File size: 16,597 Bytes
5a63fc6 0b67ff4 5a63fc6 0b67ff4 5a63fc6 3657027 5a63fc6 3657027 5a63fc6 011e48d 5a63fc6 0b67ff4 5a63fc6 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 |
import dataclasses
import pprint
import time
from functools import partial
import json
import base64
from multiprocessing import Pool
import h5py
import mlxu
from ml_collections.config_dict import config_dict
from ml_collections import ConfigDict
from tqdm import tqdm, trange
import numpy as np
from datasets import load_dataset, load_from_disk
class DatasetFactory(object):
""" Datset builder class. """
@staticmethod
def get_default_config(updates=None):
config = ConfigDict()
config.type = 'huggingface'
config.text_processor = TextProcessor.get_default_config()
config.huggingface_dataset = HuggingfaceDataset.get_default_config()
config.json_dataset = JsonDataset.get_default_config()
if updates is not None:
config.update(ConfigDict(updates).copy_and_resolve_references())
return config
@classmethod
def load_dataset(cls, config, tokenizer, **kwargs):
config = cls.get_default_config(config)
text_processor = TextProcessor(config.text_processor, tokenizer)
if config.type == 'huggingface':
return HuggingfaceDataset(
config.huggingface_dataset, tokenizer, text_processor, **kwargs
)
elif config.type == 'json':
return JsonDataset(config.json_dataset, tokenizer, text_processor, **kwargs)
else:
raise ValueError(f'Unknown dataset type: {config.type}')
def __init__(self):
raise ValueError('DatasetFactory is a static class and should not be instantiated.')
class TextProcessor(object):
""" Example processor that converts a dictionary of texts into tokens. """
@staticmethod
def get_default_config(updates=None):
config = ConfigDict()
config.fields_from_example = ''
config.fields = ''
config.subfield_separator = ' '
config.add_bos_token = True
config.add_eos_token = True
config.prepend_text = ''
config.base64_token_dtype = 'i4'
if updates is not None:
config.update(ConfigDict(updates).copy_and_resolve_references())
return config
def __init__(self, config, tokenizer):
self.config = self.get_default_config(config)
assert self.config.fields != '' or self.config.fields_from_example != '', (
'Either fields or fields_from_example must be specified.'
)
self.tokenizer = tokenizer
def __call__(self, example, has_aux=False):
if has_aux:
example, *aux = example
else:
aux = tuple()
token_buffer = []
loss_mask_buffer = []
if self.config.add_bos_token:
token_buffer.append(self.tokenizer.bos_token_id)
loss_mask_buffer.append(0.0)
if self.config.fields_from_example != '':
fields = example[self.config.fields_from_example].split(',')
else:
fields = self.config.fields.split(',')
for i, field in enumerate(fields):
if field.startswith('[') and field.endswith(']'):
# No loss for this field.
field = field[1:-1]
mask = 0.0
else:
mask = 1.0
if field.startswith('<|') and field.endswith('|>'):
# Special tokens.
field = field[2:-2]
if field == 'bos':
token_buffer.append(self.tokenizer.bos_token_id)
elif field == 'eos':
token_buffer.append(self.tokenizer.eos_token_id)
else:
# Token ID specified directly.
token_buffer.append(int(field))
loss_mask_buffer.append(mask)
elif field.startswith('{') and field.endswith('}'):
field = field[1:-1]
# Base64 encoded raw tokens.
tokens = np.frombuffer(
base64.b64decode(example[field]),
dtype=self.config.base64_token_dtype
).tolist()
token_buffer.extend(tokens)
loss_mask_buffer.extend([mask for _ in range(len(tokens))])
else:
subfields = field.split('+')
text = self.config.subfield_separator.join(
[example[subfield] for subfield in subfields]
)
if i == 0:
text = self.config.prepend_text + text
tokens = self.tokenizer.encode(text)
token_buffer.extend(tokens)
loss_mask_buffer.extend([mask for _ in range(len(tokens))])
if self.config.add_eos_token:
token_buffer.append(self.tokenizer.eos_token_id)
loss_mask_buffer.append(1.0)
return token_buffer, loss_mask_buffer, *aux
class HuggingfaceDataset(object):
""" Huggingface dataset, where the dataset is loaded using the huggingface
datasets.load_dataset() function.
"""
@staticmethod
def get_default_config(updates=None):
config = ConfigDict()
config.path = 'c4'
config.name = 'en'
config.split = 'train'
config.streaming = False
config.seq_length = 1024
config.batch_size = 8
config.always_start_with_bos = False
config.start_seek_loc = 0
config.tokens_count_at_start = 0
config.batch_token_dtype = 'i4'
config.reset_dataset_loc = False
if updates is not None:
config.update(ConfigDict(updates).copy_and_resolve_references())
return config
def __init__(self, config, tokenizer, text_processor, eval_dataset=False):
self.config = self.get_default_config(config)
name = self.config.name if self.config.name != '' else None
split = self.config.split if self.config.split != '' else None
self._tokenizer = tokenizer
self._text_processor = text_processor
self._dataset = load_from_disk(
self.config.path
)[split]
self._dataset = self._dataset.to_iterable_dataset(num_shards=128 if len(self._dataset) > 128 else len(self._dataset))
self._eval_dataset = eval_dataset
self._train_epochs = 0
self._dataset_loc = self.config.start_seek_loc
self._total_tokens = self.config.tokens_count_at_start
self._index = 0
self.reset_dataset_loc = self.config.reset_dataset_loc
def __iter__(self):
if not self._eval_dataset and self._train_epochs > 0:
self._dataset = self._dataset.shuffle(seed=42, buffer_size=10000)
chunk_size = self.config.batch_size * self.config.seq_length
while True:
token_buffer = []
loss_mask_buffer = []
if not self._eval_dataset and self._train_epochs > 0:
self._dataset.set_epoch(self._train_epochs)
for index, example in enumerate(self._dataset):
self._index = index
if not self._eval_dataset and self._dataset_loc > index:
continue
tokens, loss_masks = self.text_processor(example)
token_buffer.extend(tokens)
loss_mask_buffer.extend(loss_masks)
while len(token_buffer) > chunk_size + 1:
self._total_tokens += chunk_size
metrics = {
'dataset_example_index': index,
'dataset_total_tokens': self._total_tokens,
'epoch': self._train_epochs,
}
batch = {
'input_tokens': np.array(token_buffer[:chunk_size], dtype=self.config.batch_token_dtype).reshape(
self.config.batch_size, -1
),
'target_tokens': np.array(token_buffer[1:chunk_size + 1], dtype=self.config.batch_token_dtype).reshape(
self.config.batch_size, -1
),
'loss_masks': np.array(loss_mask_buffer[1:chunk_size + 1], dtype=np.float32).reshape(
self.config.batch_size, -1
),
}
if self.config.always_start_with_bos:
batch['input_tokens'][:, 0] = self.tokenizer.bos_token_id
yield batch, metrics
token_buffer = token_buffer[chunk_size:]
loss_mask_buffer = loss_mask_buffer[chunk_size:]
if self._eval_dataset:
break
else:
if self._train_epochs == 0:
self._dataset = self._dataset.shuffle(seed=42, buffer_size=10000)
self._dataset_loc = 0
self._train_epochs += 1
def get_state_dict(self):
return dict(
config=self.config,
dataset_loc=self._index,
total_tokens=self._total_tokens,
epochs=self._train_epochs,
)
def load_state_dict(self, state_dict):
if 'config' in state_dict:
self.config.update(ConfigDict(state_dict['config']))
self._dataset_loc = state_dict.get('dataset_loc', self.config.start_seek_loc)
self._total_tokens = state_dict.get('total_tokens', self.config.tokens_count_at_start)
self._train_epochs = state_dict.get('epochs', 0)
if self.reset_dataset_loc:
self._dataset_loc = 0
self._train_epochs = 0
@property
def seq_length(self):
return self.config.seq_length
@property
def tokenizer(self):
return self._tokenizer
@property
def text_processor(self):
return self._text_processor
@property
def dataset(self):
return self._dataset
@property
def vocab_size(self):
return len(self._tokenizer)
class JsonDataset(object):
""" JSON dataset, where each line of the data file contains a JSON
dictionary with text fields.
"""
@staticmethod
def get_default_config(updates=None):
config = ConfigDict()
config.path = ''
config.seq_length = 1024
config.batch_size = 8
config.always_start_with_bos = False
config.start_seek_loc = 0
config.example_index_at_start = 0
config.tokens_count_at_start = 0
config.tokenizer_processes = 1
config.tokenizer_parallel_chunk_size = 32
config.tokenizer_parallel_batch_size = 1024
config.throughput_average_window_size = 200
if updates is not None:
config.update(ConfigDict(updates).copy_and_resolve_references())
return config
def __init__(self, config, tokenizer, text_processor):
self.config = self.get_default_config(config)
assert self.config.path != ''
self._tokenizer = tokenizer
self._text_processor = text_processor
self._index = self.config.example_index_at_start
self._file_loc = self.config.start_seek_loc
self._total_tokens = self.config.tokens_count_at_start
def parse_json(self, line):
if not line or line == '\n':
return None
try:
data = json.loads(line)
except json.decoder.JSONDecodeError:
print(f'Error parsing json line:\n{line}')
return None
return data
def json_iterator(self):
with mlxu.open_file(self.config.path, 'r') as fin:
fin.seek(self._file_loc)
while True:
line = fin.readline()
self._file_loc = fin.tell()
if not line: # Reached EOF
self._index = 0
fin.seek(0)
continue
data = self.parse_json(line)
if data is not None:
# JSON parsing succeeded
yield data, self._file_loc, self._index
self._index += 1
def batched(self, iterator, batch_size):
batch = []
for example in iterator:
batch.append(example)
if len(batch) == batch_size:
yield batch
batch = []
if len(batch) > 0:
yield batch
def parallel_example_iterator(self):
if self.config.tokenizer_processes == 1:
for example, loc, index in self.json_iterator():
yield self.text_processor((example, loc, index), has_aux=True)
else:
process_pool = Pool(self.config.tokenizer_processes)
batched_iterator = self.batched(
self.json_iterator(), self.config.tokenizer_parallel_batch_size
)
with process_pool as pool:
map_fn = partial(self.text_processor, has_aux=True)
next_batch = pool.map_async(
map_fn, next(batched_iterator),
chunksize=self.config.tokenizer_parallel_chunk_size
)
while True:
current_batch = next_batch
next_batch = pool.map_async(
map_fn, next(batched_iterator),
chunksize=self.config.tokenizer_parallel_chunk_size
)
for example in current_batch.get():
yield example
def __iter__(self):
chunk_size = self.config.batch_size * self.config.seq_length
token_buffer = []
loss_mask_buffer = []
last_time = 0.0
step_times = []
start_time = time.time()
start_tokens = self._total_tokens
for tokens, loss_masks, loc, index in self.parallel_example_iterator():
token_buffer.extend(tokens)
loss_mask_buffer.extend(loss_masks)
while len(token_buffer) > chunk_size + 1:
self._total_tokens += chunk_size
step_times.append(time.time() - last_time)
last_time = time.time()
if len(step_times) > self.config.throughput_average_window_size:
step_times = step_times[-self.config.throughput_average_window_size:]
average_throughput = chunk_size / np.mean(step_times)
accumulated_throughput = (
(self._total_tokens - start_tokens) / (time.time() - start_time)
)
metrics = {
'dataset_file_loc': loc,
'dataset_example_index': index,
'dataset_total_tokens': self._total_tokens,
'dataset_accumulated_tps': accumulated_throughput,
'dataset_average_tps': average_throughput,
}
batch = {
'input_tokens': np.array(token_buffer[:chunk_size], dtype=np.int32).reshape(
self.config.batch_size, -1
),
'target_tokens': np.array(token_buffer[1:chunk_size + 1], dtype=np.int32).reshape(
self.config.batch_size, -1
),
'loss_masks': np.array(loss_mask_buffer[1:chunk_size + 1], dtype=np.float32).reshape(
self.config.batch_size, -1
),
}
if self.config.always_start_with_bos:
batch['input_tokens'][:, 0] = self.tokenizer.bos_token_id
yield batch, metrics
token_buffer = token_buffer[chunk_size:]
loss_mask_buffer = loss_mask_buffer[chunk_size:]
def get_state_dict(self):
return dict(
config=self.config,
index=self._index,
file_loc=self._file_loc,
total_tokens=self._total_tokens,
)
def load_state_dict(self, state_dict):
if 'config' in state_dict:
self.config.update(ConfigDict(state_dict['config']))
self._index = state_dict.get('index', self.config.example_index_at_start)
self._file_loc = state_dict.get('file_loc', self.config.start_seek_loc)
self._total_tokens = state_dict.get('total_tokens', self.config.tokens_count_at_start)
@property
def seq_length(self):
return self.config.seq_length
@property
def tokenizer(self):
return self._tokenizer
@property
def text_processor(self):
return self._text_processor
@property
def vocab_size(self):
return len(self.tokenizer)
|