modeling_csatv2.py

# modeling_csatv2.py
#
# Hugging Face Transformers용 CSATv2 래퍼
# - Config: CSATv2Config
# - Model: CSATv2ForImageClassification
#
# 사용 예:
#   from transformers import AutoImageProcessor, AutoModelForImageClassification
#   model = AutoModelForImageClassification.from_pretrained(
#       "Hyunil/CSATv2", trust_remote_code=True
#   )

from typing import Optional, Union, Tuple

import torch
import torch.nn as nn

from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import ImageClassifierOutput

from .CSATv2 import CSATv2  # 네가 올린 백본 클래스 사용


class CSATv2Config(PretrainedConfig):
    model_type = "csatv2"

    def __init__(
        self,
        image_size: int = 512,
        num_channels: int = 3,
        num_labels: int = 1000,
        drop_path_rate: float = 0.0,
        head_init_scale: float = 1.0,
        **kwargs,
    ):
        """
        HF가 사용할 설정 값들.
        """
        super().__init__(num_labels=num_labels, **kwargs)
        self.image_size = image_size
        self.num_channels = num_channels
        self.drop_path_rate = drop_path_rate
        self.head_init_scale = head_init_scale

        # label 매핑이 안 들어오면 기본값 생성
        if self.id2label is None or self.label2id is None:
            self.id2label = {i: f"LABEL_{i}" for i in range(num_labels)}
            self.label2id = {v: k for k, v in self.id2label.items()}


class CSATv2ForImageClassification(PreTrainedModel):
    """
    Hugging Face용 ImageNet 분류 모델 래퍼
    - backbone: CSATv2 (네가 구현한 모델)
    - forward(pixel_values, labels=None)
    """

    config_class = CSATv2Config

    def __init__(self, config: CSATv2Config):
        super().__init__(config)
        self.num_labels = config.num_labels

        # 네가 만든 CSATv2 백본을 그대로 사용
        self.backbone = CSATv2(
            img_size=config.image_size,
            num_classes=config.num_labels,
            drop_path_rate=config.drop_path_rate,
            head_init_scale=config.head_init_scale,
        )

        # transformers 권장: 내부 가중치 등록 후 post_init 호출
        self.post_init()

    def forward(
        self,
        pixel_values: torch.Tensor = None,
        labels: Optional[torch.Tensor] = None,
        output_hidden_states: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[ImageClassifierOutput, Tuple]:
        """
        Args:
            pixel_values: (batch, 3, H, W), ImageNet 정규화까지 된 이미지
            labels: (batch,) 0~999 class index
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if pixel_values is None:
            raise ValueError("You must provide pixel_values")

        # CSATv2는 이미 logits를 반환함
        logits = self.backbone(pixel_values)

        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(
                logits.view(-1, self.num_labels),
                labels.view(-1),
            )

        if not return_dict:
            output = (logits,)
            return ((loss,) + output) if loss is not None else output

        return ImageClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=None,
            attentions=None,
        )