|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| import json
|
| import os
|
| from types import MethodType
|
| from typing import TYPE_CHECKING, Optional, Union
|
|
|
| import torch
|
| from transformers import Trainer
|
| from typing_extensions import override
|
|
|
| from ...extras import logging
|
| from ...extras.packages import is_transformers_version_greater_than
|
| from ..callbacks import FixValueHeadModelCallback, SaveProcessorCallback
|
| from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
|
|
|
|
|
| if TYPE_CHECKING:
|
| from transformers import PreTrainedModel, ProcessorMixin
|
| from transformers.trainer import PredictionOutput
|
|
|
| from ...hparams import FinetuningArguments
|
|
|
|
|
| logger = logging.get_logger(__name__)
|
|
|
|
|
| class PairwiseTrainer(Trainer):
|
| r"""Inherits Trainer to compute pairwise loss."""
|
|
|
| def __init__(
|
| self, finetuning_args: "FinetuningArguments", processor: Optional["ProcessorMixin"], **kwargs
|
| ) -> None:
|
| if is_transformers_version_greater_than("4.46"):
|
| kwargs["processing_class"] = kwargs.pop("tokenizer")
|
|
|
| super().__init__(**kwargs)
|
| self.model_accepts_loss_kwargs = False
|
| self.finetuning_args = finetuning_args
|
| self.can_return_loss = True
|
| self.add_callback(FixValueHeadModelCallback)
|
|
|
| if processor is not None:
|
| self.add_callback(SaveProcessorCallback(processor))
|
|
|
| if finetuning_args.use_badam:
|
| from badam import BAdamCallback, clip_grad_norm_old_version
|
|
|
| self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
|
| self.add_callback(BAdamCallback)
|
|
|
| @override
|
| def create_optimizer(self) -> "torch.optim.Optimizer":
|
| if self.optimizer is None:
|
| self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
|
| return super().create_optimizer()
|
|
|
| @override
|
| def create_scheduler(
|
| self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
|
| ) -> "torch.optim.lr_scheduler.LRScheduler":
|
| create_custom_scheduler(self.args, num_training_steps, optimizer)
|
| return super().create_scheduler(num_training_steps, optimizer)
|
|
|
| @override
|
| def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
|
| if self.finetuning_args.disable_shuffling:
|
| return torch.utils.data.SequentialSampler(self.train_dataset)
|
|
|
| return super()._get_train_sampler()
|
|
|
| @override
|
| def compute_loss(
|
| self, model: "PreTrainedModel", inputs: dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
|
| ) -> Union["torch.Tensor", tuple["torch.Tensor", list["torch.Tensor"]]]:
|
| r"""Compute pairwise loss. The first n examples are chosen and the last n examples are rejected.
|
|
|
| Subclass and override to inject custom behavior.
|
|
|
| Note that the first element will be removed from the output tuple.
|
| See: https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/trainer.py#L3842
|
| """
|
| _, _, values = model(**inputs, output_hidden_states=True, return_dict=True, use_cache=False)
|
| batch_size = inputs["input_ids"].size(0) // 2
|
| chosen_masks, rejected_masks = torch.split(inputs["attention_mask"], batch_size, dim=0)
|
| chosen_rewards, rejected_rewards = torch.split(values, batch_size, dim=0)
|
| chosen_scores = chosen_rewards.gather(dim=-1, index=(chosen_masks.sum(dim=-1, keepdim=True) - 1))
|
| rejected_scores = rejected_rewards.gather(dim=-1, index=(rejected_masks.sum(dim=-1, keepdim=True) - 1))
|
| chosen_scores, rejected_scores = chosen_scores.squeeze(), rejected_scores.squeeze()
|
|
|
| loss = -torch.nn.functional.logsigmoid(chosen_scores.float() - rejected_scores.float()).mean()
|
| if return_outputs:
|
| return loss, (loss, chosen_scores, rejected_scores)
|
| else:
|
| return loss
|
|
|
| def save_predictions(self, predict_results: "PredictionOutput") -> None:
|
| r"""Save model predictions to `output_dir`.
|
|
|
| A custom behavior that not contained in Seq2SeqTrainer.
|
| """
|
| if not self.is_world_process_zero():
|
| return
|
|
|
| output_prediction_file = os.path.join(self.args.output_dir, "generated_predictions.jsonl")
|
| logger.info_rank0(f"Saving prediction results to {output_prediction_file}")
|
| chosen_scores, rejected_scores = predict_results.predictions
|
|
|
| with open(output_prediction_file, "w", encoding="utf-8") as writer:
|
| res: list[str] = []
|
| for c_score, r_score in zip(chosen_scores, rejected_scores):
|
| res.append(json.dumps({"chosen": round(float(c_score), 2), "rejected": round(float(r_score), 2)}))
|
|
|
| writer.write("\n".join(res))
|
|
|