modeling_bert_classifier.py

from typing import Dict, List, Union
from transformers import BertPreTrainedModel, BertModel,PreTrainedTokenizer
import torch.nn as nn
import torch
class BertForStorySkillClassification(BertPreTrainedModel):
    def __init__(self,config):
        super(BertForStorySkillClassification,self).__init__(config)
        self.num_labels = config.num_labels
        self.bert = BertModel(config)
        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
        self.post_init()

    def forward(self,input_ids,attention_mask=None,labels=None,**kwargs):
        outputs = self.bert(input_ids,attention_mask=attention_mask)
        cls_hidden_state = outputs.last_hidden_state[:,0,:]  ##  [batch_size,seq_len,hidden_size]
        logits = self.classifier(cls_hidden_state) ## [batch_size,num_labels]
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits.view(-1,self.num_labels),labels.view(-1))
            return loss
        return logits
    

    def predict(
            self,
            texts: Union[str, List[str]],
            tokenizer: PreTrainedTokenizer,
            batch_size: int = 32,
            return_probabilities: bool = False,
            device: Union[str, torch.device] = 'cpu',
        ) -> List[Dict]:
            """
            对输入文本进行分类预测。

            Args:
                texts: 单条文本或文本列表，例如 "故事中的角色是谁？" 或 ["问题1", "问题2"]
                tokenizer: 分词器实例（需与模型兼容）
                batch_size: 批处理大小（提升推理速度）
                return_probabilities: 是否返回概率值（默认返回标签）
                device: 指定设备（例如 "cuda" 或 "cpu"），默认自动检测模型当前设备

            Returns:
                预测结果列表，格式为：
                [{"text": "输入文本", "label": "预测标签", "score": 置信度}, ...]
            """
            # 自动获取模型所在设备
            if device is None:
                device = self.device

            # 统一输入格式为列表
            if isinstance(texts, str):
                texts = [texts]

            # 结果存储
            predictions = []

            # 批处理预测
            with torch.no_grad():
                for i in range(0, len(texts), batch_size):
                    batch_texts = texts[i : i + batch_size]

                    # 分词并转换为张量
                    inputs = tokenizer(
                        batch_texts,
                        padding=True,
                        truncation=True,
                        return_tensors="pt",
                        max_length=512,  # 与BERT最大长度一致
                    ).to(device)

                    # 模型推理
                    logits = self(**inputs)
                    probs = torch.softmax(logits, dim=-1)
                    scores, class_ids = torch.max(probs, dim=-1)

                    # 转换为标签和分数
                    for text, class_id, score in zip(batch_texts, class_ids, scores):
                        label = self.config.id2label[class_id.item()]
                        result = {"text": text, "label": label}
                        if return_probabilities:
                            result["score"] = score.item()
                        predictions.append(result)

            return predictions