Add/Update backbone checkpoints (count=6)

ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

manifest_20260210_163348.json

@@ -0,0 +1,53 @@
+{
+  "timestamp": "20260210_163348",
+  "repo_id": "dsaint31/bb_mlp_224",
+  "revision": "main",
+  "tag": null,
+  "num_labels": 3,
+  "build_device": "mps",
+  "count": 6,
+  "items": [
+    {
+      "backbone": "google/vit-base-patch16-224",
+      "subdir": "models/google__vit-base-patch16-224",
+      "dirname": "google__vit-base-patch16-224"
+    },
+    {
+      "backbone": "microsoft/swin-tiny-patch4-window7-224",
+      "subdir": "models/microsoft__swin-tiny-patch4-window7-224",
+      "dirname": "microsoft__swin-tiny-patch4-window7-224"
+    },
+    {
+      "backbone": "microsoft/resnet-50",
+      "subdir": "models/microsoft__resnet-50",
+      "dirname": "microsoft__resnet-50"
+    },
+    {
+      "backbone": "google/efficientnet-b0",
+      "subdir": "models/google__efficientnet-b0",
+      "dirname": "google__efficientnet-b0"
+    },
+    {
+      "backbone": "timm/densenet121.tv_in1k",
+      "subdir": "models/timm__densenet121.tv_in1k",
+      "dirname": "timm__densenet121.tv_in1k"
+    },
+    {
+      "backbone": "torchvision/densenet121",
+      "subdir": "models/torchvision__densenet121",
+      "dirname": "torchvision__densenet121"
+    }
+  ],
+  "root_code_included": true,
+  "root_code_files": [
+    "ds_proc.py",
+    "ds_model.py",
+    "ds_cfg.py"
+  ],
+  "subfolder_code_included": true,
+  "subfolder_code_files": [
+    "ds_proc.py",
+    "ds_model.py",
+    "ds_cfg.py"
+  ]
+}

models/google__efficientnet-b0/config.json

@@ -24,7 +24,7 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_142559",
+    "created_at": "20260210_163348",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/google__efficientnet-b0",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
@@ -40,15 +40,13 @@
     "root_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "subfolder_code_included": true,
     "subfolder_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "processor_use_fast": true
   }

models/google__efficientnet-b0/ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

models/google__efficientnet-b0/ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

models/google__efficientnet-b0/ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

models/google__efficientnet-b0/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7020f515e549776c727f92374975faf8bb9878444809463d1d1069e08f68d735
+oid sha256:05324e85fb965b74d443308262fc7c776bcc001035e3d66bf63a52b0ba4ce300
 size 17558436

models/google__vit-base-patch16-224/config.json

@@ -24,7 +24,7 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_142559",
+    "created_at": "20260210_163348",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/google__vit-base-patch16-224",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
@@ -40,15 +40,13 @@
     "root_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "subfolder_code_included": true,
     "subfolder_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "processor_use_fast": true
   }

models/google__vit-base-patch16-224/ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

models/google__vit-base-patch16-224/ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

models/google__vit-base-patch16-224/ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

models/google__vit-base-patch16-224/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c051b46d9d05d5ee0182d84ec6a01a6b2f03e4a3197e1989dbb2119b1c1554a9
+oid sha256:152041a83fb83b3877e72054e3d74e4542e1b28dba1ca7a70682b03efd68bae4
 size 346372132

models/microsoft__resnet-50/config.json

@@ -24,7 +24,7 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_142559",
+    "created_at": "20260210_163348",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/microsoft__resnet-50",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
@@ -40,15 +40,13 @@
     "root_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "subfolder_code_included": true,
     "subfolder_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "processor_use_fast": true
   }

models/microsoft__resnet-50/ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

models/microsoft__resnet-50/ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

models/microsoft__resnet-50/ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

models/microsoft__resnet-50/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:672f7ce423ab8366b6988ab17a5946d3a9191438e7b2831ee0ac03e9ee9b28f0
+oid sha256:32610cc9f181dd7de16c79732221106bf7431ae82693471b9410446efe103482
 size 96388660

models/microsoft__swin-tiny-patch4-window7-224/config.json

@@ -24,7 +24,7 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_142559",
+    "created_at": "20260210_163348",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/microsoft__swin-tiny-patch4-window7-224",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
@@ -40,15 +40,13 @@
     "root_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "subfolder_code_included": true,
     "subfolder_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "processor_use_fast": true
   }

models/microsoft__swin-tiny-patch4-window7-224/ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

models/microsoft__swin-tiny-patch4-window7-224/ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

models/microsoft__swin-tiny-patch4-window7-224/ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

models/microsoft__swin-tiny-patch4-window7-224/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e6a66b3ed3993cd9e93967288f9019ef2c355abdeacb5db056bd9cc5192b2624
+oid sha256:29c4e394ff3e53d723a82b9911acfd072ca4259b7ac667e87eadee77b645ee84
 size 111128348

models/timm__densenet121.tv_in1k/config.json

@@ -24,7 +24,7 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_142559",
+    "created_at": "20260210_163348",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/timm__densenet121.tv_in1k",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
@@ -40,15 +40,13 @@
     "root_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "subfolder_code_included": true,
     "subfolder_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "processor_use_fast": false
   }

models/timm__densenet121.tv_in1k/ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

models/timm__densenet121.tv_in1k/ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

models/timm__densenet121.tv_in1k/ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

models/timm__densenet121.tv_in1k/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eda54a9ead50e30b80b8b0e60e9024149fd0cdeada25ea7023aa27333235090f
+oid sha256:ebc8977157008d63a00e318686b5b2fa763f2ece83748f3701671ee629ab70c9
 size 29293620

models/torchvision__densenet121/config.json

@@ -24,7 +24,7 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_142559",
+    "created_at": "20260210_163348",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/torchvision__densenet121",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
@@ -40,15 +40,13 @@
     "root_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "subfolder_code_included": true,
     "subfolder_code_files": [
       "ds_proc.py",
       "ds_model.py",
-      "ds_cfg.py",
-      "ds_meta.py"
+      "ds_cfg.py"
     ],
     "processor_use_fast": false
   }

models/torchvision__densenet121/ds_cfg.py

@@ -1,6 +1,149 @@
 from transformers import PretrainedConfig
+from typing import Literal, Any
+
+# ============================================================
+# Backbone whitelist + meta registry
+# ============================================================
+
+BackboneID = Literal[
+    "google/vit-base-patch16-224",
+    "microsoft/swin-tiny-patch4-window7-224",
+    "microsoft/resnet-50",
+    "google/efficientnet-b0",
+    "timm/densenet121.tv_in1k",
+    "torchvision/densenet121",
+]
+
+# ============================================================
+# 2) Backbone metadata registry (Feature dim/rule/unfreeze rule)
+# 2) 백본 메타 레지스트리 (feature dim/rule/unfreeze rule 고정)
+# ============================================================
+# This table is the single source of truth for feature extraction and fine-tuning rules per backbone.
+# 이 테이블은 backbone별 feature 추출 및 미세조정 규칙의 단일 기준(source of truth)입니다.
+#
+# The key type is BackboneID to ensure meta keys never drift from the whitelist.
+# 키 타입을 BackboneID로 고정하여 메타 키가 화이트리스트와 어긋나지 않게 합니다.
+BACKBONE_META: dict[BackboneID, dict[str, Any]] = {
+    # -------------------------
+    # Transformers (ViT/Swin)
+    # -------------------------
+    # These backbones come from transformers and typically output hidden states and/or pooler outputs.
+    # 이 백본들은 transformers 계열이며 hidden states와 pooler 출력 등을 제공합니다.
+
+    "google/vit-base-patch16-224": {
+        # type indicates which loading/forward/extraction pathway the model code should use.
+        # type은 모델 코드가 어떤 로딩/forward/feature 추출 경로를 사용할지 결정합니다.
+        "type": "vit",
+
+        # feat_dim is the feature vector dimension consumed by the MLP head.
+        # feat_dim은 MLP head가 입력으로 받는 feature 벡터 차원입니다.
+        "feat_dim": 768,
+
+        # feat_rule defines how to get a (B, feat_dim) tensor from backbone outputs.
+        # feat_rule은 backbone 출력에서 (B, feat_dim) 텐서를 얻는 규칙을 정의합니다.
+        "feat_rule": "cls",  # Use last_hidden_state[:, 0, :] as CLS token embedding.
+                            # last_hidden_state[:, 0, :]를 CLS 토큰 임베딩으로 사용합니다.
+
+        # unfreeze defines the policy to unfreeze layers during stage2 fine-tuning.
+        # unfreeze는 stage2 미세조정에서 어떤 레이어를 풀지 정책을 정의합니다.
+        "unfreeze": "last_n",  # Unfreeze the last n encoder blocks.
+                               # encoder 블록의 마지막 n개를 unfreeze 합니다.
+
+        # has_bn indicates whether BatchNorm exists and should be handled carefully when freezing.
+        # has_bn은 BatchNorm 존재 여부이며 freeze 시 특별 취급이 필요한지 판단에 사용합니다.
+        "has_bn": False,
+    },
+
+    "microsoft/swin-tiny-patch4-window7-224": {
+        # This backbone is a Swin Transformer, which may or may not provide a pooler output depending on implementation.
+        # 이 백본은 Swin Transformer이며 구현에 따라 pooler output 제공 여부가 달라질 수 있습니다.
+        "type": "swin",
+        "feat_dim": 768,
+
+        # Prefer pooler output if available, otherwise fall back to mean pooling.
+        # pooler가 있으면 우선 사용하고, 없으면 mean pooling으로 대체합니다.
+        "feat_rule": "pool_or_mean",
+
+        # Unfreeze strategy is aligned with transformer-style encoder blocks.
+        # unfreeze 전략은 transformer 계열 encoder 블록 기준으로 맞춥니다.
+        "unfreeze": "last_n",
+        "has_bn": False,
+    },
+
+    # -------------------------
+    # Transformers (CNNs)
+    # -------------------------
+    # These backbones are CNNs exposed via transformers, usually producing pooled feature vectors or feature maps.
+    # 이 백본들은 transformers로 노출된 CNN이며 pooled feature 또는 feature map을 제공합니다.
+
+    "microsoft/resnet-50": {
+        # This entry assumes a transformers-compatible ResNet that can expose pooler or a final feature map.
+        # 이 항목은 transformers 호환 ResNet이 pooler 또는 최종 feature map을 제공할 수 있다고 가정합니다.
+        "type": "resnet",
+        "feat_dim": 2048,
+
+        # Use pooler output if the model provides it, otherwise apply global average pooling (GAP).
+        # pooler가 있으면 사용하고, 없으면 global average pooling(GAP)을 적용합니다.
+        "feat_rule": "pool_or_gap",
+
+        # CNN unfreeze policy can still be expressed as "last_n" at a block/stage granularity in your model code.
+        # CNN도 모델 코드에서 block/stage 단위로 last_n 정책을 적용할 수 있습니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    "google/efficientnet-b0": {
+        # This entry assumes a transformers-compatible EfficientNet that exposes pooled features or a final feature map.
+        # 이 항목은 transformers 호환 EfficientNet이 pooled feature 또는 최종 feature map을 제공한다고 가정합니다.
+        "type": "efficientnet",
+        "feat_dim": 1280,
+        "feat_rule": "pool_or_gap",
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # timm (DenseNet via HF Hub)
+    # -------------------------
+    # This backbone is loaded via timm using the "hf_hub:" prefix in your model loader.
+    # 이 백본은 모델 로더에서 timm의 "hf_hub:" 프리픽스를 사용해 로드합니다.
+    "timm/densenet121.tv_in1k": {
+        "type": "timm_densenet",
+
+        # DenseNet-121 final channel dimension is 1024 for the canonical architecture.
+        # DenseNet-121의 표준 아키텍처에서 최종 채널 차원은 1024입니다.
+        "feat_dim": 1024,
+
+        # timm forward_features typically returns a feature map that you then GAP to (B, C).
+        # timm의 forward_features는 보통 feature map을 반환하고 이후 GAP으로 (B, C)를 만듭니다.
+        "feat_rule": "timm_gap",
+
+        # DenseNet uses BatchNorm heavily, so freeze_bn behavior matters for stage1/stage2.
+        # DenseNet은 BatchNorm 사용이 많아 stage1/stage2에서 freeze_bn 처리가 중요합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+
+    # -------------------------
+    # torchvision (DenseNet direct)
+    # -------------------------
+    # This backbone is intended for torchvision-style loading and feature extraction, not transformers/timm.
+    # 이 백본은 transformers/timm이 아니라 torchvision 스타일 로딩 및 feature 추출을 대상으로 합니다.
+    "torchvision/densenet121": {
+        "type": "torchvision_densenet",
+        "feat_dim": 1024,
+
+        # torchvision DenseNet usually exposes .features and you apply GAP to obtain (B, C).
+        # torchvision DenseNet은 보통 .features를 노출하며 GAP으로 (B, C)를 얻습니다.
+        "feat_rule": "torchvision_densenet_gap",
+
+        # Unfreeze policy remains last_n, but the interpretation must match torchvision module naming.
+        # unfreeze 정책은 last_n을 유지하되, 해석은 torchvision 모듈 네이밍에 맞아야 합니다.
+        "unfreeze": "last_n",
+        "has_bn": True,
+    },
+}
 
-from ds_meta import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHeadConfig(PretrainedConfig):

models/torchvision__densenet121/ds_model.py

@@ -3,7 +3,7 @@
 
 # src/ds_model.py
 
-from typing import Optional, List
+from typing import Optional, List, Any, Dict
 
 import torch
 import torch.nn as nn
@@ -17,9 +17,60 @@ from transformers.modeling_outputs import ImageClassifierOutput
 # --- torchvision ---
 from torchvision import models as tv_models
 
-from ds_cfg import BackboneMLPHeadConfig
-from ds_meta import BACKBONE_META
-from mlp_head import MLPHead
+from ds_cfg import BackboneMLPHeadConfig, BACKBONE_META
+# from mlp_head import MLPHead
+
+class MLPHead(nn.Module):
+    """
+    간단한 2-layer MLP head.
+
+    Parameters
+    ----------
+    in_dim : int
+        backbone feature dim
+    num_labels : int
+        class count
+    bottleneck : int
+        hidden dim
+    p : float
+        dropout prob
+    """
+    def __init__(self, in_dim: int, num_labels: int, bottleneck: int = 256, p: float = 0.2):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dim, bottleneck)
+        self.act = nn.GELU()
+        self.drop = nn.Dropout(p)
+        self.fc2 = nn.Linear(bottleneck, num_labels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.fc2(self.drop(self.act(self.fc1(x))))
+
+# ------------------------------------------------------------
+# backbone_meta resolver
+# ------------------------------------------------------------
+def _resolve_backbone_meta(config: BackboneMLPHeadConfig, fallback_table: Dict[str, Dict[str, Any]] | None = None) -> Dict[str, Any]:
+    """
+    Resolve runtime backbone meta.
+
+    Priority:
+      1) config.backbone_meta (preferred; required for Hub runtime determinism)
+      2) fallback_table[config.backbone_name_or_path] (backward compatibility for local/dev)
+
+    Returns a dict with at least: type, feat_rule, feat_dim (and optional has_bn/unfreeze).
+    """
+    meta = getattr(config, "backbone_meta", None)
+    if isinstance(meta, dict) and len(meta) > 0:
+        return meta
+
+    bb = getattr(config, "backbone_name_or_path", None)
+    if fallback_table is not None and bb in fallback_table:
+        return fallback_table[bb]
+
+    raise ValueError(
+        "config.backbone_meta is missing/empty and no fallback meta is available. "
+        "Populate config.backbone_meta when saving to the Hub (single source of truth)."
+    )
+
 
 # ============================================================
 # (3) Model: backbone + MLP head
@@ -67,7 +118,9 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
 
         # Meta is a single source of truth for extraction and fine-tuning rules.
         # meta는 feature 추출 및 미세조정 규칙의 단일 기준입니다.
-        self._meta = BACKBONE_META[config.backbone_name_or_path]
+                # Resolve backbone meta from config (preferred) or fallback table (for backward compatibility).
+        # Prefer config.backbone_meta to keep Hub runtime self-contained.
+        self._meta = _resolve_backbone_meta(config, fallback_table=BACKBONE_META)
 
         # Backbone skeleton is always created without pretrained weights.
         # backbone skeleton은 항상 pretrained weight 없이 생성합니다.
@@ -105,7 +158,10 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
     def _build_backbone_skeleton(self, backbone_id: str) -> nn.Module:
         # Meta decides which loader path to use.
         # meta가 어떤 로더 경로를 사용할지 결정합니다.
-        meta = BACKBONE_META[backbone_id]
+        meta = self._meta if backbone_id == self.config.backbone_name_or_path else BACKBONE_META.get(backbone_id)
+        if meta is None:
+            raise KeyError(f"Unknown backbone_id={backbone_id}. Provide backbone_meta in config or extend BACKBONE_META.")
+
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -169,7 +225,7 @@ class BackboneWithMLPHeadForImageClassification(PreTrainedModel):
         from_pretrained() 이후 호출하면 체크포인트 가중치를 덮어쓰므로 절대 호출하면 안 됩니다.
         """
         bb = self.config.backbone_name_or_path
-        meta = BACKBONE_META[bb]
+        meta = self._meta
         t = meta["type"]
 
         if t == "timm_densenet":
@@ -394,7 +450,7 @@ def freeze_backbone(model: BackboneWithMLPHeadForImageClassification, freeze_bn:
     _set_requires_grad(model.backbone, False)
     _set_requires_grad(model.classifier, True)
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if freeze_bn and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -403,7 +459,7 @@ def finetune_train_mode(model: BackboneWithMLPHeadForImageClassification, keep_b
     # Stage2: train mode, optionally keeping BN layers in eval for stability.
     # stage2: train 모드로 두되 안정성을 위해 BN을 eval로 유지할 수 있습니다.
     model.train()
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if keep_bn_eval and meta.get("has_bn", False):
         set_bn_eval(model.backbone)
 
@@ -431,7 +487,7 @@ def unfreeze_last_stage(
     if n <= 0:
         return
 
-    meta = BACKBONE_META[model.config.backbone_name_or_path]
+    meta = getattr(model, "_meta", None) or getattr(model.config, "backbone_meta", None)
     if meta.get("unfreeze") != "last_n":
         raise RuntimeError(f"Unexpected unfreeze rule: {meta.get('unfreeze')} (expected 'last_n')")
 

models/torchvision__densenet121/ds_proc.py

@@ -16,7 +16,7 @@ from transformers import AutoImageProcessor, AutoConfig
 from transformers.image_processing_base import ImageProcessingMixin
 from transformers.utils.generic import TensorType
 
-from ds_meta import BackboneID, BACKBONE_META
+from ds_cfg import BackboneID, BACKBONE_META
 
 
 class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):

models/torchvision__densenet121/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e8f2bb78b2e777c1612bca3678fd638acbcba9ca4ff460616987ad3ad94dab19
+oid sha256:064a8e6356a4304a40f80f368a46d9b0ff4307b849cf6e458ab325d2fcfd9c63
 size 33394052