"qnli": 2,
"rte": 2
}
roberta_base_len_map={
"mnli":256,
"sst2":256,
"mrpc":256,
"cola":256,
"qnli":256,
"qqp":256,
"rte":256,
"stsb":256
}
roberta_large_len_map={
"mnli":256,
"sst2":256,
"mrpc":256,
"cola":256,
"qnli":256,
"qqp":256,
"rte":256,
"stsb":256
}
# load dataset and tokenization
def load_glue_data_final(tokenizer, dataset_name: str, max_seq_length: int = 256, seed=42, max_label_seq_length=5, model_type="roberta_base"):
main_dir = os.path.dirname(os.path.abspath(__file__))
if(model_type=="roberta_base"):
dataset = load_from_disk(
os.path.join(
main_dir,
"data",
dataset_name
))
max_seq_length = roberta_base_len_map[dataset_name]
elif(model_type == "roberta_large"):
dataset = load_from_disk(
os.path.join(
main_dir,
"data",
dataset_name
))
max_seq_length = roberta_large_len_map[dataset_name]
sentence1_key, sentence2_key = glue_data_keys_map[dataset_name]
dataset = dataset.map(lambda examples: tokenization(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
max_label_seq_length=max_label_seq_length,
sentence1_key=sentence1_key,
sentence2_key=sentence2_key,
model_type = model_type,
dataset_name = dataset_name
),
batched=True,
)
if(sentence2_key):
dataset = dataset.remove_columns([sentence1_key, sentence2_key, "idx", "label"])
else:
dataset = dataset.remove_columns([sentence1_key, "idx", "label"])
train_dataset = dataset["train"]
eval_dataset = dataset["validation_matched"] if dataset_name == "mnli" else dataset["validation"]
permuted_indices = np.random.RandomState(seed=seed).permutation(len(eval_dataset)).tolist()
if(dataset_name in ["mnli", "qnli", "qqp"]):
num_eval_data = 1000
elif(dataset_name in ["sst2", "cola", "stsb", "mrpc", "rte"]):
num_eval_data = int(len(eval_dataset)/2)
if(dataset_name in ["cola"]):
test_dataset = Subset(dataset=eval_dataset, indices=permuted_indices[:num_eval_data])
eval_dataset = Subset(dataset=eval_dataset, indices=permuted_indices[num_eval_data:])
else:
test_dataset = Subset(dataset=eval_dataset, indices=permuted_indices[num_eval_data:])
eval_dataset = Subset(dataset=eval_dataset, indices=permuted_indices[:num_eval_data])
num_labels = glue_data_num_labels_map[dataset_name]
return train_dataset, eval_dataset, test_dataset, num_labels
def load_glue_data_t5(tokenizer, dataset_name: str, max_seq_length: int = 256, seed=42, max_label_seq_length=5, model_type="enc_dec"):
main_dir = os.path.dirname(os.path.abspath(__file__))
print(main_dir)
if(model_type=="t5-base"):
dataset = load_from_disk(
os.path.join(
main_dir,
"data/glue_t5_base",
dataset_name
))
max_seq_length=256
sentence1_key, sentence2_key = glue_data_keys_map[dataset_name]
dataset = dataset.map(lambda examples: tokenization(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
max_label_seq_length=max_label_seq_length,
sentence1_key=sentence1_key,
sentence2_key=sentence2_key,
model_type = model_type,
dataset_name = dataset_name
),
batched=True,
)
if(sentence2_key):
dataset = dataset.remove_columns([sentence1_key, sentence2_key, "idx", "label"])
else:
dataset = dataset.remove_columns([sentence1_key, "idx", "label"])
train_dataset = dataset["train"]
permuted_indices = np.random.RandomState(seed=seed).permutation(len(train_dataset)).tolist()
eval_dataset = dataset["validation_matched"] if dataset_name == "mnli" else dataset["validation"]
if(dataset_name in ["mnli", "qnli", "qqp", "sst2"]):
num_eval_data = 1000
test_dataset = eval_dataset
eval_dataset = Subset(dataset=train_dataset, indices=permuted_indices[:num_eval_data])
train_dataset = Subset(dataset=train_dataset, indices=permuted_indices[num_eval_data:])
elif(dataset_name in ["cola", "stsb", "mrpc", "rte"]):
permuted_indices = np.random.RandomState(seed=seed).permutation(len(eval_dataset)).tolist()
num_eval_data = int(len(eval_dataset)/2)
test_dataset = Subset(dataset=eval_dataset, indices=permuted_indices[num_eval_data:])
eval_dataset = Subset(dataset=eval_dataset, indices=permuted_indices[:num_eval_data])
num_labels = glue_data_num_labels_map[dataset_name]
return train_dataset, eval_dataset, test_dataset, num_labels
# detailed function
def tokenization(examples, tokenizer, max_seq_length, sentence1_key, sentence2_key, max_label_seq_length, model_type, dataset_name):
output = tokenizer(
text=examples[sentence1_key],
text_pair=examples[sentence2_key] if sentence2_key else None,
max_length=max_seq_length,
truncation=True
)
input_ids = output.input_ids
attention_mask = output.attention_mask
labels = examples["label"]
if(dataset_name=="stsb" and model_type in ["t5-base", "dec"]):
labels = [round(label, 1) for label in labels]
if(model_type in ["t5-base", "dec"]):
labels = [str(label) for label in labels]
labels = tokenizer(
text=labels,
max_length=max_label_seq_length,
padding=True,
truncation=True
).input_ids
return {
"input_ids" : input_ids,
"attention_mask" : attention_mask,
"labels" : labels
}
def load_e2e_data(tokenizer, max_seq_length: int = 64, max_label_seq_length=100):
main_dir = os.path.dirname(os.path.abspath(__file__))
dataset = load_from_disk(
os.path.join(
main_dir,
"data/e2e_nlg"
)
)
train_dataset = dataset["train"]
eval_dataset = dataset["validation"]
test_dataset = dataset["test"]
train_dataset = train_dataset.map(lambda examples: tokenization_generation(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
max_label_seq_length=max_label_seq_length
),
batched=True,
)
train_dataset = train_dataset.remove_columns(["meaning_representation", "human_reference"])
eval_dataset = eval_dataset.map(lambda examples: tokenization_generation(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
max_label_seq_length=max_label_seq_length
),
batched=True,
)
eval_dataset = eval_dataset.remove_columns(["meaning_representation", "human_reference"])
test_dataset = test_dataset.map(lambda examples: tokenization_generation_test(
examples=examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
max_label_seq_length=max_label_seq_length
),
batched=True,
)
test_dataset = test_dataset.remove_columns(["meaning_representation", "human_reference"])
return train_dataset, eval_dataset, test_dataset
def tokenization_generation(examples, tokenizer, max_seq_length, max_label_seq_length):
max_length = max_seq_length+max_label_seq_length
bs = len(examples["meaning_representation"])
text = examples["meaning_representation"]
text = [t + tokenizer.eos_token for t in text]
test_ids = tokenizer(text = text, max_length=max_seq_length, truncation=True).input_ids
pad_id = tokenizer.pad_token_id
labels = examples["human_reference"]
labels = [l + tokenizer.eos_token for l in labels]
target_ids = tokenizer(
text = labels,
max_length=max_label_seq_length,
truncation=True
).input_ids
input_ids = [test_ids[i] + target_ids[i] for i in range(bs)]
target_ids = [[pad_id for _ in range(len(test_ids[i]))]+ target_ids[i] for i in range(bs)]
target_ids = [target_ids[i] + [pad_id for _ in range(max_length-len(input_ids[i]))] for i in range(bs)]
input_ids = [input_ids[i] + [pad_id for _ in range(max_length-len(input_ids[i]))] for i in range(bs)]
return {
"input_ids" : input_ids,
"labels" : target_ids
}
def tokenization_generation_test(examples, tokenizer, max_seq_length, max_label_seq_length):
text = examples["meaning_representation"]
text = [t + tokenizer.eos_token for t in text]
input_ids = tokenizer(text = text, max_length=max_seq_length, truncation=True).input_ids
labels = examples["human_reference"]
target_ids = tokenizer(
text = labels,
max_length=max_label_seq_length,
truncation=True
).input_ids
return {
"input_ids" : input_ids,
"labels" : target_ids
}