-
Notifications
You must be signed in to change notification settings - Fork 221
/
preprocess_data.py
215 lines (185 loc) · 8.69 KB
/
preprocess_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
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
# coding=utf-8
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Processing data for pretraining."""
import argparse
import json
import multiprocessing
import os
import sys
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__),
os.path.pardir)))
from megatron.data.indexed_dataset import best_fitting_dtype
import time
import torch
try:
import nltk
nltk_available = True
except ImportError:
nltk_available = False
from megatron.tokenizer import build_tokenizer
from megatron.data import indexed_dataset
# https://stackoverflow.com/questions/33139531/preserve-empty-lines-with-nltks-punkt-tokenizer
class CustomLanguageVars(nltk.tokenize.punkt.PunktLanguageVars):
_period_context_fmt = r"""
\S* # some word material
%(SentEndChars)s # a potential sentence ending
\s* # <-- THIS is what I changed
(?=(?P<after_tok>
%(NonWord)s # either other punctuation
|
(?P<next_tok>\S+) # <-- Normally you would have \s+ here
))"""
class IdentitySplitter(object):
def tokenize(self, *text):
return text
class Encoder(object):
def __init__(self, args):
self.args = args
def initializer(self):
# Use Encoder class as a container for global data
Encoder.tokenizer = build_tokenizer(self.args)
if self.args.split_sentences:
if not nltk_available:
print("NLTK is not available to split sentences.")
exit()
splitter = nltk.load("tokenizers/punkt/english.pickle")
if self.args.keep_newlines:
# this prevents punkt from eating newlines after sentences
Encoder.splitter = nltk.tokenize.punkt.PunktSentenceTokenizer(
train_text = splitter._params,
lang_vars = CustomLanguageVars())
else:
Encoder.splitter = splitter
else:
Encoder.splitter = IdentitySplitter()
def encode(self, json_line):
data = json.loads(json_line)
ids = {}
for key in self.args.json_keys:
text = data[key]
doc_ids = []
for sentence in Encoder.splitter.tokenize(text):
sentence_ids = Encoder.tokenizer.tokenize(sentence)
if len(sentence_ids) > 0:
doc_ids.append(sentence_ids)
if len(doc_ids) > 0 and self.args.append_eod:
doc_ids[-1].append(Encoder.tokenizer.eod)
ids[key] = doc_ids
return ids, len(json_line)
def get_args():
parser = argparse.ArgumentParser()
group = parser.add_argument_group(title='input data')
group.add_argument('--input', type=str,
help='Path to input JSON')
group.add_argument('--datasets', nargs='+', default=None,
help='Paths to one or more input datasets to merge')
group.add_argument('--json-keys', nargs='+', default=['text'],
help='space separate listed of keys to extract from json')
group.add_argument('--split-sentences', action='store_true',
help='Split documents into sentences.')
group.add_argument('--keep-newlines', action='store_true',
help='Keep newlines between sentences when splitting.')
group = parser.add_argument_group(title='tokenizer')
group.add_argument('--tokenizer-type', type=str, required=True,
choices=['BertWordPieceLowerCase','BertWordPieceCase',
'GPT2BPETokenizer', 'PretrainedFromHF'],
help='What type of tokenizer to use.')
group.add_argument('--vocab-file', type=str, default=None,
help='Path to the vocab file')
group.add_argument('--merge-file', type=str, default=None,
help='Path to the BPE merge file (if necessary).')
group.add_argument('--append-eod', action='store_true',
help='Append an <eod> token to the end of a document.')
group.add_argument("--tokenizer-name-or-path", type=str, default=None,
help="Name or path of the huggingface tokenizer.")
group.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
help='Pad the vocab size to be divisible by this value.'
'This is added for computational efficieny reasons.')
group.add_argument('--pad-vocab-size-to', type=int, default=None,
help='Pad the vocab size to be divisible by this value.'
'Value of the size of the vocabulary of the tokenizer to reach. This value must be greater than'
' the initial size of the tokenizer. If this argument is used the value of '
'`make-vocab-size-divisible-by` will be ignored.')
group = parser.add_argument_group(title='output data')
group.add_argument('--output-prefix', type=str, required=True,
help='Path to binary output file without suffix')
group.add_argument('--dataset-impl', type=str, default='mmap',
choices=['lazy', 'cached', 'mmap'])
group = parser.add_argument_group(title='runtime')
group.add_argument('--workers', type=int, default=1,
help='Number of worker processes to launch')
group.add_argument('--log-interval', type=int, default=100,
help='Interval between progress updates')
args = parser.parse_args()
args.keep_empty = False
if args.tokenizer_type.lower().startswith('bert'):
if not args.split_sentences:
print("Bert tokenizer detected, are you sure you don't want to split sentences?")
# some default/dummy values for the tokenizer
args.rank = 0
args.tensor_model_parallel_size = 1
args.vocab_extra_ids = 0
return args
def main():
args = get_args()
startup_start = time.time()
print("Opening", args.input)
fin = open(args.input, 'r', encoding='utf-8')
if nltk_available and args.split_sentences:
nltk.download("punkt", quiet=True)
encoder = Encoder(args)
tokenizer = build_tokenizer(args)
pool = multiprocessing.Pool(args.workers, initializer=encoder.initializer)
encoded_docs = pool.imap(encoder.encode, fin, 25)
#encoded_docs = map(encoder.encode, fin)
level = "document"
if args.split_sentences:
level = "sentence"
print(f"Vocab size: {tokenizer.vocab_size}")
print(f"Output prefix: {args.output_prefix}")
output_bin_files = {}
output_idx_files = {}
builders = {}
for key in args.json_keys:
output_bin_files[key] = "{}_{}_{}.bin".format(args.output_prefix,
key, level)
output_idx_files[key] = "{}_{}_{}.idx".format(args.output_prefix,
key, level)
builders[key] = indexed_dataset.make_builder(output_bin_files[key],
impl=args.dataset_impl,
dtype=best_fitting_dtype(tokenizer.vocab_size))
startup_end = time.time()
proc_start = time.time()
total_bytes_processed = 0
print("Time to startup:", startup_end - startup_start)
for i, (doc, bytes_processed) in enumerate(encoded_docs, start=1):
total_bytes_processed += bytes_processed
for key, sentences in doc.items():
if len(sentences) == 0:
continue
for sentence in sentences:
builders[key].add_item(torch.IntTensor(sentence))
builders[key].end_document()
if i % args.log_interval == 0:
current = time.time()
elapsed = current - proc_start
mbs = total_bytes_processed/elapsed/1024/1024
print(f"Processed {i} documents",
f"({i/elapsed} docs/s, {mbs} MB/s).",
file=sys.stderr)
for key in args.json_keys:
builders[key].finalize(output_idx_files[key])
if __name__ == '__main__':
main()