大模型底层 transformers源码解析之trainer.py

 

我会根据你提供的trainer.py源代码,用python的Graph包画出流程图,并着重介绍train()方法。以下是我的分析:

Trainer类是Transformers库中用于训练模型的核心类。它封装了训练循环的主要逻辑,并提供了许多功能,如混合精度训练、分布式训练、梯度累积等。

主要的输入参数有:

• model: 要训练的模型,可以是继承自PreTrainedModel的HuggingFace模型,也可以是自定义的nn.Module
• args: TrainingArguments对象,包含训练所需的各种参数
• data_collator: 用于将样本列表整理成batch的函数
• train_dataset: 用于训练的Dataset对象
• eval_dataset: 用于评估的Dataset对象
• tokenizer: 用于数据预处理的tokenizer
• model_init: 用于初始化模型的可调用对象
• compute_metrics: 用于在评估阶段计算各项指标的函数
• optimizers: 一个元组(optimizer, scheduler),如果不提供则会由Trainer自动创建

主要的属性有:

• self.model: 原始的模型对象
• self.model_wrapped: 如果模型被 wrapper(如 DP、DDP 等并行包装器)包装过,指被包装后的模型,否则与 self.model 相同
• self.optimizer: 优化器
• self.lr_scheduler: 学习率调度器
• self.state: 一个字典,用于记录训练过程中的状态信息,如当前epoch数、最佳评估结果等
• self.args: TrainingArguments参数对象
• self.callbacks: 回调函数列表

下面重点介绍train()方法的流程:

1. 根据 args 和训练环境进行一些初始化工作,如设置 logger、初始化 tensorboard 等

2. 加载检查点(如果指定了 resume_from_checkpoint)

3. 准备 train_dataloader

4. 计算总共需要训练的步数 max_steps 和 num_train_epochs

5. 创建 optimizer 和 lr_scheduler(如果没有传入的话)

6. 开始训练循环:

   For epoch in range(num_train_epochs):
   For step, inputs in enumerate(train_dataloader):
   - 将 inputs 通过 self._prepare_inputs 方法处理后传给 self.training_step
   - 在 self.training_step 中:
   - 调用 self.compute_loss 计算loss
   - 调用 self.accelerator.backward 计算梯度
   - 调用 optimizer.step() 和 lr_scheduler.step() 进行参数更新
   - 每 args.logging_steps 个步骤打印训练日志
   - 每 args.eval_steps 个步骤进行一次评估
   - 每 args.save_steps 个步骤保存一次检查点
   - 如果实现了自己的回调,则会在特定的步骤被调用

7. 训练结束后的一些清理和日志记录工作

8. 返回一个 TrainOutput 对象,包含训练损失和最终的metrics结果

核心方法：train()

def train(
        self,
        resume_from_checkpoint: Optional[Union[str, bool]] = None,
        trial: Union["optuna.Trial", Dict[str, Any]] = None,
        ignore_keys_for_eval: Optional[List[str]] = None,
        **kwargs,
    ):
        """
        Main training entry point.

        Args:
            resume_from_checkpoint (`str` or `bool`, *optional*):
                If a `str`, local path to a saved checkpoint as saved by a previous instance of [`Trainer`]. If a
                `bool` and equals `True`, load the last checkpoint in *args.output_dir* as saved by a previous instance
                of [`Trainer`]. If present, training will resume from the model/optimizer/scheduler states loaded here.
            trial (`optuna.Trial` or `Dict[str, Any]`, *optional*):
                The trial run or the hyperparameter dictionary for hyperparameter search.
            ignore_keys_for_eval (`List[str]`, *optional*)
                A list of keys in the output of your model (if it is a dictionary) that should be ignored when
                gathering predictions for evaluation during the training.
            kwargs (`Dict[str, Any]`, *optional*):
                Additional keyword arguments used to hide deprecated arguments
        """
        if resume_from_checkpoint is False:
            resume_from_checkpoint = None

        # memory metrics - must set up as early as possible
        self._memory_tracker.start()

        args = self.args

        self.is_in_train = True

        # Attach NEFTune hooks if necessary
        if self.neftune_noise_alpha is not None:
            self.model = self._activate_neftune(self.model)

        # do_train is not a reliable argument, as it might not be set and .train() still called, so
        # the following is a workaround:
        if (args.fp16_full_eval or args.bf16_full_eval) and not args.do_train:
            self._move_model_to_device(self.model, args.device)

        if "model_path" in kwargs:
            resume_from_checkpoint = kwargs.pop("model_path")
            warnings.warn(
                "`model_path` is deprecated and will be removed in a future version. Use `resume_from_checkpoint` "
                "instead.",
                FutureWarning,
            )
        if len(kwargs) > 0:
            raise TypeError(f"train() received got unexpected keyword arguments: {', '.join(list(kwargs.keys()))}.")
        # This might change the seed so needs to run first.
        self._hp_search_setup(trial)
        self._train_batch_size = self.args.train_batch_size

        # Model re-init
        model_reloaded = False
        if self.model_init is not None:
            # Seed must be set before instantiating the model when using model_init.
            enable_full_determinism(self.args.seed) if self.args.full_determinism else set_seed(self.args.seed)
            self.model = self.call_model_init(trial)
            model_reloaded = True
            # Reinitializes optimizer and scheduler
            self.optimizer, self.lr_scheduler = None, None

        # Load potential model checkpoint
        if isinstance(resume_from_checkpoint, bool) and resume_from_checkpoint:
            resume_from_checkpoint = get_last_checkpoint(args.output_dir)
            if resume_from_checkpoint is None:
                raise ValueError(f"No valid checkpoint found in output directory ({args.output_dir})")

        if resume_from_checkpoint is not None:
            if not is_sagemaker_mp_enabled() and not self.is_deepspeed_enabled and not self.is_fsdp_enabled:
                self._load_from_checkpoint(resume_from_checkpoint)
            # In case of repeating the find_executable_batch_size, set `self._train_batch_size` properly
            state = TrainerState.load_from_json(os.path.join(resume_from_checkpoint, TRAINER_STATE_NAME))
            if state.train_batch_size is not None:
                self._train_batch_size = state.train_batch_size

        # If model was re-initialized, put it on the right device and update self.model_wrapped
        if model_reloaded:
            if self.place_model_on_device:
                self._move_model_to_device(self.model, args.device)
            self.model_wrapped = self.model

        inner_training_loop = find_executable_batch_size(
            self._inner_training_loop, self._train_batch_size, args.auto_find_batch_size
        )
        if args.push_to_hub:
            try:
                # Disable progress bars when uploading models during checkpoints to avoid polluting stdout
                hf_hub_utils.disable_progress_bars()
                return inner_training_loop(
                    args=args,
                    resume_from_checkpoint=resume_from_checkpoint,
                    trial=trial,
                    ignore_keys_for_eval=ignore_keys_for_eval,
                )
            finally:
                hf_hub_utils.enable_progress_bars()
        else:
            return inner_training_loop(
                args=args,
                resume_from_checkpoint=resume_from_checkpoint,
                trial=trial,
                ignore_keys_for_eval=ignore_keys_for_eval,
            )

    def _inner_training_loop(
        self, batch_size=None, args=None, resume_from_checkpoint=None, trial=None, ignore_keys_for_eval=None
    ):
        self.accelerator.free_memory()
        self._train_batch_size = batch_size
        if self.args.auto_find_batch_size:
            if self.state.train_batch_size != self._train_batch_size:
                from accelerate.utils import release_memory

                (self.model_wrapped,) = release_memory(self.model_wrapped)
                self.model_wrapped = self.model

                # Check for DeepSpeed *after* the intial pass and modify the config
                if self.is_deepspeed_enabled:
                    # Temporarily unset `self.args.train_batch_size`
                    original_bs = self.args.per_device_train_batch_size
                    self.args.per_device_train_batch_size = self._train_batch_size // max(1, self.args.n_gpu)
                    self.propagate_args_to_deepspeed(True)
                    self.args.per_device_train_batch_size = original_bs
            self.state.train_batch_size = self._train_batch_size
        logger.debug(f"Currently training with a batch size of: {self._train_batch_size}")
        # Data loader and number of training steps
        train_dataloader = self.get_train_dataloader()
        if self.is_fsdp_xla_v2_enabled:
            train_dataloader = tpu_spmd_dataloader(train_dataloader)

        # Setting up training control variables:
        # number of training epochs: num_train_epochs
        # number of training steps per epoch: num_update_steps_per_epoch
        # total number of training steps to execute: max_steps
        total_train_batch_size = self._train_batch_size * args.gradient_accumulation_steps * args.world_size

        len_dataloader = None
        num_train_tokens = None
        if has_length(train_dataloader):
            len_dataloader = len(train_dataloader)
            num_update_steps_per_epoch = len_dataloader // args.gradient_accumulation_steps
            num_update_steps_per_epoch = max(num_update_steps_per_epoch, 1)
            num_examples = self.num_examples(train_dataloader)
            if args.max_steps > 0:
                max_steps = args.max_steps
                num_train_epochs = args.max_steps // num_update_steps_per_epoch + int(
                    args.max_steps % num_update_steps_per_epoch > 0
                )
                # May be slightly incorrect if the last batch in the training dataloader has a smaller size but it's
                # the best we can do.
                num_train_samples = args.max_steps * total_train_batch_size
                if args.include_tokens_per_second:
                    num_train_tokens = (
                        self.num_tokens(train_dataloader, args.max_steps) * args.gradient_accumulation_steps
                    )
            else:
                max_steps = math.ceil(args.num_train_epochs * num_update_steps_per_epoch)
                num_train_epochs = math.ceil(args.num_train_epochs)
                num_train_samples = self.num_examples(train_dataloader) * args.num_train_epochs
                if args.include_tokens_per_second:
                    num_train_tokens = self.num_tokens(train_dataloader) * args.num_train_epochs
        elif args.max_steps > 0:  # Rely on max_steps when dataloader does not have a working size
            max_steps = args.max_steps
            # Setting a very large number of epochs so we go as many times as necessary over the iterator.
            num_train_epochs = sys.maxsize
            num_update_steps_per_epoch = max_steps
            num_examples = total_train_batch_size * args.max_steps
            num_train_samples = args.max_steps * total_train_batch_size
            if args.include_tokens_per_second:
                num_train_tokens = self.num_tokens(train_dataloader, args.max_steps) * args.gradient_accumulation_steps
        else:
            raise ValueError(
                "args.max_steps must be set to a positive value if dataloader does not have a length, was"
                f" {args.max_steps}"
            )

        if DebugOption.UNDERFLOW_OVERFLOW in self.args.debug:
            if self.args.n_gpu > 1:
                # nn.DataParallel(model) replicates the model, creating new variables and module
                # references registered here no longer work on other gpus, breaking the module
                raise ValueError(
                    "Currently --debug underflow_overflow is not supported under DP. Please use DDP"
                    " (torchrun or torch.distributed.launch (deprecated))."
                )
            else:
                debug_overflow = DebugUnderflowOverflow(self.model)  # noqa

        delay_optimizer_creation = is_sagemaker_mp_enabled() or self.is_fsdp_xla_enabled or self.is_fsdp_enabled

        # We need to reset the scheduler, as its parameters may be different on subsequent calls
        if self._created_lr_scheduler:
            self.lr_scheduler = None
            self._created_lr_scheduler = False

        if self.is_deepspeed_enabled:
            self.optimizer, self.lr_scheduler = deepspeed_init(self, num_training_steps=max_steps)

        if not delay_optimizer_creation:
            self.create_optimizer_and_scheduler(num_training_steps=max_steps)