File size: 24,669 Bytes
c4f8e4a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 | from __future__ import annotations
from typing import Union, Optional, List, Any, TYPE_CHECKING
from pathlib import Path
import torch
import math
from copy import copy
from abc import ABC, abstractmethod
if TYPE_CHECKING:
from psi.config.config import LaunchConfig, TrainConfig, LoggingConfig
from psi.config.config import DataConfig
from accelerate import Accelerator
from transformers.trainer_utils import PredictionOutput
from transformers.optimization import get_scheduler
from torch.optim import Optimizer
from psi.utils import snake_to_pascal
import os
import re
import importlib
from torch.utils.data import DataLoader, Dataset
import torch.nn as nn
import numpy as np
import datetime
import accelerate
import random
import shutil
import json
from psi.utils import initialize_overwatch
overwatch = initialize_overwatch(__name__)
def worker_init_fn(worker_id):
# print(f"worker_init_fn called by worker {worker_id}")
worker_seed = torch.initial_seed() % 2**32
random.seed(worker_seed)
np.random.seed(worker_seed)
torch.manual_seed(worker_seed)
class Trainer(ABC):
cfg: LaunchConfig
model: Any
optimizer: torch.optim.Optimizer
lr_scheduler: torch.optim.lr_scheduler.LRScheduler
train_dataset: Dataset
val_dataset: Optional[Dataset]
def __init__(self, cfg: LaunchConfig, device: Union[torch.device, int]):
self.device = torch.device(device)
self.cfg = cfg
# For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
if cfg.train.mixed_precision == "fp16":
self.dtype = torch.float16
elif cfg.train.mixed_precision == "bf16":
self.dtype = torch.bfloat16
else:
self.dtype = torch.float32
# avoid duplicate run names in a real training run
# should read the timestamp from command line args instead
# self.timestamp = datetime.datetime.now().strftime("%y%m%d%H%M")
self.timestamp = self.cfg.timestamp
@classmethod
def instantiate(
# 把对应输入的具体的trainer实例化
cls, cfg: LaunchConfig, device: Union[torch.device, int]
) -> "Trainer":
trainer_name = cfg.train.name # "human-3d"
try:
parts = re.split(r"[-_]", trainer_name) # ["human", "3d"]
trainer_name = "_".join([p.lower() for p in parts]) # "human_3d"
module = importlib.import_module(f"psi.trainers.{trainer_name}")
trainer_clazz = getattr(module, f"{snake_to_pascal(trainer_name)}Trainer") # 帕斯卡命名法
except Exception as e:
raise ValueError(
f"fail to import {trainer_name} from psi.trainers"
) from e
return trainer_clazz(cfg, device)
@property
def train_cfg(self) -> TrainConfig:
return self.cfg.train
@property
def log_cfg(self) -> LoggingConfig:
return self.cfg.log
@property
def data_cfg(self) -> DataConfig:
return self.cfg.data
@property
def hf_token(self):
return self.cfg.train.hf_token.read_text().strip() \
if isinstance(self.cfg.train.hf_token, Path) \
else os.environ[self.cfg.train.hf_token]
def get_fsdp_plugin(self) -> accelerate.utils.FullyShardedDataParallelPlugin | None: ...
def create_datasets(self) -> tuple[Dataset, Dataset|None]:
...
def create_dataloaders(
self, train_dataset, val_dataset
) -> tuple[DataLoader, DataLoader|None]: ...
def create_optimizer(self):
optimizer_kwargs = dict(self.cfg.train.lr_scheduler_kwargs)
if self.cfg.train.optimizer_foreach is not None:
optimizer_kwargs["foreach"] = self.cfg.train.optimizer_foreach
self.optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=self.cfg.train.learning_rate,
**optimizer_kwargs, # type: ignore
)
def create_scheduler(
self, num_training_steps: int | None = None, optimizer: Optimizer | None = None
):
if num_training_steps is None:
num_training_steps = self.max_training_steps
# Creates Dummy Scheduler if `scheduler` was specified in the config file else creates `args.lr_scheduler_type` Scheduler
if (
self.accelerator.state.deepspeed_plugin is None
or "scheduler" not in self.accelerator.state.deepspeed_plugin.deepspeed_config
):
self.lr_scheduler = get_scheduler(
name=self.cfg.train.lr_scheduler_type,
optimizer=optimizer if optimizer is not None else self.optimizer,
num_warmup_steps=self.num_warmup_steps * self.world_size,
num_training_steps=num_training_steps * self.world_size,
scheduler_specific_kwargs=self.cfg.train.scheduler_specific_kwargs,
)
else:
self.lr_scheduler = accelerate.utils.DummyScheduler(
optimizer, total_num_steps=self.max_training_steps, warmup_num_steps=self.num_warmup_steps
) # type: ignore
return self.lr_scheduler
def log(
self, metrics: dict[str, Any], start_time: Optional[float] = None
) -> None:
wandb_dict = copy(metrics)
for k,v in list(wandb_dict.items()):
if isinstance(v, torch.Tensor) or isinstance(v, np.ndarray):
wandb_dict[k] = v.item()
# wandb_dict.update({"train/grad_norm": grad_norm})
# wandb_dict.update(self.get_log_kv())
self.accelerator.log(wandb_dict, step=self.global_step) # WANDB logging
def create_optimizer_and_scheduler(self, num_training_steps: int | None = None):
optimizer = self.create_optimizer()
self.create_scheduler(
num_training_steps=num_training_steps, optimizer=optimizer
)
def compute_loss(
self, batch#, return_outputs=False#, num_items_in_batch=None
) -> dict: ...
def training_step(
self,
# model: nn.Module,
batch: dict[str, Union[torch.Tensor, Any]],
) -> tuple[bool, dict[str, Any]]: ...
def prediction_step(
self,
model: nn.Module,
inputs: dict[str, Union[torch.Tensor, Any]],
prediction_loss_only: bool,
ignore_keys: Optional[list[str]] = None,
) -> tuple[
Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]
]: ...
def evaluate(
self,
# eval_dataset: Optional[Union[Dataset, dict[str, Dataset]]] = None,
# ignore_keys: Optional[list[str]] = None,
# metric_key_prefix: str = "eval",
) -> dict[str, float] | None:
...
def predict(
self,
test_dataset: Dataset,
ignore_keys: Optional[list[str]] = None,
metric_key_prefix: str = "test",
) -> PredictionOutput: ...
@property
def world_size(self) -> int:
"""Returns the number of processes in the current distributed group."""
return overwatch.world_size()
@property
def device_train_batch_size(self) -> int:
"""Returns the training batch size per process."""
return self.cfg.train.train_batch_size
@property
def gradient_accumulation_steps(self) -> int:
"""Returns the number of gradient accumulation steps."""
return self.cfg.train.gradient_accumulation_steps
@property
def global_train_batch_size(self) -> int:
"""Returns the global training batch size."""
return (
self.device_train_batch_size
* self.world_size
* self.gradient_accumulation_steps
)
@property
def len_train_dataset(self) -> int:
"""Returns the length of the training dataset."""
if hasattr(self, "_len_train_dataset") and self._len_train_dataset: # type: ignore
return self._len_train_dataset # type: ignore
return len(self.train_dataset) # type: ignore
@len_train_dataset.setter
def len_train_dataset(self, length: int):
self._len_train_dataset = length
@property
def len_val_dataset(self) -> int:
"""Returns the length of the training dataset."""
if hasattr(self, "_len_train_dataset") and self._len_val_dataset: # type: ignore
return self._len_val_dataset # type: ignore
if isinstance(self.val_dataset, dict):
return sum(len(dataset) for dataset in self.val_dataset.values()) # type: ignore
return len(self.val_dataset) # type: ignore
@len_val_dataset.setter
def len_val_dataset(self, length: int):
self._len_val_dataset = length
@property
def len_train_dataloader(self) -> int:
total = self.len_train_dataset / (
self.device_train_batch_size * self.world_size
)
if hasattr(self, "train_dataloader_drop_last") and self.train_dataloader_drop_last: # type: ignore
return math.floor(total) # drop the last incomplete batch
else:
return math.ceil(total) # include the last incomplete batch
@property
def len_val_dataloader(self) -> int:
if not hasattr(self, "val_dataloader") or self.val_dataloader is None:
return 0
if isinstance(self.val_dataloader, dict):
return sum(len(dataloader) for dataloader in self.val_dataloader.values())
total = self.len_val_dataset / (
self.device_train_batch_size * self.world_size
)
if hasattr(self, "val_dataloader_drop_last") and self.val_dataloader_drop_last: # type: ignore
return math.floor(total) # drop the last incomplete batch
else:
return math.ceil(total) # include the last incomplete batch
@property
def num_steps_per_epoch(self) -> int:
"""number of global steps (sync gradients) per epoch"""
# assert self.len_train_dataloader > self.gradient_accumulation_steps
return max(self.len_train_dataloader // self.gradient_accumulation_steps, 1)
@property
def max_training_steps(self) -> int:
"""Returns the maximum number of training steps."""
if self.cfg.train.max_training_steps is not None:
return self.cfg.train.max_training_steps
else:
assert self.cfg.train.num_train_epochs is not None
return self.num_steps_per_epoch * self.world_size * self.cfg.train.num_train_epochs
@property
def max_training_epochs(self) -> int:
return math.ceil(self.max_training_steps / self.num_steps_per_epoch)
@property
def num_warmup_steps(self) -> int:
"""
Get number of steps used for a linear warmup.
"""
warmup_steps = (
self.cfg.train.warmup_steps
if self.cfg.train.warmup_steps is not None
else math.ceil(self.max_training_steps * self.cfg.train.warmup_ratio) # type: ignore
)
return warmup_steps # because accelerate divides the steps by world_size when preparing the scheduler
@property
@abstractmethod
def task_run_name(self) -> str:
"""Returns the task-specific run name."""
@property
def run_name(self) -> str:
run_name = (
f"{self.cfg.exp}{self.task_run_name}"
f".b{self.global_train_batch_size}.gpus{overwatch.world_size()}"
)
run_name = f"{run_name}.{self.timestamp}"
if self.cfg.debug:
run_name = f"debug-{run_name}"
return run_name
@property
def project_dir(self) -> str:
"""Returns the project directory for saving checkpoints and logs."""
return os.path.join(
self.cfg.train.output_dir, self.cfg.train.name, self.run_name
)
def next_epoch(self, epoch):
"""called between epochs, e.g. for resetting the distributed samplers"""
if hasattr(self, "train_sampler") and self.train_sampler is not None: # type: ignore
self.train_sampler.set_epoch(epoch) # type: ignore
@abstractmethod
def init_models(self):
"""Initialize the models for training."""
# @abstractmethod
def set_train(self):
self.model.train()
# @abstractmethod
def set_eval(self):
self.model.eval()
# @abstractmethod
# def create_optimizers(self):
# ...
# @abstractmethod
# def create_lr_schedulers(self):
# ...
# @abstractmethod
# def train_one_step(self, batch_input, global_step, local_step, accelerator=None):
# ...
def step(self, batch_input, global_step, local_step) -> tuple[bool, dict[str, Any]]:
""" Perform a single training step. """
self.local_step = local_step
self.global_step = global_step
sync_gradients, losses = self.training_step(batch_input)
return sync_gradients, losses
@abstractmethod
def prepare(self, accelerator: Accelerator) -> DataLoader:
self.optimizer, self.lr_scheduler = accelerator.prepare(
self.optimizer, self.lr_scheduler
)
self.train_dataloader = accelerator.prepare(self.train_dataloader)
if self.cfg.train.overfit_single_batch:
overwatch.warning("Overfitting a single batch: reusing first batch every step. set cfg.data.image_aug = False for true memorization.")
first_batch = next(iter(self.train_dataloader))
class SingleBatchLoader:
def __iter__(self):
while True:
yield first_batch
def __len__(self):
return 1
self.train_dataloader = SingleBatchLoader()
# FIXME if self.train_dataloader.dataset is IterableDataset:
# assert (
# abs(len(self.train_dataloader) - self.len_train_dataloader)
# <= 1 # because of drop_last option is lost in prepare.
# ), f"check calculations again, {len(self.train_dataloader)} != {self.len_train_dataloader}"
val_dataloader = getattr(self, "val_dataloader", None)
if val_dataloader is not None: # not using if self.val_dataloader to avoid DataLoader.__len__() being called on iterable dataset
self.val_dataloader = accelerator.prepare(self.val_dataloader)
""" NOTE:
We have to manually record how many steps we have accumulated gradients
because the accelerator does not do it for us.
"""
self.accelerator = accelerator
return self.train_dataloader # type: ignore
def save_checkpoint(self, global_step: int) -> str | None:
save_dir = os.path.join(self.project_dir, "checkpoints")
os.makedirs(save_dir, exist_ok=True)
ckpt_dir = os.path.join(save_dir, f"ckpt_{global_step}")
self.accelerator.save_state(ckpt_dir)
self.accelerator.wait_for_everyone()
if self.accelerator.is_main_process:
# Keep only the latest max_checkpoints_to_keep checkpoints
max_to_keep = self.train_cfg.max_checkpoints_to_keep or 100
if max_to_keep is not None and max_to_keep > 0:
# List all checkpoint directories matching ckpt_*
ckpt_dirs = [d for d in os.listdir(save_dir) if d.startswith("ckpt_") and os.path.isdir(os.path.join(save_dir, d))]
# Extract step numbers and sort by step (assume ckpt_{step})
def extract_step(d):
try:
return int(d.split("ckpt_")[-1])
except Exception:
return -1
ckpt_dirs_sorted = sorted(ckpt_dirs, key=extract_step, reverse=True)
# Remove older checkpoints if exceeding max_to_keep
for old_ckpt in ckpt_dirs_sorted[max_to_keep:]:
old_ckpt_path = os.path.join(save_dir, old_ckpt)
# if "0000" in old_ckpt:
# # force keeping ckpt saved every 10k
# continue
try:
shutil.rmtree(old_ckpt_path)
overwatch.info(f"Removed old checkpoint: {old_ckpt_path}")
except Exception as e:
overwatch.warning(f"Failed to remove old checkpoint {old_ckpt_path}: {e}")
self.accelerator.wait_for_everyone()
return ckpt_dir
def save_best_checkpoint(self, eval_metrics: dict[str, float], global_step: int) -> str | None:
# <modify> 根据eval loss保存当前最优ckpt,独立于按步数保存的ckpt,不参与max_checkpoints_to_keep清理。
metric_name = "loss"
if metric_name not in eval_metrics:
return None
save_dir = os.path.join(self.project_dir, "checkpoints")
os.makedirs(save_dir, exist_ok=True)
best_info_path = os.path.join(save_dir, "best_eval_metric.json")
best_ckpt_dir = os.path.join(save_dir, "best_ckpt")
tmp_ckpt_dir = os.path.join(save_dir, f"best_ckpt_tmp_{global_step}")
metric_value = float(eval_metrics[metric_name])
# <modify> 跳过NaN/Inf eval metric,避免训练炸掉后把正常best_ckpt覆盖成坏ckpt。
if not math.isfinite(metric_value):
overwatch.warning(
f"Skip saving best checkpoint because {metric_name} is not finite: {metric_value}"
)
return None
best_metric = float("inf")
if os.path.exists(best_info_path):
try:
with open(best_info_path, "r") as f:
best_info = json.load(f)
best_metric = float(best_info.get("metric_value", best_metric))
# <modify> 如果历史best记录已经是NaN/Inf,当作没有历史best,避免后续比较异常。
if not math.isfinite(best_metric):
overwatch.warning(
f"Ignore non-finite previous best metric from {best_info_path}: {best_metric}"
)
best_metric = float("inf")
except Exception as e:
overwatch.warning(f"Failed to read best eval metric from {best_info_path}: {e}")
if metric_value >= best_metric:
return None
if self.accelerator.is_main_process and os.path.exists(tmp_ckpt_dir):
shutil.rmtree(tmp_ckpt_dir)
self.accelerator.wait_for_everyone()
self.accelerator.save_state(tmp_ckpt_dir)
self.accelerator.wait_for_everyone()
if self.accelerator.is_main_process:
if os.path.exists(best_ckpt_dir):
shutil.rmtree(best_ckpt_dir)
os.rename(tmp_ckpt_dir, best_ckpt_dir)
best_info = {
"metric_name": metric_name,
"metric_value": metric_value,
"global_step": global_step,
"checkpoint_dir": best_ckpt_dir,
}
with open(best_info_path, "w") as f:
json.dump(best_info, f, indent=2)
overwatch.info(
f"Saved best checkpoint: {best_ckpt_dir} "
f"({metric_name}={metric_value:.6f}, previous_best={best_metric:.6f})"
)
self.accelerator.wait_for_everyone()
return best_ckpt_dir
def resume_from_checkpoint(self) -> tuple[int, Optional[str]]:
""" resume from a checkpoint if specified in the config.
the checkpoint path can be either:
1) full path to a checkpoint folder, e.g.
.runs/trainer_name/run_name/checkpoints/ckpt_xxxxxx
3) or .runs/trainer_name/run_name (latest)
"""
if self.cfg.train.resume_from_checkpoint is None:
return 0, None
# <modify> 传入run目录时默认读取best_ckpt;普通ckpt只保留ckpt_数字目录参与排序。
# 原逻辑:
# if "ckpt_" in self.cfg.train.resume_from_checkpoint:
# path = os.path.basename(self.cfg.train.resume_from_checkpoint)
# else:
# if os.path.exists(f"{self.cfg.train.resume_from_checkpoint}/checkpoints"):
# # Get the most recent checkpoint
# dirs = os.listdir(f"{self.cfg.train.resume_from_checkpoint}/checkpoints")
# dirs = sorted(dirs, key=lambda x: int(x.split("_")[1]))
# path = dirs[-1] if len(dirs) > 0 else None
# else:
# path = None
resume_path = self.cfg.train.resume_from_checkpoint
best_info_path = None
if os.path.exists(os.path.join(resume_path, "model.safetensors")):
load_path = resume_path
path = os.path.basename(load_path)
best_info_path = os.path.join(os.path.dirname(load_path), "best_eval_metric.json")
elif os.path.exists(os.path.join(resume_path, "checkpoints")):
checkpoint_dir = os.path.join(resume_path, "checkpoints")
best_path = os.path.join(checkpoint_dir, "best_ckpt")
best_info_path = os.path.join(checkpoint_dir, "best_eval_metric.json")
if os.path.exists(best_path):
path = "best_ckpt"
load_path = best_path
else:
dirs = os.listdir(checkpoint_dir)
dirs = [d for d in dirs if re.match(r"^ckpt_\d+$", d)]
dirs = sorted(dirs, key=lambda x: int(x.split("_")[1]))
path = dirs[-1] if len(dirs) > 0 else None
load_path = os.path.join(checkpoint_dir, path) if path is not None else None
else:
path = None
load_path = None
if path is None:
overwatch.critical(
f"Checkpoint '{self.cfg.train.resume_from_checkpoint}' does not exist. Starting a new training run."
)
self.cfg.train.resume_from_checkpoint = None
initial_global_step = 0
else:
overwatch.info(f"Resuming from checkpoint {load_path}")
if self.cfg.train.resume_load_full_state:
self.accelerator.load_state(load_path)
overwatch.info(f"Loaded full training state from {load_path}")
else:
# <modify> 默认 resume 只加载模型权重,optimizer/scheduler 按当前脚本配置重新开始。
from safetensors.torch import load_file
model_path = os.path.join(load_path, "model.safetensors")
self.accelerator.unwrap_model(self.model).load_state_dict(load_file(model_path, device="cpu"))
overwatch.info(f"Loaded model weights from {model_path}")
if path == "best_ckpt" and best_info_path is not None and os.path.exists(best_info_path):
with open(best_info_path, "r") as f:
best_info = json.load(f)
initial_global_step = int(best_info["global_step"]) + 1
else:
initial_global_step = int(path.split("_")[1]) + 1 # prevent from saving to the same checkpoint again
return initial_global_step, load_path
@property
def lr(self):
return self.get_lr()
def get_lr(self):
return self.lr_scheduler.get_last_lr()[0]
def get_total_grad_norm(self, params=None, norm_type=2):
if "DeepSpeedEngine" in self.model.__class__.__name__:
grad_norm = self.model.get_global_grad_norm()
return grad_norm
else:
total_norm = 0.0
if params is None:
params = self.model.parameters()
for p in params:
if p.grad is not None:
param_norm = p.grad.data.norm(norm_type)
total_norm += param_norm.item() ** norm_type
total_norm = total_norm ** (1. / norm_type)
return total_norm
def unwrap_model(self):
# Function for unwrapping if model was compiled with `torch.compile`.
model = self.accelerator.unwrap_model(self.model)
from diffusers.utils.torch_utils import is_compiled_module
model = model._orig_mod if is_compiled_module(model) else model
# songlin: revert original forward which is changed by accelerate.prepare to always return fp32
# if hasattr(model, "_original_forward"):
# model.forward = model._original_forward
return model
# def log(
# self, ):
# ...
# def evaluate(self, global_step: int, accelerator: Accelerator):
# ...
def finalize(self):
...
|