Source code for paddlenlp.trainer.trainer_callback

# Copyright 2020-present the HuggingFace Inc. team.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# This file is modified from
#  https://github.com/huggingface/transformers/blob/main/src/transformers/trainer_callback.py
"""
Callbacks to use with the Trainer class and customize the training loop.
"""
import dataclasses
import json
from dataclasses import dataclass
from typing import Dict, List, Optional, Union

import numpy as np
from tqdm.auto import tqdm

from .trainer_utils import IntervalStrategy, has_length
from .training_args import TrainingArguments
from paddlenlp.utils.log import logger

__all__ = [
    "TrainerState",
    "TrainerControl",
    "TrainerCallback",
    "CallbackHandler",
    "DefaultFlowCallback",
    "ProgressCallback",
    "PrinterCallback",
    "EarlyStoppingCallback",
]


[docs]@dataclass class TrainerState: """ A class containing the [`Trainer`] inner state that will be saved along the model and optimizer when checkpointing and passed to the [`TrainerCallback`]. <Tip> In all this class, one step is to be understood as one update step. When using gradient accumulation, one update step may require several forward and backward passes: if you use `gradient_accumulation_steps=n`, then one update step requires going through *n* batches. </Tip> Args: epoch (`float`, *optional*): Only set during training, will represent the epoch the training is at (the decimal part being the percentage of the current epoch completed). global_step (`int`, *optional*, defaults to 0): During training, represents the number of update steps completed. max_steps (`int`, *optional*, defaults to 0): The number of update steps to do during the current training. total_flos (`float`, *optional*, defaults to 0): The total number of floating operations done by the model since the beginning of training (stored as floats to avoid overflow). log_history (`List[Dict[str, float]]`, *optional*): The list of logs done since the beginning of training. best_metric (`float`, *optional*): When tracking the best model, the value of the best metric encountered so far. best_model_checkpoint (`str`, *optional*): When tracking the best model, the value of the name of the checkpoint for the best model encountered so far. is_local_process_zero (`bool`, *optional*, defaults to `True`): Whether or not this process is the local (e.g., on one machine if training in a distributed fashion on several machines) main process. is_world_process_zero (`bool`, *optional*, defaults to `True`): Whether or not this process is the global main process (when training in a distributed fashion on several machines, this is only going to be `True` for one process). """ epoch: Optional[float] = None global_step: int = 0 max_steps: int = 0 num_train_epochs: int = 0 total_flos: float = 0 log_history: List[Dict[str, float]] = None best_metric: Optional[float] = None best_model_checkpoint: Optional[str] = None is_local_process_zero: bool = True is_world_process_zero: bool = True trial_name: str = None trial_params: Dict[str, Union[str, float, int, bool]] = None def __post_init__(self): if self.log_history is None: self.log_history = []
[docs] def save_to_json(self, json_path: str): """Save the content of this instance in JSON format inside `json_path`.""" json_string = json.dumps( dataclasses.asdict(self), indent=2, sort_keys=True) + "\n" with open(json_path, "w", encoding="utf-8") as f: f.write(json_string)
[docs] @classmethod def load_from_json(cls, json_path: str): """Create an instance from the content of `json_path`.""" with open(json_path, "r", encoding="utf-8") as f: text = f.read() return cls(**json.loads(text))
[docs]@dataclass class TrainerControl: """ A class that handles the [`Trainer`] control flow. This class is used by the [`TrainerCallback`] to activate some switches in the training loop. Args: should_training_stop (`bool`, *optional*, defaults to `False`): Whether or not the training should be interrupted. If `True`, this variable will not be set back to `False`. The training will just stop. should_epoch_stop (`bool`, *optional*, defaults to `False`): Whether or not the current epoch should be interrupted. If `True`, this variable will be set back to `False` at the beginning of the next epoch. should_save (`bool`, *optional*, defaults to `False`): Whether or not the model should be saved at this step. If `True`, this variable will be set back to `False` at the beginning of the next step. should_evaluate (`bool`, *optional*, defaults to `False`): Whether or not the model should be evaluated at this step. If `True`, this variable will be set back to `False` at the beginning of the next step. should_log (`bool`, *optional*, defaults to `False`): Whether or not the logs should be reported at this step. If `True`, this variable will be set back to `False` at the beginning of the next step. """ should_training_stop: bool = False should_epoch_stop: bool = False should_save: bool = False should_evaluate: bool = False should_log: bool = False def _new_training(self): """Internal method that resets the variable for a new training.""" self.should_training_stop = False def _new_epoch(self): """Internal method that resets the variable for a new epoch.""" self.should_epoch_stop = False def _new_step(self): """Internal method that resets the variable for a new step.""" self.should_save = False self.should_evaluate = False self.should_log = False
[docs]class TrainerCallback: """ A class for objects that will inspect the state of the training loop at some events and take some decisions. At each of those events the following arguments are available: Args: args ([`TrainingArguments`]): The training arguments used to instantiate the [`Trainer`]. state ([`TrainerState`]): The current state of the [`Trainer`]. control ([`TrainerControl`]): The object that is returned to the [`Trainer`] and can be used to make some decisions. model ([`PreTrainedModel`] or `paddle.nn.Layer`): The model being trained. tokenizer ([`PreTrainedTokenizer`]): The tokenizer used for encoding the data. optimizer (`paddle.optimizer.Optimizer`): The optimizer used for the training steps. lr_scheduler (`paddle.optimizer.lr.LRScheduler`): The scheduler used for setting the learning rate. train_dataloader (`paddle.io.DataLoader`, *optional*): The current dataloader used for training. eval_dataloader (`paddle.io.DataLoader`, *optional*): The current dataloader used for training. metrics (`Dict[str, float]`): The metrics computed by the last evaluation phase. Those are only accessible in the event `on_evaluate`. logs (`Dict[str, float]`): The values to log. Those are only accessible in the event `on_log`. The `control` object is the only one that can be changed by the callback, in which case the event that changes it should return the modified version. The argument `args`, `state` and `control` are positionals for all events, all the others are grouped in `kwargs`. You can unpack the ones you need in the signature of the event using them. As an example, see the code of the simple [`~transformer.PrinterCallback`]. Example: ```python class PrinterCallback(TrainerCallback): def on_log(self, args, state, control, logs=None, **kwargs): _ = logs.pop("total_flos", None) if state.is_local_process_zero: logger.info(logs) ```"""
[docs] def on_init_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the end of the initialization of the [`Trainer`]. """ pass
[docs] def on_train_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the beginning of training. """ pass
[docs] def on_train_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the end of training. """ pass
[docs] def on_epoch_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the beginning of an epoch. """ pass
[docs] def on_epoch_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the end of an epoch. """ pass
[docs] def on_step_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the beginning of a training step. If using gradient accumulation, one training step might take several inputs. """ pass
[docs] def on_substep_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the end of an substep during gradient accumulation. """ pass
[docs] def on_step_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called at the end of a training step. If using gradient accumulation, one training step might take several inputs. """ pass
[docs] def on_evaluate(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called after an evaluation phase. """ pass
[docs] def on_save(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called after a checkpoint save. """ pass
[docs] def on_log(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called after logging the last logs. """ pass
[docs] def on_prediction_step(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): """ Event called after a prediction step. """ pass
[docs]class CallbackHandler(TrainerCallback): """Internal class that just calls the list of callbacks in order.""" def __init__(self, callbacks, model, tokenizer, optimizer, lr_scheduler): self.callbacks = [] for cb in callbacks: self.add_callback(cb) self.model = model self.tokenizer = tokenizer self.optimizer = optimizer self.lr_scheduler = lr_scheduler self.train_dataloader = None self.eval_dataloader = None if not any( isinstance(cb, DefaultFlowCallback) for cb in self.callbacks): logger.warning( "The Trainer will not work properly if you don't have a `DefaultFlowCallback` in its callbacks. You\n" + "should add one before training with `trainer.add_callback(DefaultFlowCallback). The current list of" + "callbacks is\n:" + self.callback_list) def add_callback(self, callback): cb = callback() if isinstance(callback, type) else callback cb_class = callback if isinstance(callback, type) else callback.__class__ if cb_class in [c.__class__ for c in self.callbacks]: logger.warning( f"You are adding a {cb_class} to the callbacks of this Trainer, but there is already one. The current" + "list of callbacks is\n:" + self.callback_list) self.callbacks.append(cb) def pop_callback(self, callback): if isinstance(callback, type): for cb in self.callbacks: if isinstance(cb, callback): self.callbacks.remove(cb) return cb else: for cb in self.callbacks: if cb == callback: self.callbacks.remove(cb) return cb def remove_callback(self, callback): if isinstance(callback, type): for cb in self.callbacks: if isinstance(cb, callback): self.callbacks.remove(cb) return else: self.callbacks.remove(callback) @property def callback_list(self): return "\n".join(cb.__class__.__name__ for cb in self.callbacks)
[docs] def on_init_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): return self.call_event("on_init_end", args, state, control)
[docs] def on_train_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): control.should_training_stop = False return self.call_event("on_train_begin", args, state, control)
[docs] def on_train_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): return self.call_event("on_train_end", args, state, control)
[docs] def on_epoch_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): control.should_epoch_stop = False return self.call_event("on_epoch_begin", args, state, control)
[docs] def on_epoch_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): return self.call_event("on_epoch_end", args, state, control)
[docs] def on_step_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): control.should_log = False control.should_evaluate = False control.should_save = False return self.call_event("on_step_begin", args, state, control)
[docs] def on_substep_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): return self.call_event("on_substep_end", args, state, control)
[docs] def on_step_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): return self.call_event("on_step_end", args, state, control)
[docs] def on_evaluate(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, metrics): control.should_evaluate = False return self.call_event("on_evaluate", args, state, control, metrics=metrics)
[docs] def on_save(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): control.should_save = False return self.call_event("on_save", args, state, control)
[docs] def on_log(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, logs): control.should_log = False return self.call_event("on_log", args, state, control, logs=logs)
[docs] def on_prediction_step(self, args: TrainingArguments, state: TrainerState, control: TrainerControl): return self.call_event("on_prediction_step", args, state, control)
def call_event(self, event, args, state, control, **kwargs): for callback in self.callbacks: result = getattr(callback, event)( args, state, control, model=self.model, tokenizer=self.tokenizer, optimizer=self.optimizer, lr_scheduler=self.lr_scheduler, train_dataloader=self.train_dataloader, eval_dataloader=self.eval_dataloader, **kwargs, ) # A Callback can skip the return of `control` if it doesn't change it. if result is not None: control = result return control
[docs]class DefaultFlowCallback(TrainerCallback): """ A [`TrainerCallback`] that handles the default flow of the training loop for logs, evaluation and checkpoints. """
[docs] def on_step_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): # Log if state.global_step == 1 and args.logging_first_step: control.should_log = True if args.logging_strategy == IntervalStrategy.STEPS and state.global_step % args.logging_steps == 0: control.should_log = True # Evaluate if args.evaluation_strategy == IntervalStrategy.STEPS and state.global_step % args.eval_steps == 0: control.should_evaluate = True # Save if (args.save_strategy == IntervalStrategy.STEPS and args.save_steps > 0 and state.global_step % args.save_steps == 0): control.should_save = True # End training if state.global_step >= state.max_steps: control.should_training_stop = True # Log and save on end if args.logging_strategy == IntervalStrategy.STEPS and state.global_step >= args.logging_steps: control.should_log = True if args.evaluation_strategy == IntervalStrategy.STEPS and state.global_step >= args.eval_steps: control.should_evaluate = True if args.save_strategy == IntervalStrategy.STEPS and args.save_steps > 0 and state.global_step >= args.save_steps: control.should_save = True return control
[docs] def on_epoch_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs): # Log if args.logging_strategy == IntervalStrategy.EPOCH: control.should_log = True # Evaluate if args.evaluation_strategy == IntervalStrategy.EPOCH: control.should_evaluate = True # Save if args.save_strategy == IntervalStrategy.EPOCH: control.should_save = True return control
[docs]class ProgressCallback(TrainerCallback): """ A [`TrainerCallback`] that displays the progress of training or evaluation. """ def __init__(self): self.training_bar = None self.prediction_bar = None
[docs] def on_train_begin(self, args, state, control, **kwargs): if state.is_local_process_zero: self.training_bar = tqdm(total=state.max_steps) self.current_step = 0
[docs] def on_step_end(self, args, state, control, **kwargs): if state.is_local_process_zero: self.training_bar.update(state.global_step - self.current_step) self.current_step = state.global_step
[docs] def on_prediction_step(self, args, state, control, eval_dataloader=None, **kwargs): if state.is_local_process_zero and has_length(eval_dataloader.dataset): if self.prediction_bar is None: self.prediction_bar = tqdm(total=len(eval_dataloader), leave=self.training_bar is None) self.prediction_bar.update(1)
[docs] def on_evaluate(self, args, state, control, **kwargs): if state.is_local_process_zero: if self.prediction_bar is not None: self.prediction_bar.close() self.prediction_bar = None
[docs] def on_log(self, args, state, control, logs=None, **kwargs): if state.is_local_process_zero and self.training_bar is not None: _ = logs.pop("total_flos", None) if type(logs) is dict: logs_str = ", ".join(f"{k}: {v}" for k, v in logs.items()) else: logs_str = str(logs) self.training_bar.write(logs_str)
[docs] def on_train_end(self, args, state, control, **kwargs): if state.is_local_process_zero: self.training_bar.close() self.training_bar = None
[docs]class PrinterCallback(TrainerCallback): """ A bare [`TrainerCallback`] that just prints the logs. """
[docs] def on_log(self, args, state, control, logs=None, **kwargs): _ = logs.pop("total_flos", None) if state.is_local_process_zero: if type(logs) is dict: logger.info(", ".join(f"{k}: {v}" for k, v in logs.items())) else: logger.info(logs)
[docs]class EarlyStoppingCallback(TrainerCallback): """ A [`TrainerCallback`] that handles early stopping. Args: early_stopping_patience (`int`): Use with `metric_for_best_model` to stop training when the specified metric worsens for `early_stopping_patience` evaluation calls. early_stopping_threshold(`float`, *optional*): Use with TrainingArguments `metric_for_best_model` and `early_stopping_patience` to denote how much the specified metric must improve to satisfy early stopping conditions. ` This callback depends on [`TrainingArguments`] argument *load_best_model_at_end* functionality to set best_metric in [`TrainerState`]. """ def __init__(self, early_stopping_patience: int = 1, early_stopping_threshold: Optional[float] = 0.0): self.early_stopping_patience = early_stopping_patience self.early_stopping_threshold = early_stopping_threshold # early_stopping_patience_counter denotes the number of times validation metrics failed to improve. self.early_stopping_patience_counter = 0 def check_metric_value(self, args, state, control, metric_value): # best_metric is set by code for load_best_model operator = np.greater if args.greater_is_better else np.less if state.best_metric is None or ( operator(metric_value, state.best_metric) and abs(metric_value - state.best_metric) > self.early_stopping_threshold): self.early_stopping_patience_counter = 0 else: self.early_stopping_patience_counter += 1
[docs] def on_train_begin(self, args, state, control, **kwargs): assert args.load_best_model_at_end, "EarlyStoppingCallback requires load_best_model_at_end = True" assert ( args.metric_for_best_model is not None ), "EarlyStoppingCallback requires metric_for_best_model is defined" assert ( args.evaluation_strategy != IntervalStrategy.NO ), "EarlyStoppingCallback requires IntervalStrategy of steps or epoch"
[docs] def on_evaluate(self, args, state, control, metrics, **kwargs): metric_to_check = args.metric_for_best_model if not metric_to_check.startswith("eval_"): metric_to_check = f"eval_{metric_to_check}" metric_value = metrics.get(metric_to_check) if metric_value is None: logger.warning( f"early stopping required metric_for_best_model, but did not find {metric_to_check} so early stopping is disabled" ) return self.check_metric_value(args, state, control, metric_value) if self.early_stopping_patience_counter >= self.early_stopping_patience: control.should_training_stop = True