Spaces:
Runtime error
Runtime error
feat: handle streaming
Browse files- dev/seq2seq/do_big_run.sh +6 -2
- dev/seq2seq/do_small_run.sh +6 -3
- dev/seq2seq/run_seq2seq_flax.py +402 -268
dev/seq2seq/do_big_run.sh
CHANGED
|
@@ -1,7 +1,11 @@
|
|
| 1 |
python run_seq2seq_flax.py \
|
| 2 |
--max_source_length 128 \
|
| 3 |
-
--
|
| 4 |
-
--
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
--output_dir output \
|
| 6 |
--per_device_train_batch_size 56 \
|
| 7 |
--per_device_eval_batch_size 56 \
|
|
|
|
| 1 |
python run_seq2seq_flax.py \
|
| 2 |
--max_source_length 128 \
|
| 3 |
+
--dataset_repo_or_path dalle-mini/encoded \
|
| 4 |
+
--train_file **/train/*/*.jsonl \
|
| 5 |
+
--validation_file **/valid/*/*.jsonl \
|
| 6 |
+
--streaming \
|
| 7 |
+
--len_train 1000000 \
|
| 8 |
+
--len_eval 100 \
|
| 9 |
--output_dir output \
|
| 10 |
--per_device_train_batch_size 56 \
|
| 11 |
--per_device_eval_batch_size 56 \
|
dev/seq2seq/do_small_run.sh
CHANGED
|
@@ -1,7 +1,10 @@
|
|
| 1 |
python run_seq2seq_flax.py \
|
| 2 |
-
--
|
| 3 |
-
--train_file
|
| 4 |
-
--validation_file
|
|
|
|
|
|
|
|
|
|
| 5 |
--output_dir output \
|
| 6 |
--per_device_train_batch_size 56 \
|
| 7 |
--per_device_eval_batch_size 56 \
|
|
|
|
| 1 |
python run_seq2seq_flax.py \
|
| 2 |
+
--dataset_repo_or_path dalle-mini/encoded \
|
| 3 |
+
--train_file **/train/*/*.jsonl \
|
| 4 |
+
--validation_file **/valid/*/*.jsonl \
|
| 5 |
+
--streaming \
|
| 6 |
+
--len_train 1000000 \
|
| 7 |
+
--len_eval 1000 \
|
| 8 |
--output_dir output \
|
| 9 |
--per_device_train_batch_size 56 \
|
| 10 |
--per_device_eval_batch_size 56 \
|
dev/seq2seq/run_seq2seq_flax.py
CHANGED
|
@@ -20,9 +20,8 @@ Script adapted from run_summarization_flax.py
|
|
| 20 |
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
|
| 21 |
|
| 22 |
import os
|
| 23 |
-
import logging as pylogging
|
| 24 |
import sys
|
| 25 |
-
import time
|
| 26 |
from dataclasses import dataclass, field
|
| 27 |
from functools import partial
|
| 28 |
from pathlib import Path
|
|
@@ -30,7 +29,6 @@ from typing import Callable, Optional
|
|
| 30 |
import json
|
| 31 |
|
| 32 |
import datasets
|
| 33 |
-
import nltk # Here to have a nice missing dependency error message early on
|
| 34 |
import numpy as np
|
| 35 |
from datasets import Dataset, load_dataset, load_metric
|
| 36 |
from tqdm import tqdm
|
|
@@ -47,9 +45,7 @@ from flax.jax_utils import unreplicate
|
|
| 47 |
from flax.training import train_state
|
| 48 |
from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
|
| 49 |
from transformers import (
|
| 50 |
-
CONFIG_MAPPING,
|
| 51 |
FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
|
| 52 |
-
AutoConfig,
|
| 53 |
AutoTokenizer,
|
| 54 |
FlaxAutoModelForSeq2SeqLM,
|
| 55 |
FlaxBartForConditionalGeneration,
|
|
@@ -61,17 +57,9 @@ from transformers.file_utils import is_offline_mode
|
|
| 61 |
|
| 62 |
import wandb
|
| 63 |
|
| 64 |
-
|
| 65 |
|
| 66 |
-
|
| 67 |
-
nltk.data.find("tokenizers/punkt")
|
| 68 |
-
except (LookupError, OSError):
|
| 69 |
-
if is_offline_mode():
|
| 70 |
-
raise LookupError(
|
| 71 |
-
"Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
|
| 72 |
-
)
|
| 73 |
-
with FileLock(".lock") as lock:
|
| 74 |
-
nltk.download("punkt", quiet=True)
|
| 75 |
|
| 76 |
|
| 77 |
MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.keys())
|
|
@@ -83,7 +71,7 @@ MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
|
|
| 83 |
OUTPUT_VOCAB_SIZE = 16384 + 1 # encoded image token space + 1 for bos
|
| 84 |
OUTPUT_LENGTH = 256 + 1 # number of encoded tokens + 1 for bos
|
| 85 |
BOS_TOKEN_ID = 16384
|
| 86 |
-
BASE_MODEL =
|
| 87 |
|
| 88 |
|
| 89 |
@dataclass
|
|
@@ -101,20 +89,34 @@ class ModelArguments:
|
|
| 101 |
)
|
| 102 |
model_type: Optional[str] = field(
|
| 103 |
default=None,
|
| 104 |
-
metadata={
|
|
|
|
|
|
|
|
|
|
| 105 |
)
|
| 106 |
config_name: Optional[str] = field(
|
| 107 |
-
default=None,
|
|
|
|
|
|
|
|
|
|
| 108 |
)
|
| 109 |
tokenizer_name: Optional[str] = field(
|
| 110 |
-
default=None,
|
|
|
|
|
|
|
|
|
|
| 111 |
)
|
| 112 |
cache_dir: Optional[str] = field(
|
| 113 |
-
default=None,
|
|
|
|
|
|
|
|
|
|
| 114 |
)
|
| 115 |
use_fast_tokenizer: bool = field(
|
| 116 |
default=True,
|
| 117 |
-
metadata={
|
|
|
|
|
|
|
| 118 |
)
|
| 119 |
dtype: Optional[str] = field(
|
| 120 |
default="float32",
|
|
@@ -137,27 +139,51 @@ class DataTrainingArguments:
|
|
| 137 |
"""
|
| 138 |
|
| 139 |
dataset_name: Optional[str] = field(
|
| 140 |
-
default=None,
|
|
|
|
| 141 |
)
|
| 142 |
dataset_config_name: Optional[str] = field(
|
| 143 |
-
default=None,
|
|
|
|
|
|
|
|
|
|
| 144 |
)
|
| 145 |
text_column: Optional[str] = field(
|
| 146 |
-
default=
|
| 147 |
-
metadata={
|
|
|
|
|
|
|
| 148 |
)
|
| 149 |
encoding_column: Optional[str] = field(
|
| 150 |
-
default=
|
| 151 |
-
metadata={
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
)
|
| 153 |
-
train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
|
| 154 |
validation_file: Optional[str] = field(
|
| 155 |
default=None,
|
| 156 |
-
metadata={
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
)
|
| 158 |
-
|
| 159 |
default=None,
|
| 160 |
-
metadata={"help": "
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
)
|
| 162 |
max_source_length: Optional[int] = field(
|
| 163 |
default=128,
|
|
@@ -167,7 +193,8 @@ class DataTrainingArguments:
|
|
| 167 |
},
|
| 168 |
)
|
| 169 |
no_decay: bool = field(
|
| 170 |
-
default=False,
|
|
|
|
| 171 |
)
|
| 172 |
max_target_length: Optional[int] = field(
|
| 173 |
default=OUTPUT_LENGTH,
|
|
@@ -199,60 +226,65 @@ class DataTrainingArguments:
|
|
| 199 |
"value if set."
|
| 200 |
},
|
| 201 |
)
|
| 202 |
-
|
| 203 |
-
default=
|
| 204 |
-
metadata={
|
| 205 |
-
"help": "For debugging purposes or quicker training, truncate the number of prediction examples to this "
|
| 206 |
-
"value if set."
|
| 207 |
-
},
|
| 208 |
)
|
| 209 |
preprocessing_num_workers: Optional[int] = field(
|
| 210 |
-
default=80,
|
| 211 |
metadata={"help": "The number of processes to use for the preprocessing."},
|
| 212 |
)
|
| 213 |
source_prefix: Optional[str] = field(
|
| 214 |
-
default=None, metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
|
| 215 |
-
)
|
| 216 |
-
predict_with_generate: bool = field(
|
| 217 |
-
default=False, metadata={"help": "Whether to use generate to calculate generative metrics."}
|
| 218 |
-
)
|
| 219 |
-
num_beams: Optional[int] = field(
|
| 220 |
default=None,
|
| 221 |
metadata={
|
| 222 |
-
"help": "
|
| 223 |
-
"which is used during evaluation."
|
| 224 |
},
|
| 225 |
)
|
| 226 |
overwrite_cache: bool = field(
|
| 227 |
-
default=False,
|
|
|
|
| 228 |
)
|
| 229 |
log_interval: Optional[int] = field(
|
| 230 |
default=40,
|
| 231 |
-
metadata={
|
| 232 |
-
"help": "For debugging purposes or quicker training, truncate the number of training examples to this "
|
| 233 |
-
"value if set."
|
| 234 |
-
},
|
| 235 |
)
|
| 236 |
log_model: bool = field(
|
| 237 |
-
default=False,
|
|
|
|
| 238 |
)
|
| 239 |
save_model_steps: Optional[int] = field(
|
| 240 |
-
default=3000,
|
| 241 |
metadata={
|
| 242 |
"help": "For logging the model more frequently. Used only when `log_model` is set."
|
| 243 |
},
|
| 244 |
)
|
| 245 |
|
| 246 |
def __post_init__(self):
|
| 247 |
-
if
|
| 248 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
else:
|
| 250 |
if self.train_file is not None:
|
| 251 |
extension = self.train_file.split(".")[-1]
|
| 252 |
-
assert extension in [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 253 |
if self.validation_file is not None:
|
| 254 |
extension = self.validation_file.split(".")[-1]
|
| 255 |
-
assert extension in [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 256 |
if self.val_max_target_length is None:
|
| 257 |
self.val_max_target_length = self.max_target_length
|
| 258 |
|
|
@@ -263,14 +295,20 @@ class TrainState(train_state.TrainState):
|
|
| 263 |
optimizer_step: int
|
| 264 |
|
| 265 |
def replicate(self):
|
| 266 |
-
return jax_utils.replicate(self).replace(
|
|
|
|
|
|
|
| 267 |
|
| 268 |
|
| 269 |
class CustomFlaxBartModule(FlaxBartModule):
|
| 270 |
def setup(self):
|
| 271 |
# check config is valid, otherwise set default values
|
| 272 |
-
self.config.vocab_size_output = getattr(
|
| 273 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 274 |
|
| 275 |
# we keep shared to easily load pre-trained weights
|
| 276 |
self.shared = nn.Embed(
|
|
@@ -286,18 +324,29 @@ class CustomFlaxBartModule(FlaxBartModule):
|
|
| 286 |
embedding_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
| 287 |
dtype=self.dtype,
|
| 288 |
)
|
| 289 |
-
self.encoder = FlaxBartEncoder(
|
|
|
|
|
|
|
| 290 |
|
| 291 |
# the decoder has a different config
|
| 292 |
decoder_config = BartConfig(self.config.to_dict())
|
| 293 |
-
decoder_config.max_position_embeddings =
|
|
|
|
|
|
|
| 294 |
decoder_config.vocab_size = self.config.vocab_size_output
|
| 295 |
-
self.decoder = FlaxBartDecoder(
|
|
|
|
|
|
|
|
|
|
| 296 |
|
| 297 |
-
class CustomFlaxBartForConditionalGenerationModule(
|
|
|
|
|
|
|
| 298 |
def setup(self):
|
| 299 |
# check config is valid, otherwise set default values
|
| 300 |
-
self.config.vocab_size_output = getattr(
|
|
|
|
|
|
|
| 301 |
|
| 302 |
self.model = CustomFlaxBartModule(config=self.config, dtype=self.dtype)
|
| 303 |
self.lm_head = nn.Dense(
|
|
@@ -306,13 +355,18 @@ class CustomFlaxBartForConditionalGenerationModule(FlaxBartForConditionalGenerat
|
|
| 306 |
dtype=self.dtype,
|
| 307 |
kernel_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
| 308 |
)
|
| 309 |
-
self.final_logits_bias = self.param(
|
|
|
|
|
|
|
|
|
|
| 310 |
|
| 311 |
class CustomFlaxBartForConditionalGeneration(FlaxBartForConditionalGeneration):
|
| 312 |
module_class = CustomFlaxBartForConditionalGenerationModule
|
| 313 |
-
|
| 314 |
|
| 315 |
-
|
|
|
|
|
|
|
|
|
|
| 316 |
"""
|
| 317 |
Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
|
| 318 |
Shuffle batches if `shuffle` is `True`.
|
|
@@ -330,33 +384,58 @@ def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuf
|
|
| 330 |
for idx in batch_idx:
|
| 331 |
batch = dataset[idx]
|
| 332 |
batch = {k: jnp.array(v) for k, v in batch.items()}
|
| 333 |
-
|
| 334 |
batch = shard(batch)
|
| 335 |
-
|
| 336 |
yield batch
|
| 337 |
|
| 338 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 339 |
def create_learning_rate_fn(
|
| 340 |
-
train_ds_size: int,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 341 |
) -> Callable[[int], jnp.array]:
|
| 342 |
"""Returns a linear warmup, linear_decay learning rate function."""
|
| 343 |
steps_per_epoch = train_ds_size // train_batch_size
|
| 344 |
num_train_steps = steps_per_epoch * num_train_epochs
|
| 345 |
-
warmup_fn = optax.linear_schedule(
|
|
|
|
|
|
|
| 346 |
if no_decay:
|
| 347 |
return warmup_fn
|
| 348 |
decay_fn = optax.linear_schedule(
|
| 349 |
-
init_value=learning_rate,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 350 |
)
|
| 351 |
-
schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
|
| 352 |
return schedule_fn
|
| 353 |
|
| 354 |
|
| 355 |
def wandb_log(metrics, step=None, prefix=None):
|
| 356 |
if jax.process_index() == 0:
|
| 357 |
-
log_metrics = {
|
|
|
|
|
|
|
|
|
|
| 358 |
if step is not None:
|
| 359 |
-
log_metrics[
|
| 360 |
wandb.log(log_metrics)
|
| 361 |
|
| 362 |
|
|
@@ -365,11 +444,15 @@ def main():
|
|
| 365 |
# or by passing the --help flag to this script.
|
| 366 |
# We now keep distinct sets of args, for a cleaner separation of concerns.
|
| 367 |
|
| 368 |
-
parser = HfArgumentParser(
|
|
|
|
|
|
|
| 369 |
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
| 370 |
# If we pass only one argument to the script and it's the path to a json file,
|
| 371 |
# let's parse it to get our arguments.
|
| 372 |
-
model_args, data_args, training_args = parser.parse_json_file(
|
|
|
|
|
|
|
| 373 |
else:
|
| 374 |
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
| 375 |
|
|
@@ -383,18 +466,18 @@ def main():
|
|
| 383 |
f"Output directory ({training_args.output_dir}) already exists and is not empty."
|
| 384 |
"Use --overwrite_output_dir to overcome."
|
| 385 |
)
|
| 386 |
-
|
| 387 |
# Set up wandb run
|
| 388 |
wandb.init(
|
| 389 |
-
entity=
|
| 390 |
-
project=
|
| 391 |
-
job_type=
|
| 392 |
-
config=parser.parse_args()
|
| 393 |
)
|
| 394 |
|
| 395 |
# set default x-axis as 'train/step'
|
| 396 |
-
wandb.define_metric(
|
| 397 |
-
wandb.define_metric(
|
| 398 |
|
| 399 |
# Make one log on every process with the configuration for debugging.
|
| 400 |
pylogging.basicConfig(
|
|
@@ -418,16 +501,13 @@ def main():
|
|
| 418 |
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
|
| 419 |
# (the dataset will be downloaded automatically from the datasets Hub).
|
| 420 |
#
|
| 421 |
-
data_files = {
|
| 422 |
-
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
dataset = load_dataset("csv", data_files=data_files, cache_dir=model_args.cache_dir, delimiter="\t")
|
| 429 |
-
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
|
| 430 |
-
# https://huggingface.co/docs/datasets/loading_datasets.html.
|
| 431 |
|
| 432 |
# Set up items to load or create
|
| 433 |
tokenizer = None
|
|
@@ -435,17 +515,17 @@ def main():
|
|
| 435 |
|
| 436 |
def restore_state(state, artifact_dir):
|
| 437 |
# restore optimizer state
|
| 438 |
-
with (Path(artifact_dir) /
|
| 439 |
opt_state = from_bytes(state.opt_state, f.read())
|
| 440 |
-
|
| 441 |
# restore steps
|
| 442 |
-
with (Path(artifact_dir) /
|
| 443 |
training_state = json.load(f)
|
| 444 |
-
step = training_state[
|
| 445 |
optimizer_step = step // training_args.gradient_accumulation_steps
|
| 446 |
|
| 447 |
return step, optimizer_step, opt_state
|
| 448 |
-
|
| 449 |
if model_args.from_checkpoint is not None:
|
| 450 |
artifact = wandb.run.use_artifact(model_args.from_checkpoint)
|
| 451 |
artifact_dir = artifact.download()
|
|
@@ -461,40 +541,54 @@ def main():
|
|
| 461 |
config = model.config
|
| 462 |
|
| 463 |
# load tokenizer if present
|
| 464 |
-
if (Path(artifact_dir) /
|
| 465 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 466 |
-
|
| 467 |
-
|
|
|
|
|
|
|
| 468 |
|
| 469 |
else:
|
| 470 |
base_model = FlaxAutoModelForSeq2SeqLM.from_pretrained(
|
| 471 |
-
model_args.model_name_or_path,
|
|
|
|
|
|
|
| 472 |
)
|
| 473 |
# Set up our new model config
|
| 474 |
config = BartConfig.from_pretrained(model_args.model_name_or_path)
|
| 475 |
config.tie_word_embeddings = False
|
| 476 |
config.decoder_start_token_id = BOS_TOKEN_ID # for first token
|
| 477 |
-
config.bos_token_id =
|
| 478 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 479 |
config.eos_token_id = BOS_TOKEN_ID + 1 # unreachable
|
| 480 |
config.forced_bos_token_id = None # we don't need this token
|
| 481 |
config.forced_eos_token_id = None # we don't need this token
|
| 482 |
-
config.force_bos_token_to_be_generated =
|
|
|
|
|
|
|
| 483 |
config.min_length = data_args.max_target_length
|
| 484 |
config.max_length = data_args.max_target_length
|
| 485 |
|
| 486 |
# Create a custom model and initialize it randomly
|
| 487 |
-
model = CustomFlaxBartForConditionalGeneration(
|
|
|
|
|
|
|
| 488 |
|
| 489 |
# Use pre-trained weights for encoder
|
| 490 |
-
model.params[
|
| 491 |
-
model.params[
|
| 492 |
del base_model
|
| 493 |
|
| 494 |
# Load tokenizer if it has not been set
|
| 495 |
if tokenizer is None:
|
| 496 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 497 |
-
model_args.model_name_or_path,
|
|
|
|
|
|
|
| 498 |
)
|
| 499 |
|
| 500 |
print(f"TPUs: {jax.device_count()}")
|
|
@@ -504,23 +598,11 @@ def main():
|
|
| 504 |
|
| 505 |
# Preprocessing the datasets.
|
| 506 |
# We need to tokenize inputs and targets.
|
| 507 |
-
if training_args.do_train:
|
| 508 |
-
column_names = dataset["train"].column_names
|
| 509 |
-
elif training_args.do_eval:
|
| 510 |
-
column_names = dataset["validation"].column_names
|
| 511 |
-
elif training_args.do_predict:
|
| 512 |
-
column_names = dataset["test"].column_names
|
| 513 |
-
else:
|
| 514 |
-
logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
|
| 515 |
-
return
|
| 516 |
|
| 517 |
# Get the column names for input/target.
|
| 518 |
text_column = data_args.text_column
|
| 519 |
encoding_column = data_args.encoding_column
|
| 520 |
|
| 521 |
-
# Temporarily set max_target_length for training.
|
| 522 |
-
max_target_length = data_args.max_target_length
|
| 523 |
-
|
| 524 |
def shift_tokens_right(input_ids: np.array, decoder_start_token_id: int):
|
| 525 |
"""
|
| 526 |
Shift input ids one token to the right.
|
|
@@ -530,18 +612,28 @@ def main():
|
|
| 530 |
shifted_input_ids[:, 0] = decoder_start_token_id
|
| 531 |
return shifted_input_ids
|
| 532 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 533 |
def preprocess_function(examples):
|
| 534 |
inputs = examples[text_column]
|
| 535 |
-
inputs = [prefix + inp for inp in inputs]
|
| 536 |
-
|
| 537 |
model_inputs = tokenizer(
|
| 538 |
-
inputs,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 539 |
)
|
| 540 |
|
| 541 |
# set up targets
|
| 542 |
# Note: labels correspond to our target indices
|
| 543 |
# decoder input ids are the same but shifted to the right with bos at the beginning (and without last token)
|
| 544 |
-
labels =
|
| 545 |
labels = np.asarray(labels)
|
| 546 |
|
| 547 |
# We need the labels, in addition to the decoder_input_ids, for the compute_loss function
|
|
@@ -558,46 +650,75 @@ def main():
|
|
| 558 |
raise ValueError("--do_train requires a train dataset")
|
| 559 |
train_dataset = dataset["train"]
|
| 560 |
if data_args.max_train_samples is not None:
|
| 561 |
-
train_dataset =
|
| 562 |
-
|
| 563 |
-
|
| 564 |
-
|
| 565 |
-
|
| 566 |
-
|
| 567 |
-
|
| 568 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 569 |
)
|
| 570 |
|
| 571 |
if training_args.do_eval:
|
| 572 |
-
max_target_length = data_args.val_max_target_length
|
| 573 |
if "validation" not in dataset:
|
| 574 |
raise ValueError("--do_eval requires a validation dataset")
|
| 575 |
eval_dataset = dataset["validation"]
|
| 576 |
if data_args.max_eval_samples is not None:
|
| 577 |
-
eval_dataset =
|
| 578 |
-
|
| 579 |
-
|
| 580 |
-
|
| 581 |
-
|
| 582 |
-
|
| 583 |
-
|
| 584 |
-
|
| 585 |
-
|
| 586 |
-
|
| 587 |
-
|
| 588 |
-
|
| 589 |
-
|
| 590 |
-
|
| 591 |
-
|
| 592 |
-
|
| 593 |
-
|
| 594 |
-
|
| 595 |
-
|
| 596 |
-
|
| 597 |
-
|
| 598 |
-
|
| 599 |
-
|
| 600 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 601 |
)
|
| 602 |
|
| 603 |
# Initialize our training
|
|
@@ -606,21 +727,40 @@ def main():
|
|
| 606 |
|
| 607 |
# Store some constant
|
| 608 |
num_epochs = int(training_args.num_train_epochs)
|
| 609 |
-
train_batch_size =
|
|
|
|
|
|
|
| 610 |
total_batch_size = int(train_batch_size) * training_args.gradient_accumulation_steps
|
| 611 |
eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
|
| 612 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 613 |
total_steps = steps_per_epoch * num_epochs
|
| 614 |
-
total_optimization_steps = (
|
| 615 |
|
| 616 |
# Create learning rate schedule
|
| 617 |
linear_decay_lr_schedule_fn = create_learning_rate_fn(
|
| 618 |
-
|
| 619 |
total_batch_size,
|
| 620 |
training_args.num_train_epochs,
|
| 621 |
training_args.warmup_steps,
|
| 622 |
training_args.learning_rate,
|
| 623 |
-
data_args.no_decay
|
| 624 |
)
|
| 625 |
|
| 626 |
# We use Optax's "masking" functionality to not apply weight decay
|
|
@@ -633,9 +773,17 @@ def main():
|
|
| 633 |
def decay_mask_fn(params):
|
| 634 |
flat_params = traverse_util.flatten_dict(params)
|
| 635 |
layer_norm_params = [
|
| 636 |
-
(name, "scale")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 637 |
]
|
| 638 |
-
flat_mask = {
|
|
|
|
|
|
|
|
|
|
| 639 |
return traverse_util.unflatten_dict(flat_mask)
|
| 640 |
|
| 641 |
# create adam optimizer
|
|
@@ -667,7 +815,9 @@ def main():
|
|
| 667 |
if model_args.from_checkpoint is not None:
|
| 668 |
# restore optimizer state, step and optimizer_step
|
| 669 |
step, optimizer_step, opt_state = restore_state(state, artifact_dir)
|
| 670 |
-
state = state.replace(
|
|
|
|
|
|
|
| 671 |
|
| 672 |
# label smoothed cross entropy
|
| 673 |
def loss_fn(logits, labels):
|
|
@@ -681,7 +831,9 @@ def main():
|
|
| 681 |
|
| 682 |
def compute_loss(params):
|
| 683 |
labels = batch.pop("labels")
|
| 684 |
-
logits = state.apply_fn(
|
|
|
|
|
|
|
| 685 |
loss = loss_fn(logits, labels)
|
| 686 |
return loss
|
| 687 |
|
|
@@ -690,10 +842,14 @@ def main():
|
|
| 690 |
grad_accum = jax.tree_multimap(lambda x, y: x + y, grads, state.grad_accum)
|
| 691 |
|
| 692 |
def update_fn():
|
| 693 |
-
grads = jax.tree_map(
|
|
|
|
|
|
|
| 694 |
grads = jax.lax.pmean(grads, "batch")
|
| 695 |
new_state = state.apply_gradients(
|
| 696 |
-
grads=grads,
|
|
|
|
|
|
|
| 697 |
)
|
| 698 |
return new_state
|
| 699 |
|
|
@@ -704,7 +860,10 @@ def main():
|
|
| 704 |
None,
|
| 705 |
)
|
| 706 |
|
| 707 |
-
metrics = {
|
|
|
|
|
|
|
|
|
|
| 708 |
metrics = jax.lax.pmean(metrics, axis_name="batch")
|
| 709 |
|
| 710 |
return new_state.replace(dropout_rng=new_dropout_rng), metrics
|
|
@@ -720,39 +879,25 @@ def main():
|
|
| 720 |
metrics = jax.lax.pmean(metrics, axis_name="batch")
|
| 721 |
return metrics
|
| 722 |
|
| 723 |
-
# Define generation function
|
| 724 |
-
max_length = (
|
| 725 |
-
data_args.val_max_target_length if data_args.val_max_target_length is not None else model.config.max_length
|
| 726 |
-
)
|
| 727 |
-
num_beams = data_args.num_beams if data_args.num_beams is not None else model.config.num_beams
|
| 728 |
-
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
|
| 729 |
-
|
| 730 |
-
def generate_step(params, batch):
|
| 731 |
-
model.params = params
|
| 732 |
-
output_ids = model.generate(batch["input_ids"], attention_mask=batch["attention_mask"], **gen_kwargs)
|
| 733 |
-
return output_ids.sequences
|
| 734 |
-
|
| 735 |
# Create parallel version of the train and eval step
|
| 736 |
-
p_train_step = jax.pmap(
|
| 737 |
-
train_step, "batch", donate_argnums=(0,)
|
| 738 |
-
)
|
| 739 |
p_eval_step = jax.pmap(eval_step, "batch")
|
| 740 |
-
p_generate_step = jax.pmap(generate_step, "batch")
|
| 741 |
|
| 742 |
# Replicate the train state on each device
|
| 743 |
state = state.replicate()
|
| 744 |
|
| 745 |
logger.info("***** Running training *****")
|
| 746 |
-
logger.info(f" Num examples = {
|
| 747 |
logger.info(f" Num Epochs = {num_epochs}")
|
| 748 |
-
logger.info(
|
|
|
|
|
|
|
| 749 |
logger.info(
|
| 750 |
f" Total train batch size (w. parallel & distributed) = {train_batch_size * training_args.gradient_accumulation_steps}"
|
| 751 |
)
|
| 752 |
logger.info(f" Total global steps = {total_steps}")
|
| 753 |
logger.info(f" Total optimization steps = {total_optimization_steps}")
|
| 754 |
|
| 755 |
-
train_time = 0
|
| 756 |
epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
|
| 757 |
global_step = 0
|
| 758 |
|
|
@@ -760,31 +905,31 @@ def main():
|
|
| 760 |
# ======================== Evaluating ==============================
|
| 761 |
eval_metrics = []
|
| 762 |
if training_args.do_eval:
|
| 763 |
-
|
| 764 |
-
|
| 765 |
-
|
| 766 |
-
|
| 767 |
-
eval_steps =
|
| 768 |
-
for
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 769 |
# Model forward
|
| 770 |
-
batch = next(eval_loader)
|
| 771 |
-
labels = batch["labels"]
|
| 772 |
-
|
| 773 |
metrics = p_eval_step(state.params, batch)
|
| 774 |
eval_metrics.append(metrics)
|
| 775 |
|
| 776 |
-
# generation
|
| 777 |
-
if data_args.predict_with_generate:
|
| 778 |
-
generated_ids = p_generate_step(state.params, batch)
|
| 779 |
-
eval_preds.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
|
| 780 |
-
eval_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
|
| 781 |
-
|
| 782 |
# normalize eval metrics
|
|
|
|
| 783 |
eval_metrics = get_metrics(eval_metrics)
|
|
|
|
| 784 |
eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
|
|
|
|
| 785 |
|
| 786 |
# log metrics
|
| 787 |
-
wandb_log(eval_metrics, step=global_step, prefix=
|
| 788 |
|
| 789 |
# Print metrics and update progress bar
|
| 790 |
desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']})"
|
|
@@ -808,28 +953,42 @@ def main():
|
|
| 808 |
|
| 809 |
# save state
|
| 810 |
state = unreplicate(state)
|
| 811 |
-
with (Path(training_args.output_dir) /
|
| 812 |
f.write(to_bytes(state.opt_state))
|
| 813 |
-
with (Path(training_args.output_dir) /
|
| 814 |
-
|
|
|
|
|
|
|
| 815 |
|
| 816 |
# save to W&B
|
| 817 |
if data_args.log_model:
|
| 818 |
-
metadata = {
|
| 819 |
if eval_metrics is not None:
|
| 820 |
-
metadata[
|
| 821 |
artifact = wandb.Artifact(
|
| 822 |
name=f"model-{wandb.run.id}", type="bart_model", metadata=metadata
|
| 823 |
)
|
| 824 |
-
artifact.add_file(
|
| 825 |
-
|
| 826 |
-
|
| 827 |
-
artifact.add_file(str(Path(training_args.output_dir) /
|
| 828 |
-
artifact.add_file(
|
| 829 |
-
|
| 830 |
-
|
| 831 |
-
artifact.add_file(
|
| 832 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 833 |
wandb.run.log_artifact(artifact)
|
| 834 |
|
| 835 |
# save some space
|
|
@@ -843,39 +1002,47 @@ def main():
|
|
| 843 |
params=params,
|
| 844 |
push_to_hub=training_args.push_to_hub,
|
| 845 |
commit_message=f"Saving weights and logs of epoch {epoch+1}",
|
| 846 |
-
temp_dir=True # avoid issues with being in a repository
|
| 847 |
)
|
| 848 |
-
|
| 849 |
for epoch in epochs:
|
| 850 |
# ======================== Training ================================
|
| 851 |
-
train_start = time.time()
|
| 852 |
|
| 853 |
# Create sampling rng
|
| 854 |
rng, input_rng = jax.random.split(rng)
|
| 855 |
|
| 856 |
# Generate an epoch by shuffling sampling indices from the train dataset
|
| 857 |
-
|
| 858 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 859 |
# train
|
| 860 |
-
for
|
| 861 |
-
|
| 862 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 863 |
state, train_metric = p_train_step(state, batch)
|
| 864 |
|
| 865 |
if global_step % data_args.log_interval == 0 and jax.process_index() == 0:
|
| 866 |
# log metrics
|
| 867 |
-
wandb_log(unreplicate(train_metric), step=global_step, prefix=
|
| 868 |
|
| 869 |
if training_args.eval_steps and global_step % training_args.eval_steps == 0:
|
| 870 |
run_evaluation()
|
| 871 |
-
|
| 872 |
if global_step % data_args.save_model_steps == 0:
|
| 873 |
run_save_model(state, global_step, epoch)
|
| 874 |
-
|
| 875 |
# log final train metrics
|
| 876 |
-
wandb_log(unreplicate(train_metric), step=global_step, prefix=
|
| 877 |
|
| 878 |
-
train_time += time.time() - train_start
|
| 879 |
train_metric = unreplicate(train_metric)
|
| 880 |
epochs.write(
|
| 881 |
f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
|
|
@@ -888,38 +1055,5 @@ def main():
|
|
| 888 |
run_save_model(state, global_step, epoch, eval_metrics)
|
| 889 |
|
| 890 |
|
| 891 |
-
# ======================== Prediction loop ==============================
|
| 892 |
-
if training_args.do_predict:
|
| 893 |
-
logger.info("*** Predict ***")
|
| 894 |
-
|
| 895 |
-
pred_metrics = []
|
| 896 |
-
pred_generations = []
|
| 897 |
-
pred_labels = []
|
| 898 |
-
|
| 899 |
-
pred_loader = data_loader(input_rng, predict_dataset, eval_batch_size)
|
| 900 |
-
pred_steps = len(predict_dataset) // eval_batch_size
|
| 901 |
-
for _ in tqdm(range(pred_steps), desc="Predicting...", position=2, leave=False):
|
| 902 |
-
# Model forward
|
| 903 |
-
batch = next(pred_loader)
|
| 904 |
-
labels = batch["labels"]
|
| 905 |
-
|
| 906 |
-
metrics = p_eval_step(state.params, batch)
|
| 907 |
-
pred_metrics.append(metrics)
|
| 908 |
-
|
| 909 |
-
# generation
|
| 910 |
-
if data_args.predict_with_generate:
|
| 911 |
-
generated_ids = p_generate_step(state.params, batch)
|
| 912 |
-
pred_generations.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
|
| 913 |
-
pred_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
|
| 914 |
-
|
| 915 |
-
# normalize prediction metrics
|
| 916 |
-
pred_metrics = get_metrics(pred_metrics)
|
| 917 |
-
pred_metrics = jax.tree_map(jnp.mean, pred_metrics)
|
| 918 |
-
|
| 919 |
-
# Print metrics
|
| 920 |
-
desc = f"Predict Loss: {pred_metrics['loss']})"
|
| 921 |
-
logger.info(desc)
|
| 922 |
-
|
| 923 |
-
|
| 924 |
if __name__ == "__main__":
|
| 925 |
main()
|
|
|
|
| 20 |
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
|
| 21 |
|
| 22 |
import os
|
| 23 |
+
import logging as pylogging # To avoid collision with transformers.utils.logging
|
| 24 |
import sys
|
|
|
|
| 25 |
from dataclasses import dataclass, field
|
| 26 |
from functools import partial
|
| 27 |
from pathlib import Path
|
|
|
|
| 29 |
import json
|
| 30 |
|
| 31 |
import datasets
|
|
|
|
| 32 |
import numpy as np
|
| 33 |
from datasets import Dataset, load_dataset, load_metric
|
| 34 |
from tqdm import tqdm
|
|
|
|
| 45 |
from flax.training import train_state
|
| 46 |
from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
|
| 47 |
from transformers import (
|
|
|
|
| 48 |
FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
|
|
|
|
| 49 |
AutoTokenizer,
|
| 50 |
FlaxAutoModelForSeq2SeqLM,
|
| 51 |
FlaxBartForConditionalGeneration,
|
|
|
|
| 57 |
|
| 58 |
import wandb
|
| 59 |
|
| 60 |
+
from dalle_mini.text import TextNormalizer
|
| 61 |
|
| 62 |
+
logger = pylogging.getLogger(__name__)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
|
| 64 |
|
| 65 |
MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.keys())
|
|
|
|
| 71 |
OUTPUT_VOCAB_SIZE = 16384 + 1 # encoded image token space + 1 for bos
|
| 72 |
OUTPUT_LENGTH = 256 + 1 # number of encoded tokens + 1 for bos
|
| 73 |
BOS_TOKEN_ID = 16384
|
| 74 |
+
BASE_MODEL = "facebook/bart-large-cnn" # we currently have issues with bart-large
|
| 75 |
|
| 76 |
|
| 77 |
@dataclass
|
|
|
|
| 89 |
)
|
| 90 |
model_type: Optional[str] = field(
|
| 91 |
default=None,
|
| 92 |
+
metadata={
|
| 93 |
+
"help": "If training from scratch, pass a model type from the list: "
|
| 94 |
+
+ ", ".join(MODEL_TYPES)
|
| 95 |
+
},
|
| 96 |
)
|
| 97 |
config_name: Optional[str] = field(
|
| 98 |
+
default=None,
|
| 99 |
+
metadata={
|
| 100 |
+
"help": "Pretrained config name or path if not the same as model_name"
|
| 101 |
+
},
|
| 102 |
)
|
| 103 |
tokenizer_name: Optional[str] = field(
|
| 104 |
+
default=None,
|
| 105 |
+
metadata={
|
| 106 |
+
"help": "Pretrained tokenizer name or path if not the same as model_name"
|
| 107 |
+
},
|
| 108 |
)
|
| 109 |
cache_dir: Optional[str] = field(
|
| 110 |
+
default=None,
|
| 111 |
+
metadata={
|
| 112 |
+
"help": "Where do you want to store the pretrained models downloaded from s3"
|
| 113 |
+
},
|
| 114 |
)
|
| 115 |
use_fast_tokenizer: bool = field(
|
| 116 |
default=True,
|
| 117 |
+
metadata={
|
| 118 |
+
"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."
|
| 119 |
+
},
|
| 120 |
)
|
| 121 |
dtype: Optional[str] = field(
|
| 122 |
default="float32",
|
|
|
|
| 139 |
"""
|
| 140 |
|
| 141 |
dataset_name: Optional[str] = field(
|
| 142 |
+
default=None,
|
| 143 |
+
metadata={"help": "The name of the dataset to use (via the datasets library)."},
|
| 144 |
)
|
| 145 |
dataset_config_name: Optional[str] = field(
|
| 146 |
+
default=None,
|
| 147 |
+
metadata={
|
| 148 |
+
"help": "The configuration name of the dataset to use (via the datasets library)."
|
| 149 |
+
},
|
| 150 |
)
|
| 151 |
text_column: Optional[str] = field(
|
| 152 |
+
default="caption",
|
| 153 |
+
metadata={
|
| 154 |
+
"help": "The name of the column in the datasets containing the full texts (for summarization)."
|
| 155 |
+
},
|
| 156 |
)
|
| 157 |
encoding_column: Optional[str] = field(
|
| 158 |
+
default="encoding",
|
| 159 |
+
metadata={
|
| 160 |
+
"help": "The name of the column in the datasets containing the image encodings."
|
| 161 |
+
},
|
| 162 |
+
)
|
| 163 |
+
dataset_repo_or_path: Optional[str] = field(
|
| 164 |
+
default=None,
|
| 165 |
+
metadata={"help": "The dataset repository containing encoded files."},
|
| 166 |
+
)
|
| 167 |
+
train_file: Optional[str] = field(
|
| 168 |
+
default=None, metadata={"help": "The input training data file (a text file)."}
|
| 169 |
)
|
|
|
|
| 170 |
validation_file: Optional[str] = field(
|
| 171 |
default=None,
|
| 172 |
+
metadata={
|
| 173 |
+
"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."
|
| 174 |
+
},
|
| 175 |
+
)
|
| 176 |
+
streaming: bool = field(
|
| 177 |
+
default=False,
|
| 178 |
+
metadata={"help": "Whether to stream the dataset."},
|
| 179 |
)
|
| 180 |
+
len_train: Optional[int] = field(
|
| 181 |
default=None,
|
| 182 |
+
metadata={"help": "Length of training dataset, required for streaming"},
|
| 183 |
+
)
|
| 184 |
+
len_eval: Optional[int] = field(
|
| 185 |
+
default=None,
|
| 186 |
+
metadata={"help": "Length of validation dataset, required for streaming"},
|
| 187 |
)
|
| 188 |
max_source_length: Optional[int] = field(
|
| 189 |
default=128,
|
|
|
|
| 193 |
},
|
| 194 |
)
|
| 195 |
no_decay: bool = field(
|
| 196 |
+
default=False,
|
| 197 |
+
metadata={"help": "Whether to use decay in the learning rate scheduler."},
|
| 198 |
)
|
| 199 |
max_target_length: Optional[int] = field(
|
| 200 |
default=OUTPUT_LENGTH,
|
|
|
|
| 226 |
"value if set."
|
| 227 |
},
|
| 228 |
)
|
| 229 |
+
normalize_text: bool = field(
|
| 230 |
+
default=False,
|
| 231 |
+
metadata={"help": "Normalize/Simplify text"},
|
|
|
|
|
|
|
|
|
|
| 232 |
)
|
| 233 |
preprocessing_num_workers: Optional[int] = field(
|
| 234 |
+
default=80, # ensure we have the same datasets cached data and avoid using too much space
|
| 235 |
metadata={"help": "The number of processes to use for the preprocessing."},
|
| 236 |
)
|
| 237 |
source_prefix: Optional[str] = field(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 238 |
default=None,
|
| 239 |
metadata={
|
| 240 |
+
"help": "A prefix to add before every source text (useful for T5 models)."
|
|
|
|
| 241 |
},
|
| 242 |
)
|
| 243 |
overwrite_cache: bool = field(
|
| 244 |
+
default=False,
|
| 245 |
+
metadata={"help": "Overwrite the cached training and evaluation sets"},
|
| 246 |
)
|
| 247 |
log_interval: Optional[int] = field(
|
| 248 |
default=40,
|
| 249 |
+
metadata={"help": "Log frequency for metrics"},
|
|
|
|
|
|
|
|
|
|
| 250 |
)
|
| 251 |
log_model: bool = field(
|
| 252 |
+
default=False,
|
| 253 |
+
metadata={"help": "Overwrite the cached training and evaluation sets"},
|
| 254 |
)
|
| 255 |
save_model_steps: Optional[int] = field(
|
| 256 |
+
default=3000, # about once every hour in our experiments
|
| 257 |
metadata={
|
| 258 |
"help": "For logging the model more frequently. Used only when `log_model` is set."
|
| 259 |
},
|
| 260 |
)
|
| 261 |
|
| 262 |
def __post_init__(self):
|
| 263 |
+
if (
|
| 264 |
+
self.dataset_name is None
|
| 265 |
+
and self.train_file is None
|
| 266 |
+
and self.validation_file is None
|
| 267 |
+
):
|
| 268 |
+
raise ValueError(
|
| 269 |
+
"Need either a dataset name or a training/validation file."
|
| 270 |
+
)
|
| 271 |
else:
|
| 272 |
if self.train_file is not None:
|
| 273 |
extension = self.train_file.split(".")[-1]
|
| 274 |
+
assert extension in [
|
| 275 |
+
"tsv",
|
| 276 |
+
"csv",
|
| 277 |
+
"json",
|
| 278 |
+
"jsonl",
|
| 279 |
+
], "`train_file` should be a tsv, csv or json file."
|
| 280 |
if self.validation_file is not None:
|
| 281 |
extension = self.validation_file.split(".")[-1]
|
| 282 |
+
assert extension in [
|
| 283 |
+
"tsv",
|
| 284 |
+
"csv",
|
| 285 |
+
"json",
|
| 286 |
+
"jsonl",
|
| 287 |
+
], "`validation_file` should be a tsv, csv or json file."
|
| 288 |
if self.val_max_target_length is None:
|
| 289 |
self.val_max_target_length = self.max_target_length
|
| 290 |
|
|
|
|
| 295 |
optimizer_step: int
|
| 296 |
|
| 297 |
def replicate(self):
|
| 298 |
+
return jax_utils.replicate(self).replace(
|
| 299 |
+
dropout_rng=shard_prng_key(self.dropout_rng)
|
| 300 |
+
)
|
| 301 |
|
| 302 |
|
| 303 |
class CustomFlaxBartModule(FlaxBartModule):
|
| 304 |
def setup(self):
|
| 305 |
# check config is valid, otherwise set default values
|
| 306 |
+
self.config.vocab_size_output = getattr(
|
| 307 |
+
self.config, "vocab_size_output", OUTPUT_VOCAB_SIZE
|
| 308 |
+
)
|
| 309 |
+
self.config.max_position_embeddings_decoder = getattr(
|
| 310 |
+
self.config, "max_position_embeddings_decoder", OUTPUT_LENGTH
|
| 311 |
+
)
|
| 312 |
|
| 313 |
# we keep shared to easily load pre-trained weights
|
| 314 |
self.shared = nn.Embed(
|
|
|
|
| 324 |
embedding_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
| 325 |
dtype=self.dtype,
|
| 326 |
)
|
| 327 |
+
self.encoder = FlaxBartEncoder(
|
| 328 |
+
self.config, dtype=self.dtype, embed_tokens=self.shared
|
| 329 |
+
)
|
| 330 |
|
| 331 |
# the decoder has a different config
|
| 332 |
decoder_config = BartConfig(self.config.to_dict())
|
| 333 |
+
decoder_config.max_position_embeddings = (
|
| 334 |
+
self.config.max_position_embeddings_decoder
|
| 335 |
+
)
|
| 336 |
decoder_config.vocab_size = self.config.vocab_size_output
|
| 337 |
+
self.decoder = FlaxBartDecoder(
|
| 338 |
+
decoder_config, dtype=self.dtype, embed_tokens=self.decoder_embed
|
| 339 |
+
)
|
| 340 |
+
|
| 341 |
|
| 342 |
+
class CustomFlaxBartForConditionalGenerationModule(
|
| 343 |
+
FlaxBartForConditionalGenerationModule
|
| 344 |
+
):
|
| 345 |
def setup(self):
|
| 346 |
# check config is valid, otherwise set default values
|
| 347 |
+
self.config.vocab_size_output = getattr(
|
| 348 |
+
self.config, "vocab_size_output", OUTPUT_VOCAB_SIZE
|
| 349 |
+
)
|
| 350 |
|
| 351 |
self.model = CustomFlaxBartModule(config=self.config, dtype=self.dtype)
|
| 352 |
self.lm_head = nn.Dense(
|
|
|
|
| 355 |
dtype=self.dtype,
|
| 356 |
kernel_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),
|
| 357 |
)
|
| 358 |
+
self.final_logits_bias = self.param(
|
| 359 |
+
"final_logits_bias", self.bias_init, (1, self.config.vocab_size_output)
|
| 360 |
+
)
|
| 361 |
+
|
| 362 |
|
| 363 |
class CustomFlaxBartForConditionalGeneration(FlaxBartForConditionalGeneration):
|
| 364 |
module_class = CustomFlaxBartForConditionalGenerationModule
|
|
|
|
| 365 |
|
| 366 |
+
|
| 367 |
+
def data_loader(
|
| 368 |
+
rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False
|
| 369 |
+
):
|
| 370 |
"""
|
| 371 |
Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
|
| 372 |
Shuffle batches if `shuffle` is `True`.
|
|
|
|
| 384 |
for idx in batch_idx:
|
| 385 |
batch = dataset[idx]
|
| 386 |
batch = {k: jnp.array(v) for k, v in batch.items()}
|
|
|
|
| 387 |
batch = shard(batch)
|
|
|
|
| 388 |
yield batch
|
| 389 |
|
| 390 |
|
| 391 |
+
def data_loader_streaming(dataset: Dataset, batch_size: int):
|
| 392 |
+
keys = ["input_ids", "attention_mask", "labels", "decoder_input_ids"]
|
| 393 |
+
batch = {k: [] for k in keys}
|
| 394 |
+
for item in dataset:
|
| 395 |
+
for k, v in item.items():
|
| 396 |
+
batch[k].append(v)
|
| 397 |
+
if len(batch[keys[0]]) == batch_size:
|
| 398 |
+
batch = {k: jnp.array(v) for k, v in batch.items()}
|
| 399 |
+
batch = shard(batch)
|
| 400 |
+
yield batch
|
| 401 |
+
batch = {k: [] for k in keys}
|
| 402 |
+
|
| 403 |
+
|
| 404 |
def create_learning_rate_fn(
|
| 405 |
+
train_ds_size: int,
|
| 406 |
+
train_batch_size: int,
|
| 407 |
+
num_train_epochs: int,
|
| 408 |
+
num_warmup_steps: int,
|
| 409 |
+
learning_rate: float,
|
| 410 |
+
no_decay: bool,
|
| 411 |
) -> Callable[[int], jnp.array]:
|
| 412 |
"""Returns a linear warmup, linear_decay learning rate function."""
|
| 413 |
steps_per_epoch = train_ds_size // train_batch_size
|
| 414 |
num_train_steps = steps_per_epoch * num_train_epochs
|
| 415 |
+
warmup_fn = optax.linear_schedule(
|
| 416 |
+
init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps
|
| 417 |
+
)
|
| 418 |
if no_decay:
|
| 419 |
return warmup_fn
|
| 420 |
decay_fn = optax.linear_schedule(
|
| 421 |
+
init_value=learning_rate,
|
| 422 |
+
end_value=0,
|
| 423 |
+
transition_steps=num_train_steps - num_warmup_steps,
|
| 424 |
+
)
|
| 425 |
+
schedule_fn = optax.join_schedules(
|
| 426 |
+
schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps]
|
| 427 |
)
|
|
|
|
| 428 |
return schedule_fn
|
| 429 |
|
| 430 |
|
| 431 |
def wandb_log(metrics, step=None, prefix=None):
|
| 432 |
if jax.process_index() == 0:
|
| 433 |
+
log_metrics = {
|
| 434 |
+
f"{prefix}/{k}" if prefix is not None else k: jax.device_get(v)
|
| 435 |
+
for k, v in metrics.items()
|
| 436 |
+
}
|
| 437 |
if step is not None:
|
| 438 |
+
log_metrics["train/step"] = step
|
| 439 |
wandb.log(log_metrics)
|
| 440 |
|
| 441 |
|
|
|
|
| 444 |
# or by passing the --help flag to this script.
|
| 445 |
# We now keep distinct sets of args, for a cleaner separation of concerns.
|
| 446 |
|
| 447 |
+
parser = HfArgumentParser(
|
| 448 |
+
(ModelArguments, DataTrainingArguments, TrainingArguments)
|
| 449 |
+
)
|
| 450 |
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
|
| 451 |
# If we pass only one argument to the script and it's the path to a json file,
|
| 452 |
# let's parse it to get our arguments.
|
| 453 |
+
model_args, data_args, training_args = parser.parse_json_file(
|
| 454 |
+
json_file=os.path.abspath(sys.argv[1])
|
| 455 |
+
)
|
| 456 |
else:
|
| 457 |
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
|
| 458 |
|
|
|
|
| 466 |
f"Output directory ({training_args.output_dir}) already exists and is not empty."
|
| 467 |
"Use --overwrite_output_dir to overcome."
|
| 468 |
)
|
| 469 |
+
|
| 470 |
# Set up wandb run
|
| 471 |
wandb.init(
|
| 472 |
+
entity="dalle-mini",
|
| 473 |
+
project="dalle-mini",
|
| 474 |
+
job_type="Seq2Seq",
|
| 475 |
+
config=parser.parse_args(),
|
| 476 |
)
|
| 477 |
|
| 478 |
# set default x-axis as 'train/step'
|
| 479 |
+
wandb.define_metric("train/step")
|
| 480 |
+
wandb.define_metric("*", step_metric="train/step")
|
| 481 |
|
| 482 |
# Make one log on every process with the configuration for debugging.
|
| 483 |
pylogging.basicConfig(
|
|
|
|
| 501 |
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
|
| 502 |
# (the dataset will be downloaded automatically from the datasets Hub).
|
| 503 |
#
|
| 504 |
+
data_files = {
|
| 505 |
+
"train": data_args.train_file,
|
| 506 |
+
"validation": data_args.validation_file,
|
| 507 |
+
}
|
| 508 |
+
dataset = load_dataset(
|
| 509 |
+
data_args.dataset_repo_or_path, data_files=data_files, streaming=True
|
| 510 |
+
)
|
|
|
|
|
|
|
|
|
|
| 511 |
|
| 512 |
# Set up items to load or create
|
| 513 |
tokenizer = None
|
|
|
|
| 515 |
|
| 516 |
def restore_state(state, artifact_dir):
|
| 517 |
# restore optimizer state
|
| 518 |
+
with (Path(artifact_dir) / "opt_state.msgpack").open("rb") as f:
|
| 519 |
opt_state = from_bytes(state.opt_state, f.read())
|
| 520 |
+
|
| 521 |
# restore steps
|
| 522 |
+
with (Path(artifact_dir) / "training_state.json").open("r") as f:
|
| 523 |
training_state = json.load(f)
|
| 524 |
+
step = training_state["step"]
|
| 525 |
optimizer_step = step // training_args.gradient_accumulation_steps
|
| 526 |
|
| 527 |
return step, optimizer_step, opt_state
|
| 528 |
+
|
| 529 |
if model_args.from_checkpoint is not None:
|
| 530 |
artifact = wandb.run.use_artifact(model_args.from_checkpoint)
|
| 531 |
artifact_dir = artifact.download()
|
|
|
|
| 541 |
config = model.config
|
| 542 |
|
| 543 |
# load tokenizer if present
|
| 544 |
+
if (Path(artifact_dir) / "tokenizer_config.json").exists():
|
| 545 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 546 |
+
model_args.model_name_or_path,
|
| 547 |
+
cache_dir=model_args.cache_dir,
|
| 548 |
+
use_fast=model_args.use_fast_tokenizer,
|
| 549 |
+
)
|
| 550 |
|
| 551 |
else:
|
| 552 |
base_model = FlaxAutoModelForSeq2SeqLM.from_pretrained(
|
| 553 |
+
model_args.model_name_or_path,
|
| 554 |
+
seed=training_args.seed,
|
| 555 |
+
dtype=getattr(jnp, model_args.dtype),
|
| 556 |
)
|
| 557 |
# Set up our new model config
|
| 558 |
config = BartConfig.from_pretrained(model_args.model_name_or_path)
|
| 559 |
config.tie_word_embeddings = False
|
| 560 |
config.decoder_start_token_id = BOS_TOKEN_ID # for first token
|
| 561 |
+
config.bos_token_id = (
|
| 562 |
+
BOS_TOKEN_ID # should not be used (due to forced_bos_token_id)
|
| 563 |
+
)
|
| 564 |
+
config.pos_token_id = (
|
| 565 |
+
BOS_TOKEN_ID # should not be needed (as we generate until max_length)
|
| 566 |
+
)
|
| 567 |
config.eos_token_id = BOS_TOKEN_ID + 1 # unreachable
|
| 568 |
config.forced_bos_token_id = None # we don't need this token
|
| 569 |
config.forced_eos_token_id = None # we don't need this token
|
| 570 |
+
config.force_bos_token_to_be_generated = (
|
| 571 |
+
False # otherwise it sets bos_token_id at loading
|
| 572 |
+
)
|
| 573 |
config.min_length = data_args.max_target_length
|
| 574 |
config.max_length = data_args.max_target_length
|
| 575 |
|
| 576 |
# Create a custom model and initialize it randomly
|
| 577 |
+
model = CustomFlaxBartForConditionalGeneration(
|
| 578 |
+
config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
|
| 579 |
+
)
|
| 580 |
|
| 581 |
# Use pre-trained weights for encoder
|
| 582 |
+
model.params["model"]["encoder"] = base_model.params["model"]["encoder"]
|
| 583 |
+
model.params["model"]["shared"] = base_model.params["model"]["shared"]
|
| 584 |
del base_model
|
| 585 |
|
| 586 |
# Load tokenizer if it has not been set
|
| 587 |
if tokenizer is None:
|
| 588 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 589 |
+
model_args.model_name_or_path,
|
| 590 |
+
cache_dir=model_args.cache_dir,
|
| 591 |
+
use_fast=model_args.use_fast_tokenizer,
|
| 592 |
)
|
| 593 |
|
| 594 |
print(f"TPUs: {jax.device_count()}")
|
|
|
|
| 598 |
|
| 599 |
# Preprocessing the datasets.
|
| 600 |
# We need to tokenize inputs and targets.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 601 |
|
| 602 |
# Get the column names for input/target.
|
| 603 |
text_column = data_args.text_column
|
| 604 |
encoding_column = data_args.encoding_column
|
| 605 |
|
|
|
|
|
|
|
|
|
|
| 606 |
def shift_tokens_right(input_ids: np.array, decoder_start_token_id: int):
|
| 607 |
"""
|
| 608 |
Shift input ids one token to the right.
|
|
|
|
| 612 |
shifted_input_ids[:, 0] = decoder_start_token_id
|
| 613 |
return shifted_input_ids
|
| 614 |
|
| 615 |
+
text_normalizer = TextNormalizer() if data_args.normalize_text else None
|
| 616 |
+
|
| 617 |
+
def normalize_text(example):
|
| 618 |
+
example[text_column] = text_normalizer(example[text_column])
|
| 619 |
+
return example
|
| 620 |
+
|
| 621 |
def preprocess_function(examples):
|
| 622 |
inputs = examples[text_column]
|
| 623 |
+
inputs = [prefix + inp for inp in inputs] if prefix else inputs
|
| 624 |
+
# Setting padding="max_length" as we need fixed length inputs for jitted functions
|
| 625 |
model_inputs = tokenizer(
|
| 626 |
+
inputs,
|
| 627 |
+
max_length=data_args.max_source_length,
|
| 628 |
+
padding="max_length",
|
| 629 |
+
truncation=True,
|
| 630 |
+
return_tensors="np",
|
| 631 |
)
|
| 632 |
|
| 633 |
# set up targets
|
| 634 |
# Note: labels correspond to our target indices
|
| 635 |
# decoder input ids are the same but shifted to the right with bos at the beginning (and without last token)
|
| 636 |
+
labels = examples[encoding_column]
|
| 637 |
labels = np.asarray(labels)
|
| 638 |
|
| 639 |
# We need the labels, in addition to the decoder_input_ids, for the compute_loss function
|
|
|
|
| 650 |
raise ValueError("--do_train requires a train dataset")
|
| 651 |
train_dataset = dataset["train"]
|
| 652 |
if data_args.max_train_samples is not None:
|
| 653 |
+
train_dataset = (
|
| 654 |
+
train_dataset.take(data_args.max_train_samples)
|
| 655 |
+
if data_args.streaming
|
| 656 |
+
else train_dataset.select(range(data_args.max_train_samples))
|
| 657 |
+
)
|
| 658 |
+
if data_args.streaming:
|
| 659 |
+
train_dataset = train_dataset.shuffle(1000, training_args.seed)
|
| 660 |
+
if data_args.normalize_text:
|
| 661 |
+
train_dataset = (
|
| 662 |
+
train_dataset.map(text_normalizer)
|
| 663 |
+
if data_args.streaming
|
| 664 |
+
else train_dataset.map(
|
| 665 |
+
normalize_text,
|
| 666 |
+
num_proc=data_args.preprocessing_num_workers,
|
| 667 |
+
load_from_cache_file=not data_args.overwrite_cache,
|
| 668 |
+
desc="Normalizing the validation dataset",
|
| 669 |
+
)
|
| 670 |
+
)
|
| 671 |
+
train_dataset = (
|
| 672 |
+
train_dataset.map(
|
| 673 |
+
preprocess_function,
|
| 674 |
+
batched=True,
|
| 675 |
+
)
|
| 676 |
+
if data_args.streaming
|
| 677 |
+
else train_dataset.map(
|
| 678 |
+
preprocess_function,
|
| 679 |
+
batched=True,
|
| 680 |
+
num_proc=data_args.preprocessing_num_workers,
|
| 681 |
+
remove_columns=train_dataset.column_names,
|
| 682 |
+
load_from_cache_file=not data_args.overwrite_cache,
|
| 683 |
+
desc="Running tokenizer on validation dataset",
|
| 684 |
+
)
|
| 685 |
)
|
| 686 |
|
| 687 |
if training_args.do_eval:
|
|
|
|
| 688 |
if "validation" not in dataset:
|
| 689 |
raise ValueError("--do_eval requires a validation dataset")
|
| 690 |
eval_dataset = dataset["validation"]
|
| 691 |
if data_args.max_eval_samples is not None:
|
| 692 |
+
eval_dataset = (
|
| 693 |
+
eval_dataset.take(data_args.max_train_samples)
|
| 694 |
+
if data_args.streaming
|
| 695 |
+
else eval_dataset.select(range(data_args.max_train_samples))
|
| 696 |
+
)
|
| 697 |
+
if data_args.normalize_text:
|
| 698 |
+
eval_dataset = (
|
| 699 |
+
eval_dataset.map(text_normalizer)
|
| 700 |
+
if data_args.streaming
|
| 701 |
+
else eval_dataset.map(
|
| 702 |
+
normalize_text,
|
| 703 |
+
num_proc=data_args.preprocessing_num_workers,
|
| 704 |
+
load_from_cache_file=not data_args.overwrite_cache,
|
| 705 |
+
desc="Normalizing the validation dataset",
|
| 706 |
+
)
|
| 707 |
+
)
|
| 708 |
+
eval_dataset = (
|
| 709 |
+
eval_dataset.map(
|
| 710 |
+
preprocess_function,
|
| 711 |
+
batched=True,
|
| 712 |
+
)
|
| 713 |
+
if data_args.streaming
|
| 714 |
+
else eval_dataset.map(
|
| 715 |
+
preprocess_function,
|
| 716 |
+
batched=True,
|
| 717 |
+
num_proc=data_args.preprocessing_num_workers,
|
| 718 |
+
remove_columns=eval_dataset.column_names,
|
| 719 |
+
load_from_cache_file=not data_args.overwrite_cache,
|
| 720 |
+
desc="Running tokenizer on validation dataset",
|
| 721 |
+
)
|
| 722 |
)
|
| 723 |
|
| 724 |
# Initialize our training
|
|
|
|
| 727 |
|
| 728 |
# Store some constant
|
| 729 |
num_epochs = int(training_args.num_train_epochs)
|
| 730 |
+
train_batch_size = (
|
| 731 |
+
int(training_args.per_device_train_batch_size) * jax.device_count()
|
| 732 |
+
)
|
| 733 |
total_batch_size = int(train_batch_size) * training_args.gradient_accumulation_steps
|
| 734 |
eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
|
| 735 |
+
if data_args.streaming:
|
| 736 |
+
len_train_dataset = data_args.len_train
|
| 737 |
+
if (
|
| 738 |
+
data_args.max_train_samples is not None
|
| 739 |
+
and data_args.max_train_samples < len_train_dataset
|
| 740 |
+
):
|
| 741 |
+
len_train_dataset = data_args.max_train_samples
|
| 742 |
+
|
| 743 |
+
len_eval_dataset = data_args.len_eval
|
| 744 |
+
if (
|
| 745 |
+
data_args.max_eval_samples is not None
|
| 746 |
+
and data_args.max_eval_samples < len_eval_dataset
|
| 747 |
+
):
|
| 748 |
+
len_eval_dataset = data_args.max_eval_samples
|
| 749 |
+
else:
|
| 750 |
+
len_train_dataset = len(train_dataset)
|
| 751 |
+
len_eval_dataset = len(eval_dataset)
|
| 752 |
+
steps_per_epoch = len_train_dataset // train_batch_size
|
| 753 |
total_steps = steps_per_epoch * num_epochs
|
| 754 |
+
total_optimization_steps = (len_train_dataset // total_batch_size) * num_epochs
|
| 755 |
|
| 756 |
# Create learning rate schedule
|
| 757 |
linear_decay_lr_schedule_fn = create_learning_rate_fn(
|
| 758 |
+
len_train_dataset,
|
| 759 |
total_batch_size,
|
| 760 |
training_args.num_train_epochs,
|
| 761 |
training_args.warmup_steps,
|
| 762 |
training_args.learning_rate,
|
| 763 |
+
data_args.no_decay,
|
| 764 |
)
|
| 765 |
|
| 766 |
# We use Optax's "masking" functionality to not apply weight decay
|
|
|
|
| 773 |
def decay_mask_fn(params):
|
| 774 |
flat_params = traverse_util.flatten_dict(params)
|
| 775 |
layer_norm_params = [
|
| 776 |
+
(name, "scale")
|
| 777 |
+
for name in [
|
| 778 |
+
"self_attn_layer_norm",
|
| 779 |
+
"layernorm_embedding",
|
| 780 |
+
"final_layer_norm",
|
| 781 |
+
]
|
| 782 |
]
|
| 783 |
+
flat_mask = {
|
| 784 |
+
path: (path[-1] != "bias" and path[-2:] not in layer_norm_params)
|
| 785 |
+
for path in flat_params
|
| 786 |
+
}
|
| 787 |
return traverse_util.unflatten_dict(flat_mask)
|
| 788 |
|
| 789 |
# create adam optimizer
|
|
|
|
| 815 |
if model_args.from_checkpoint is not None:
|
| 816 |
# restore optimizer state, step and optimizer_step
|
| 817 |
step, optimizer_step, opt_state = restore_state(state, artifact_dir)
|
| 818 |
+
state = state.replace(
|
| 819 |
+
step=step, optimizer_step=optimizer_step, opt_state=opt_state
|
| 820 |
+
)
|
| 821 |
|
| 822 |
# label smoothed cross entropy
|
| 823 |
def loss_fn(logits, labels):
|
|
|
|
| 831 |
|
| 832 |
def compute_loss(params):
|
| 833 |
labels = batch.pop("labels")
|
| 834 |
+
logits = state.apply_fn(
|
| 835 |
+
**batch, params=params, dropout_rng=dropout_rng, train=True
|
| 836 |
+
)[0]
|
| 837 |
loss = loss_fn(logits, labels)
|
| 838 |
return loss
|
| 839 |
|
|
|
|
| 842 |
grad_accum = jax.tree_multimap(lambda x, y: x + y, grads, state.grad_accum)
|
| 843 |
|
| 844 |
def update_fn():
|
| 845 |
+
grads = jax.tree_map(
|
| 846 |
+
lambda x: x / training_args.gradient_accumulation_steps, grad_accum
|
| 847 |
+
)
|
| 848 |
grads = jax.lax.pmean(grads, "batch")
|
| 849 |
new_state = state.apply_gradients(
|
| 850 |
+
grads=grads,
|
| 851 |
+
grad_accum=jax.tree_map(jnp.zeros_like, grads),
|
| 852 |
+
optimizer_step=state.optimizer_step + 1,
|
| 853 |
)
|
| 854 |
return new_state
|
| 855 |
|
|
|
|
| 860 |
None,
|
| 861 |
)
|
| 862 |
|
| 863 |
+
metrics = {
|
| 864 |
+
"loss": loss,
|
| 865 |
+
"learning_rate": linear_decay_lr_schedule_fn(state.optimizer_step),
|
| 866 |
+
}
|
| 867 |
metrics = jax.lax.pmean(metrics, axis_name="batch")
|
| 868 |
|
| 869 |
return new_state.replace(dropout_rng=new_dropout_rng), metrics
|
|
|
|
| 879 |
metrics = jax.lax.pmean(metrics, axis_name="batch")
|
| 880 |
return metrics
|
| 881 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 882 |
# Create parallel version of the train and eval step
|
| 883 |
+
p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
|
|
|
|
|
|
|
| 884 |
p_eval_step = jax.pmap(eval_step, "batch")
|
|
|
|
| 885 |
|
| 886 |
# Replicate the train state on each device
|
| 887 |
state = state.replicate()
|
| 888 |
|
| 889 |
logger.info("***** Running training *****")
|
| 890 |
+
logger.info(f" Num examples = {len_train_dataset}")
|
| 891 |
logger.info(f" Num Epochs = {num_epochs}")
|
| 892 |
+
logger.info(
|
| 893 |
+
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
|
| 894 |
+
)
|
| 895 |
logger.info(
|
| 896 |
f" Total train batch size (w. parallel & distributed) = {train_batch_size * training_args.gradient_accumulation_steps}"
|
| 897 |
)
|
| 898 |
logger.info(f" Total global steps = {total_steps}")
|
| 899 |
logger.info(f" Total optimization steps = {total_optimization_steps}")
|
| 900 |
|
|
|
|
| 901 |
epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
|
| 902 |
global_step = 0
|
| 903 |
|
|
|
|
| 905 |
# ======================== Evaluating ==============================
|
| 906 |
eval_metrics = []
|
| 907 |
if training_args.do_eval:
|
| 908 |
+
if data_args.streaming:
|
| 909 |
+
eval_loader = data_loader_streaming(eval_dataset, eval_batch_size)
|
| 910 |
+
else:
|
| 911 |
+
eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
|
| 912 |
+
eval_steps = len_eval_dataset // eval_batch_size
|
| 913 |
+
for batch in tqdm(
|
| 914 |
+
eval_loader,
|
| 915 |
+
desc="Evaluating...",
|
| 916 |
+
position=2,
|
| 917 |
+
leave=False,
|
| 918 |
+
total=eval_steps,
|
| 919 |
+
):
|
| 920 |
# Model forward
|
|
|
|
|
|
|
|
|
|
| 921 |
metrics = p_eval_step(state.params, batch)
|
| 922 |
eval_metrics.append(metrics)
|
| 923 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 924 |
# normalize eval metrics
|
| 925 |
+
breakpoint()
|
| 926 |
eval_metrics = get_metrics(eval_metrics)
|
| 927 |
+
breakpoint()
|
| 928 |
eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
|
| 929 |
+
breakpoint()
|
| 930 |
|
| 931 |
# log metrics
|
| 932 |
+
wandb_log(eval_metrics, step=global_step, prefix="eval")
|
| 933 |
|
| 934 |
# Print metrics and update progress bar
|
| 935 |
desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']})"
|
|
|
|
| 953 |
|
| 954 |
# save state
|
| 955 |
state = unreplicate(state)
|
| 956 |
+
with (Path(training_args.output_dir) / "opt_state.msgpack").open("wb") as f:
|
| 957 |
f.write(to_bytes(state.opt_state))
|
| 958 |
+
with (Path(training_args.output_dir) / "training_state.json").open(
|
| 959 |
+
"w"
|
| 960 |
+
) as f:
|
| 961 |
+
json.dump({"step": state.step.item()}, f)
|
| 962 |
|
| 963 |
# save to W&B
|
| 964 |
if data_args.log_model:
|
| 965 |
+
metadata = {"step": step, "epoch": epoch}
|
| 966 |
if eval_metrics is not None:
|
| 967 |
+
metadata["eval/loss"] = eval_metrics["loss"]
|
| 968 |
artifact = wandb.Artifact(
|
| 969 |
name=f"model-{wandb.run.id}", type="bart_model", metadata=metadata
|
| 970 |
)
|
| 971 |
+
artifact.add_file(
|
| 972 |
+
str(Path(training_args.output_dir) / "flax_model.msgpack")
|
| 973 |
+
)
|
| 974 |
+
artifact.add_file(str(Path(training_args.output_dir) / "config.json"))
|
| 975 |
+
artifact.add_file(
|
| 976 |
+
str(Path(training_args.output_dir) / "tokenizer.json")
|
| 977 |
+
)
|
| 978 |
+
artifact.add_file(
|
| 979 |
+
str(Path(training_args.output_dir) / "tokenizer_config.json")
|
| 980 |
+
)
|
| 981 |
+
artifact.add_file(str(Path(training_args.output_dir) / "vocab.json"))
|
| 982 |
+
artifact.add_file(str(Path(training_args.output_dir) / "merges.txt"))
|
| 983 |
+
artifact.add_file(
|
| 984 |
+
str(Path(training_args.output_dir) / "special_tokens_map.json")
|
| 985 |
+
)
|
| 986 |
+
artifact.add_file(
|
| 987 |
+
str(Path(training_args.output_dir) / "opt_state.msgpack")
|
| 988 |
+
)
|
| 989 |
+
artifact.add_file(
|
| 990 |
+
str(Path(training_args.output_dir) / "training_state.json")
|
| 991 |
+
)
|
| 992 |
wandb.run.log_artifact(artifact)
|
| 993 |
|
| 994 |
# save some space
|
|
|
|
| 1002 |
params=params,
|
| 1003 |
push_to_hub=training_args.push_to_hub,
|
| 1004 |
commit_message=f"Saving weights and logs of epoch {epoch+1}",
|
| 1005 |
+
temp_dir=True, # avoid issues with being in a repository
|
| 1006 |
)
|
| 1007 |
+
|
| 1008 |
for epoch in epochs:
|
| 1009 |
# ======================== Training ================================
|
|
|
|
| 1010 |
|
| 1011 |
# Create sampling rng
|
| 1012 |
rng, input_rng = jax.random.split(rng)
|
| 1013 |
|
| 1014 |
# Generate an epoch by shuffling sampling indices from the train dataset
|
| 1015 |
+
if data_args.streaming:
|
| 1016 |
+
train_dataset.set_epoch(epoch)
|
| 1017 |
+
train_loader = data_loader_streaming(train_dataset, train_batch_size)
|
| 1018 |
+
else:
|
| 1019 |
+
train_loader = data_loader(
|
| 1020 |
+
input_rng, train_dataset, train_batch_size, shuffle=True
|
| 1021 |
+
)
|
| 1022 |
# train
|
| 1023 |
+
for batch in tqdm(
|
| 1024 |
+
train_loader,
|
| 1025 |
+
desc="Training...",
|
| 1026 |
+
position=1,
|
| 1027 |
+
leave=False,
|
| 1028 |
+
total=steps_per_epoch,
|
| 1029 |
+
):
|
| 1030 |
+
global_step += 1
|
| 1031 |
state, train_metric = p_train_step(state, batch)
|
| 1032 |
|
| 1033 |
if global_step % data_args.log_interval == 0 and jax.process_index() == 0:
|
| 1034 |
# log metrics
|
| 1035 |
+
wandb_log(unreplicate(train_metric), step=global_step, prefix="train")
|
| 1036 |
|
| 1037 |
if training_args.eval_steps and global_step % training_args.eval_steps == 0:
|
| 1038 |
run_evaluation()
|
| 1039 |
+
|
| 1040 |
if global_step % data_args.save_model_steps == 0:
|
| 1041 |
run_save_model(state, global_step, epoch)
|
| 1042 |
+
|
| 1043 |
# log final train metrics
|
| 1044 |
+
wandb_log(unreplicate(train_metric), step=global_step, prefix="train")
|
| 1045 |
|
|
|
|
| 1046 |
train_metric = unreplicate(train_metric)
|
| 1047 |
epochs.write(
|
| 1048 |
f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
|
|
|
|
| 1055 |
run_save_model(state, global_step, epoch, eval_metrics)
|
| 1056 |
|
| 1057 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1058 |
if __name__ == "__main__":
|
| 1059 |
main()
|