-
Notifications
You must be signed in to change notification settings - Fork 202
/
Copy pathconfigure_data.py
180 lines (165 loc) · 8.25 KB
/
configure_data.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
import os
import copy
import data_utils
class DataConfig(object):
def __init__(self, parser, defaults={}):
super(DataConfig,self).__init__()
self.parser = parser
self.defaults = defaults
def apply(self, opt):
print('configuring data')
self.apply_defaults(opt)
return make_loaders(opt)
def set_defaults(self, **kwargs):
for k, v in kwargs.items():
self.defaults[k] = v
def apply_defaults(self, opt):
for k, v in self.defaults.items():
k = k.replace('-', '_')
if not hasattr(opt, k):
setattr(opt, k, v)
def make_loaders(opt):
"""makes training/val/test"""
batch_size = opt.batch_size * opt.world_size
eval_batch_size = opt.eval_batch_size * opt.world_size
seq_length = opt.seq_length
if seq_length < 0:
seq_length = seq_length * opt.world_size
eval_seq_length = opt.eval_seq_length
if opt.eval_seq_length < 0:
eval_seq_length = eval_seq_length * opt.world_size
# data_loader_args = {'num_workers': 0, 'shuffle': opt.shuffle, 'batch_size': batch_size,
data_loader_args = {'num_workers': 4, 'shuffle': opt.shuffle, 'batch_size': batch_size,
# data_loader_args = {'num_workers': 1, 'shuffle': opt.shuffle, 'batch_size': batch_size,
'pin_memory': True, 'transpose': opt.transpose, 'distributed': opt.world_size > 1,
'rank': opt.rank, 'world_size': opt.world_size, 'drop_last': opt.world_size > 1}
if opt.data_set_type == 'L2R':
loader_type = data_utils.ShardLoader
data_loader_args.update({'seq_len': seq_length, 'persist_state': opt.persist_state, 'samples_per_shard': opt.samples_per_shard})
else:
loader_type = data_utils.DataLoader
split = get_split(opt)
data_set_args = {
'path': opt.data, 'seq_length': seq_length, 'lazy': opt.lazy, 'delim': opt.delim,
'text_key': opt.text_key, 'label_key': opt.label_key, 'preprocess': opt.preprocess,
'ds_type': opt.data_set_type, 'split': split, 'loose': opt.loose_json,
'tokenizer_type': opt.tokenizer_type, 'tokenizer_model_path': opt.tokenizer_path,
'vocab_size': opt.vocab_size, 'model_type': opt.tokenizer_model_type,
'non_binary_cols': opt.non_binary_cols, 'process_fn': opt.process_fn}
eval_loader_args = copy.copy(data_loader_args)
eval_set_args = copy.copy(data_set_args)
eval_set_args['split']=[1.]
# if optional eval args were set then replace their equivalent values in the arg dict
if opt.eval_batch_size != 0:
eval_loader_args['batch_size'] = eval_batch_size
if opt.eval_seq_length != 0:
eval_set_args['seq_length'] = eval_seq_length
if opt.data_set_type == 'L2R':
eval_loader_args['seq_len'] = eval_seq_length
if opt.eval_text_key is not None:
eval_set_args['text_key'] = opt.eval_text_key
if opt.eval_label_key is not None:
eval_set_args['label_key'] = opt.eval_label_key
train = None
valid = None
test = None
if opt.data is not None:
train, tokenizer = data_utils.make_dataset(**data_set_args)
if should_split(split):
train, valid, test = train
eval_set_args['tokenizer'] = tokenizer
if opt.valid is not None:
eval_set_args['path'] = opt.valid
valid, _ = data_utils.make_dataset(**eval_set_args)
if test is None and opt.test is not None:
eval_set_args['path'] = opt.test
test, _ = data_utils.make_dataset(**eval_set_args)
if train is not None and opt.batch_size > 0:
train = loader_type(train, **data_loader_args)
if valid is not None:
valid = loader_type(valid, **eval_loader_args)
if test is not None:
test = loader_type(test, **eval_loader_args)
return (train, valid, test), tokenizer
def should_split(split):
return max(split) != 1.
def get_split(opt):
splits = []
if opt.split.find(',') != -1:
splits = [float(s) for s in opt.split.split(',')]
elif opt.split.find('/') != -1:
splits = [float(s) for s in opt.split.split('/')]
else:
splits = [float(opt.split)]
split_total = sum(splits)
if split_total < 1.:
splits.append(1-split_total)
while len(splits) < 3:
splits.append(0.)
splits = splits[:3]
if opt.valid is not None:
splits[1] = 0.
if opt.test is not None:
splits[2] = 0.
final_sum = sum(splits)
return [s/final_sum for s in splits]
def configure_data(parser):
"""add cmdline flags for configuring datasets"""
main_parser = parser
group = parser.add_argument_group('data options')
group.add_argument('--data', nargs='+', default=['./data/imdb/unsup.json'],
help="""Filename for training""")
group.add_argument('--valid', nargs='*', default=None,
help="""Filename for validation""")
group.add_argument('--test', nargs='*', default=None,
help="""Filename for testing""")
group.add_argument('--process-fn', type=str, default='process_str', choices=['process_str', 'process_tweet'],
help='what preprocessing function to use to process text. One of [process_str, process_tweet].')
group.add_argument('--batch-size', type=int, default=128,
help='Data Loader batch size')
group.add_argument('--eval-batch-size', type=int, default=0,
help='Data Loader batch size for evaluation datasets')
group.add_argument('--data-size', type=int, default=256,
help='number of tokens in data')
group.add_argument('--loose-json', action='store_true',
help='Use loose json (one json-formatted string per newline), instead of tight json (data file is one json string)')
group.add_argument('--preprocess', action='store_true',
help='force preprocessing of datasets')
group.add_argument('--delim', default=',',
help='delimiter used to parse csv testfiles')
group.add_argument('--non-binary-cols', nargs='*', default=None,
help='labels for columns to non-binary dataset [only works for csv datasets]')
group.add_argument('--split', default='1.',
help='comma-separated list of proportions for training, validation, and test split')
group.add_argument('--text-key', default='sentence',
help='key to use to extract text from json/csv')
group.add_argument('--label-key', default='label',
help='key to use to extract labels from json/csv')
group.add_argument('--eval-text-key', default=None,
help='key to use to extract text from json/csv evaluation datasets')
group.add_argument('--eval-label-key', default=None,
help='key to use to extract labels from json/csv evaluation datasets')
# tokenizer arguments
group.add_argument('--tokenizer-type', type=str, default='CharacterLevelTokenizer', choices=['CharacterLevelTokenizer', 'SentencePieceTokenizer'],
help='what type of tokenizer to use')
group.add_argument('--tokenizer-model-type', type=str, default='bpe', choices=['bpe', 'char', 'unigram', 'word'],
help='Model type to use for sentencepiece tokenization')
group.add_argument('--vocab-size', type=int, default=256,
help='vocab size to use for non-character-level tokenization')
group.add_argument('--tokenizer-path', type=str, default='tokenizer.model',
help='path used to save/load sentencepiece tokenization models')
# These are options that are relevant to data loading functionality, but are not meant to be exposed to the command line user.
# These options are intneded to be set in code by specific scripts.
defaults = {
'world_size': 1,
'rank': -1,
'persist_state': 0,
'lazy': False,
'shuffle': False,
'transpose': False,
'data_set_type': 'supervised',
'seq_length': 256,
'eval_seq_length': 256,
'samples_per_shard': 100
}
return DataConfig(main_parser, defaults=defaults), group