-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathpreprocess.py
executable file
·309 lines (260 loc) · 11.7 KB
/
preprocess.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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
Pre-process Data / features files and build vocabulary
"""
import codecs
import glob
import sys
import gc
import torch
from functools import partial
from collections import Counter, defaultdict
from onmt.utils.logging import init_logger, logger
from onmt.utils.misc import split_corpus
import onmt.inputters as inputters
import onmt.opts as opts
from onmt.utils.parse import ArgumentParser
from onmt.inputters.inputter import _build_fields_vocab,\
_load_vocab
def check_existing_pt_files(opt):
""" Check if there are existing .pt files to avoid overwriting them """
pattern = opt.save_data + '.{}*.pt'
for t in ['train', 'valid']:
path = pattern.format(t)
if glob.glob(path):
sys.stderr.write("Please backup existing pt files: %s, "
"to avoid overwriting them!\n" % path)
sys.exit(1)
def build_save_dataset(corpus_type, fields, src_reader, tgt_reader, opt):
assert corpus_type in ['train', 'valid']
if corpus_type == 'train':
counters = defaultdict(Counter)
srcs = opt.train_src
tgts = opt.train_tgt
ids = opt.train_ids
if opt.data_type == "graph":
node1s = opt.train_node1
node2s = opt.train_node2
else:
srcs = [opt.valid_src]
tgts = [opt.valid_tgt]
ids = [None]
if opt.data_type == "graph":
node1s = [opt.valid_node1]
node2s = [opt.valid_node2]
if opt.data_type == "graph":
zips = zip(srcs, node1s, node2s, tgts, ids)
else:
zips = zip(srcs, tgts, ids)
for idx, z in enumerate(zips):
if opt.data_type == "graph":
src, node1, node2, tgt, maybe_id = z
logger.info("Reading source, node1, node2 and target files: \n %s \n %s \n %s \n %s." % (src, node1, node2, tgt))
src_shards = split_corpus(src, opt.shard_size)
tgt_shards = split_corpus(tgt, opt.shard_size)
node1_shards = split_corpus(node1, opt.shard_size)
node2_shards = split_corpus(node2, opt.shard_size)
shard_pairs = zip(src_shards, node1_shards, node2_shards, tgt_shards)
else:
src, tgt, maybe_id = z
logger.info("Reading source and target files: %s %s." % (src, tgt))
src_shards = split_corpus(src, opt.shard_size)
tgt_shards = split_corpus(tgt, opt.shard_size)
shard_pairs = zip(src_shards, tgt_shards)
dataset_paths = []
if (corpus_type == "train" or opt.filter_valid) and tgt is not None:
filter_pred = partial(
inputters.filter_example, use_src_len=opt.data_type == "text",
max_src_len=opt.src_seq_length, max_tgt_len=opt.tgt_seq_length)
else:
filter_pred = None
if corpus_type == "train":
existing_fields = None
if opt.src_vocab != "":
try:
logger.info("Using existing vocabulary...")
existing_fields = torch.load(opt.src_vocab)
logger.info('done!! - success')
except torch.serialization.pickle.UnpicklingError:
logger.info("Building vocab from text file...")
src_vocab, src_vocab_size = _load_vocab(
opt.src_vocab, "src", counters,
opt.src_words_min_frequency)
else:
src_vocab = None
if opt.tgt_vocab != "":
tgt_vocab, tgt_vocab_size = _load_vocab(
opt.tgt_vocab, "tgt", counters,
opt.tgt_words_min_frequency)
else:
tgt_vocab = None
if opt.data_type == "graph":
for i, (src_shard, node1_shard, node2_shard, tgt_shard) in enumerate(shard_pairs):
assert len(src_shard) == len(tgt_shard) == \
len(node1_shard) == len(node2_shard)
logger.info("Building shard %d." % i)
dataset = inputters.Dataset(
fields,
readers=([src_reader, src_reader, src_reader, tgt_reader]
if tgt_reader else [src_reader]),
data=([("src", src_shard), ("node1", node1_shard),
("node2", node2_shard), ("tgt", tgt_shard)]
if tgt_reader else [("src", src_shard)]),
dirs=([opt.src_dir, None, None, None]
if tgt_reader else [opt.src_dir]),
sort_key=inputters.str2sortkey[opt.data_type],
filter_pred=filter_pred
)
if corpus_type == "train" and existing_fields is None:
for ex in dataset.examples:
for name, field in fields.items():
try:
f_iter = iter(field)
except TypeError:
f_iter = [(name, field)]
all_data = [getattr(ex, name, None)]
else:
all_data = getattr(ex, name)
for (sub_n, sub_f), fd in zip(
f_iter, all_data):
has_vocab = (sub_n == 'src' and
src_vocab is not None) or \
(sub_n == 'tgt' and
tgt_vocab is not None)
if (hasattr(sub_f, 'sequential')
and sub_f.sequential and not has_vocab):
val = fd
counters[sub_n].update(val)
if maybe_id:
shard_base = corpus_type + "_" + maybe_id
else:
shard_base = corpus_type
data_path = "{:s}.{:s}.{:d}.pt".\
format(opt.save_data, shard_base, i)
dataset_paths.append(data_path)
logger.info(" * saving %sth %s data shard to %s."
% (i, shard_base, data_path))
dataset.save(data_path)
del dataset.examples
gc.collect()
del dataset
gc.collect()
else:
for i, (src_shard, tgt_shard) in enumerate(shard_pairs):
assert len(src_shard) == len(tgt_shard)
logger.info("Building shard %d." % i)
dataset = inputters.Dataset(
fields,
readers=([src_reader, tgt_reader]
if tgt_reader else [src_reader]),
data=([("src", src_shard), ("tgt", tgt_shard)]
if tgt_reader else [("src", src_shard)]),
dirs=([opt.src_dir, None]
if tgt_reader else [opt.src_dir]),
sort_key=inputters.str2sortkey[opt.data_type],
filter_pred=filter_pred
)
if corpus_type == "train" and existing_fields is None:
for ex in dataset.examples:
for name, field in fields.items():
try:
f_iter = iter(field)
except TypeError:
f_iter = [(name, field)]
all_data = [getattr(ex, name, None)]
else:
all_data = getattr(ex, name)
for (sub_n, sub_f), fd in zip(
f_iter, all_data):
has_vocab = (sub_n == 'src' and
src_vocab is not None) or \
(sub_n == 'tgt' and
tgt_vocab is not None)
if (hasattr(sub_f, 'sequential')
and sub_f.sequential and not has_vocab):
val = fd
counters[sub_n].update(val)
if maybe_id:
shard_base = corpus_type + "_" + maybe_id
else:
shard_base = corpus_type
data_path = "{:s}.{:s}.{:d}.pt". \
format(opt.save_data, shard_base, i)
dataset_paths.append(data_path)
logger.info(" * saving %sth %s data shard to %s."
% (i, shard_base, data_path))
dataset.save(data_path)
del dataset.examples
gc.collect()
del dataset
gc.collect()
if corpus_type == "train":
vocab_path = opt.save_data + '.vocab.pt'
if existing_fields is None:
fields = _build_fields_vocab(
fields, counters, opt.data_type,
opt.share_vocab, opt.vocab_size_multiple,
opt.src_vocab_size, opt.src_words_min_frequency,
opt.tgt_vocab_size, opt.tgt_words_min_frequency)
else:
fields = existing_fields
torch.save(fields, vocab_path)
def build_save_vocab(train_dataset, fields, opt):
fields = inputters.build_vocab(
train_dataset, fields, opt.data_type, opt.share_vocab,
opt.src_vocab, opt.src_vocab_size, opt.src_words_min_frequency,
opt.tgt_vocab, opt.tgt_vocab_size, opt.tgt_words_min_frequency,
vocab_size_multiple=opt.vocab_size_multiple
)
vocab_path = opt.save_data + '.vocab.pt'
torch.save(fields, vocab_path)
def count_features(path):
"""
path: location of a corpus file with whitespace-delimited tokens and
￨-delimited features within the token
returns: the number of features in the dataset
"""
with codecs.open(path, "r", "utf-8") as f:
first_tok = f.readline().split(None, 1)[0]
return len(first_tok.split(u"￨")) - 1
def main(opt):
ArgumentParser.validate_preprocess_args(opt)
torch.manual_seed(opt.seed)
if not(opt.overwrite):
check_existing_pt_files(opt)
init_logger(opt.log_file)
logger.info("Extracting features...")
src_nfeats = 0
tgt_nfeats = 0
for src, tgt in zip(opt.train_src, opt.train_tgt):
src_nfeats += count_features(src) if opt.data_type in ('text', 'graph') \
else 0
tgt_nfeats += count_features(tgt) # tgt always text so far
logger.info(" * number of source features: %d." % src_nfeats)
logger.info(" * number of target features: %d." % tgt_nfeats)
logger.info("Building `Fields` object...")
fields = inputters.get_fields(
opt.data_type,
src_nfeats,
tgt_nfeats,
dynamic_dict=opt.dynamic_dict,
src_truncate=opt.src_seq_length_trunc,
tgt_truncate=opt.tgt_seq_length_trunc)
src_reader = inputters.str2reader[opt.data_type].from_opt(opt)
tgt_reader = inputters.str2reader["text"].from_opt(opt)
logger.info("Building & saving training data...")
build_save_dataset(
'train', fields, src_reader, tgt_reader, opt)
if opt.valid_src and opt.valid_tgt:
logger.info("Building & saving validation data...")
build_save_dataset('valid', fields, src_reader, tgt_reader, opt)
def _get_parser():
parser = ArgumentParser(description='preprocess.py')
opts.config_opts(parser)
opts.preprocess_opts(parser)
return parser
if __name__ == "__main__":
parser = _get_parser()
opt = parser.parse_args()
main(opt)