modeling_my_roberta.py

# modeling_my_roberta.py
from typing import Optional, Tuple
import torch
import torch.nn as nn
from transformers import RobertaModel, RobertaPreTrainedModel

class MyRobertaForSequenceClassification(RobertaPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = getattr(config, "num_labels", 4)
        self.roberta = RobertaModel(config, add_pooling_layer=False)
        self.classifier = nn.Linear(config.hidden_size, self.num_labels)
        self.loss_fn = nn.CrossEntropyLoss()
        self.post_init()  # init_weights

    def _get_pad_id(self, input_ids):
        # Roberta pad = 1가 일반적이지만 토크나이저에 맞춰 attention_mask 직접 받는 게 더 안전
        return 1

    def _pool(self, last_hidden_state, attention_mask, model_type: str = "roberta"):
        # 너의 로직: roberta면 CLS, 그 외 mask mean
        if last_hidden_state.dim() == 3:
            if model_type in {"bert", "roberta", "deberta", "xlm-roberta", "electra"}:
                return last_hidden_state[:, 0, :]
            mask = attention_mask.unsqueeze(-1).float()
            summed = (last_hidden_state * mask).sum(dim=1)
            denom = mask.sum(dim=1).clamp(min=1e-9)
            return summed / denom
        elif last_hidden_state.dim() == 2:
            return last_hidden_state
        else:
            raise ValueError(f"Unexpected hidden dim: {last_hidden_state.dim()}")

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        labels: Optional[torch.LongTensor] = None,
        **kwargs
    ):
        if attention_mask is None and input_ids is not None:
            pad_id = self._get_pad_id(input_ids)
            attention_mask = input_ids.ne(pad_id)

        outputs = self.roberta(input_ids=input_ids, attention_mask=attention_mask, return_dict=True)
        hidden = outputs.last_hidden_state
        pooled = self._pool(hidden, attention_mask, "roberta")
        logits = self.classifier(pooled)

        loss = None
        if labels is not None:
            loss = self.loss_fn(logits, labels.long())

        return {"loss": loss, "logits": logits}