Source code for FlagEmbedding.finetune.reranker.encoder_only.base.modeling

from transformers import PreTrainedModel, AutoTokenizer
import logging

from FlagEmbedding.abc.finetune.reranker import AbsRerankerModel

logger = logging.getLogger(__name__)




class CrossEncoderModel(AbsRerankerModel):
    """Model class for reranker.

    Args:
        base_model (PreTrainedModel): The underlying pre-trained model used for encoding and scoring input pairs.
        tokenizer (AutoTokenizer, optional): The tokenizer for encoding input text. Defaults to ``None``.
        train_batch_size (int, optional): The batch size to use. Defaults to ``4``.
    """
    def __init__(
        self,
        base_model: PreTrainedModel,
        tokenizer: AutoTokenizer = None,
        train_batch_size: int = 4,
    ):
        super().__init__(
            base_model,
            tokenizer=tokenizer,
            train_batch_size=train_batch_size,
        )



    def encode(self, features):
        """Encodes input features to logits.

        Args:
            features (dict): Dictionary with input features.

        Returns:
            torch.Tensor: The logits output from the model.
        """
        return self.model(**features, return_dict=True).logits