Add/Update backbone checkpoints (count=6)

manifest_20260210_113951.json

@@ -0,0 +1,42 @@
+{
+  "timestamp": "20260210_113951",
+  "device": "mps",
+  "repo_id": "dsaint31/bb_mlp_224",
+  "revision": "main",
+  "tag": null,
+  "num_labels": 3,
+  "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"
+    }
+  ],
+  "processor_use_fast": false
+}

models/google__efficientnet-b0/config.json

@@ -24,16 +24,18 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_111913",
+    "created_at": "20260210_113951",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/google__efficientnet-b0",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
     "processor_class": "BackboneMLPHead224ImageProcessor",
     "backbone_id": "google/efficientnet-b0",
+    "meta_type": "efficientnet",
     "pretrained_injected": true,
     "policy_init_pretrained_forbidden": true,
     "policy_reload_no_pretrained_injection": true,
     "num_labels": 3,
-    "device": "mps"
+    "device": "mps",
+    "processor_use_fast": false
   }
 }

models/google__efficientnet-b0/ds_proc.py

@@ -35,13 +35,22 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
     3) Runtime objects are rebuilt at init/load time based on backbone meta.
        런타임 객체는 backbone meta에 따라 init/load 시점에 재구성됩니다.
+
+    4) For reproducibility, use_fast must be explicitly persisted and honored on load.
+       재현성을 위해 use_fast는 명시적으로 저장되고, 로드시 반드시 반영되어야 합니다.
     """
 
     # HF vision models conventionally expect "pixel_values" as the primary input key.
     # HF vision 모델은 관례적으로 입력 키로 "pixel_values"를 기대합니다.
     model_input_names = ["pixel_values"]
 
-    def __init__(self, backbone_name_or_path: BackboneID, is_training: bool = False, **kwargs):
+    def __init__(
+        self,
+        backbone_name_or_path: BackboneID,
+        is_training: bool = False,
+        use_fast: bool = False,
+        **kwargs,
+    ):
         # ImageProcessingMixin stores extra kwargs and manages auto_map metadata.
         # ImageProcessingMixin은 추가 kwargs를 저장하고 auto_map 메타를 관리합니다.
         super().__init__(**kwargs)
@@ -59,6 +68,10 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         self.backbone_name_or_path = backbone_name_or_path
         self.is_training = bool(is_training)
 
+        # Reproducibility switch for transformers processors.
+        # transformers processor의 fast/slow 선택을 재현 가능하게 고정합니다.
+        self.use_fast = bool(use_fast)
+
         # Runtime-only fields: must never be serialized.
         # 런타임 전용 필드: 절대 직렬화되면 안 됩니다.
         self._meta = None
@@ -109,7 +122,13 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # Default: transformers backbone delegates to its official AutoImageProcessor.
         # 기본: transformers 백본은 공식 AutoImageProcessor에 위임합니다.
-        self._delegate = AutoImageProcessor.from_pretrained(self.backbone_name_or_path)
+        #
+        # IMPORTANT:
+        # - use_fast는 transformers 기본값 변경에 흔들리지 않도록 반드시 명시적으로 전달합니다.
+        self._delegate = AutoImageProcessor.from_pretrained(
+            self.backbone_name_or_path,
+            use_fast=self.use_fast,
+        )
 
     @staticmethod
     def _build_timm_transform(*, backbone_id: str, is_training: bool):
@@ -197,6 +216,7 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         d["image_processor_type"] = self.__class__.__name__
         d["backbone_name_or_path"] = self.backbone_name_or_path
         d["is_training"] = self.is_training
+        d["use_fast"] = self.use_fast
 
         # Remove any runtime-only fields defensively.
         # 런타임 전용 필드는 보수적으로 제거합니다.
@@ -216,7 +236,14 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
             raise ValueError("preprocessor_config.json missing key: backbone_name_or_path")
 
         is_training = bool(image_processor_dict.get("is_training", False))
-        return cls(backbone_name_or_path=backbone, is_training=is_training, **kwargs)
+        use_fast = bool(image_processor_dict.get("use_fast", False))
+
+        return cls(
+            backbone_name_or_path=backbone,
+            is_training=is_training,
+            use_fast=use_fast,
+            **kwargs,
+        )
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
@@ -237,7 +264,12 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # is_training is runtime-only and should default to False for inference/serving.
         # is_training은 런타임 전용이며 추론/서빙 기본값은 False가 맞습니다.
-        return cls(backbone_name_or_path=backbone, is_training=False)
+        #
+        # IMPORTANT:
+        # - use_fast는 kwargs로 전달될 수 있으므로, 있으면 반영합니다.
+        use_fast = bool(kwargs.pop("use_fast", False))
+
+        return cls(backbone_name_or_path=backbone, is_training=False, use_fast=use_fast)
 
     # ============================================================
     # Call interface

models/google__efficientnet-b0/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:01add885928661dde020c6cc5adaf1e6b8d6684a9effa7ef2903ac8d03fc9a5c
+oid sha256:6b89c592a265df3c45219f864a9f40156e1d72c043aa284abccadd56a7ec41d2
 size 17558436

models/google__efficientnet-b0/preprocessor_config.json

@@ -4,5 +4,6 @@
   },
   "backbone_name_or_path": "google/efficientnet-b0",
   "image_processor_type": "BackboneMLPHead224ImageProcessor",
-  "is_training": false
+  "is_training": false,
+  "use_fast": false
 }

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

@@ -24,16 +24,18 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_111913",
+    "created_at": "20260210_113951",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/google__vit-base-patch16-224",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
     "processor_class": "BackboneMLPHead224ImageProcessor",
     "backbone_id": "google/vit-base-patch16-224",
+    "meta_type": "vit",
     "pretrained_injected": true,
     "policy_init_pretrained_forbidden": true,
     "policy_reload_no_pretrained_injection": true,
     "num_labels": 3,
-    "device": "mps"
+    "device": "mps",
+    "processor_use_fast": false
   }
 }

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

@@ -35,13 +35,22 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
     3) Runtime objects are rebuilt at init/load time based on backbone meta.
        런타임 객체는 backbone meta에 따라 init/load 시점에 재구성됩니다.
+
+    4) For reproducibility, use_fast must be explicitly persisted and honored on load.
+       재현성을 위해 use_fast는 명시적으로 저장되고, 로드시 반드시 반영되어야 합니다.
     """
 
     # HF vision models conventionally expect "pixel_values" as the primary input key.
     # HF vision 모델은 관례적으로 입력 키로 "pixel_values"를 기대합니다.
     model_input_names = ["pixel_values"]
 
-    def __init__(self, backbone_name_or_path: BackboneID, is_training: bool = False, **kwargs):
+    def __init__(
+        self,
+        backbone_name_or_path: BackboneID,
+        is_training: bool = False,
+        use_fast: bool = False,
+        **kwargs,
+    ):
         # ImageProcessingMixin stores extra kwargs and manages auto_map metadata.
         # ImageProcessingMixin은 추가 kwargs를 저장하고 auto_map 메타를 관리합니다.
         super().__init__(**kwargs)
@@ -59,6 +68,10 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         self.backbone_name_or_path = backbone_name_or_path
         self.is_training = bool(is_training)
 
+        # Reproducibility switch for transformers processors.
+        # transformers processor의 fast/slow 선택을 재현 가능하게 고정합니다.
+        self.use_fast = bool(use_fast)
+
         # Runtime-only fields: must never be serialized.
         # 런타임 전용 필드: 절대 직렬화되면 안 됩니다.
         self._meta = None
@@ -109,7 +122,13 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # Default: transformers backbone delegates to its official AutoImageProcessor.
         # 기본: transformers 백본은 공식 AutoImageProcessor에 위임합니다.
-        self._delegate = AutoImageProcessor.from_pretrained(self.backbone_name_or_path)
+        #
+        # IMPORTANT:
+        # - use_fast는 transformers 기본값 변경에 흔들리지 않도록 반드시 명시적으로 전달합니다.
+        self._delegate = AutoImageProcessor.from_pretrained(
+            self.backbone_name_or_path,
+            use_fast=self.use_fast,
+        )
 
     @staticmethod
     def _build_timm_transform(*, backbone_id: str, is_training: bool):
@@ -197,6 +216,7 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         d["image_processor_type"] = self.__class__.__name__
         d["backbone_name_or_path"] = self.backbone_name_or_path
         d["is_training"] = self.is_training
+        d["use_fast"] = self.use_fast
 
         # Remove any runtime-only fields defensively.
         # 런타임 전용 필드는 보수적으로 제거합니다.
@@ -216,7 +236,14 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
             raise ValueError("preprocessor_config.json missing key: backbone_name_or_path")
 
         is_training = bool(image_processor_dict.get("is_training", False))
-        return cls(backbone_name_or_path=backbone, is_training=is_training, **kwargs)
+        use_fast = bool(image_processor_dict.get("use_fast", False))
+
+        return cls(
+            backbone_name_or_path=backbone,
+            is_training=is_training,
+            use_fast=use_fast,
+            **kwargs,
+        )
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
@@ -237,7 +264,12 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # is_training is runtime-only and should default to False for inference/serving.
         # is_training은 런타임 전용이며 추론/서빙 기본값은 False가 맞습니다.
-        return cls(backbone_name_or_path=backbone, is_training=False)
+        #
+        # IMPORTANT:
+        # - use_fast는 kwargs로 전달될 수 있으므로, 있으면 반영합니다.
+        use_fast = bool(kwargs.pop("use_fast", False))
+
+        return cls(backbone_name_or_path=backbone, is_training=False, use_fast=use_fast)
 
     # ============================================================
     # Call interface

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

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5ca403809fcb589b9b176f4712c6f6aa3c31d2cb7465aaaa6cd4391d46947e8e
+oid sha256:0ca9d22f81f9858ff4c0734428a65604dd178d6ceb937d5759fcef7e15aba8bc
 size 346372132

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

@@ -4,5 +4,6 @@
   },
   "backbone_name_or_path": "google/vit-base-patch16-224",
   "image_processor_type": "BackboneMLPHead224ImageProcessor",
-  "is_training": false
+  "is_training": false,
+  "use_fast": false
 }

models/microsoft__resnet-50/config.json

@@ -24,16 +24,18 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_111913",
+    "created_at": "20260210_113951",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/microsoft__resnet-50",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
     "processor_class": "BackboneMLPHead224ImageProcessor",
     "backbone_id": "microsoft/resnet-50",
+    "meta_type": "resnet",
     "pretrained_injected": true,
     "policy_init_pretrained_forbidden": true,
     "policy_reload_no_pretrained_injection": true,
     "num_labels": 3,
-    "device": "mps"
+    "device": "mps",
+    "processor_use_fast": false
   }
 }

models/microsoft__resnet-50/ds_proc.py

@@ -35,13 +35,22 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
     3) Runtime objects are rebuilt at init/load time based on backbone meta.
        런타임 객체는 backbone meta에 따라 init/load 시점에 재구성됩니다.
+
+    4) For reproducibility, use_fast must be explicitly persisted and honored on load.
+       재현성을 위해 use_fast는 명시적으로 저장되고, 로드시 반드시 반영되어야 합니다.
     """
 
     # HF vision models conventionally expect "pixel_values" as the primary input key.
     # HF vision 모델은 관례적으로 입력 키로 "pixel_values"를 기대합니다.
     model_input_names = ["pixel_values"]
 
-    def __init__(self, backbone_name_or_path: BackboneID, is_training: bool = False, **kwargs):
+    def __init__(
+        self,
+        backbone_name_or_path: BackboneID,
+        is_training: bool = False,
+        use_fast: bool = False,
+        **kwargs,
+    ):
         # ImageProcessingMixin stores extra kwargs and manages auto_map metadata.
         # ImageProcessingMixin은 추가 kwargs를 저장하고 auto_map 메타를 관리합니다.
         super().__init__(**kwargs)
@@ -59,6 +68,10 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         self.backbone_name_or_path = backbone_name_or_path
         self.is_training = bool(is_training)
 
+        # Reproducibility switch for transformers processors.
+        # transformers processor의 fast/slow 선택을 재현 가능하게 고정합니다.
+        self.use_fast = bool(use_fast)
+
         # Runtime-only fields: must never be serialized.
         # 런타임 전용 필드: 절대 직렬화되면 안 됩니다.
         self._meta = None
@@ -109,7 +122,13 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # Default: transformers backbone delegates to its official AutoImageProcessor.
         # 기본: transformers 백본은 공식 AutoImageProcessor에 위임합니다.
-        self._delegate = AutoImageProcessor.from_pretrained(self.backbone_name_or_path)
+        #
+        # IMPORTANT:
+        # - use_fast는 transformers 기본값 변경에 흔들리지 않도록 반드시 명시적으로 전달합니다.
+        self._delegate = AutoImageProcessor.from_pretrained(
+            self.backbone_name_or_path,
+            use_fast=self.use_fast,
+        )
 
     @staticmethod
     def _build_timm_transform(*, backbone_id: str, is_training: bool):
@@ -197,6 +216,7 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         d["image_processor_type"] = self.__class__.__name__
         d["backbone_name_or_path"] = self.backbone_name_or_path
         d["is_training"] = self.is_training
+        d["use_fast"] = self.use_fast
 
         # Remove any runtime-only fields defensively.
         # 런타임 전용 필드는 보수적으로 제거합니다.
@@ -216,7 +236,14 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
             raise ValueError("preprocessor_config.json missing key: backbone_name_or_path")
 
         is_training = bool(image_processor_dict.get("is_training", False))
-        return cls(backbone_name_or_path=backbone, is_training=is_training, **kwargs)
+        use_fast = bool(image_processor_dict.get("use_fast", False))
+
+        return cls(
+            backbone_name_or_path=backbone,
+            is_training=is_training,
+            use_fast=use_fast,
+            **kwargs,
+        )
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
@@ -237,7 +264,12 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # is_training is runtime-only and should default to False for inference/serving.
         # is_training은 런타임 전용이며 추론/서빙 기본값은 False가 맞습니다.
-        return cls(backbone_name_or_path=backbone, is_training=False)
+        #
+        # IMPORTANT:
+        # - use_fast는 kwargs로 전달될 수 있으므로, 있으면 반영합니다.
+        use_fast = bool(kwargs.pop("use_fast", False))
+
+        return cls(backbone_name_or_path=backbone, is_training=False, use_fast=use_fast)
 
     # ============================================================
     # Call interface

models/microsoft__resnet-50/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:229f77319aec60efa8fff986455e304b9f4f2d46440550b69490da095a11d170
+oid sha256:5ee43d5448edadf03eb7a01e7ef9248ebf4713d31babfb6f59985b0401b5d1f2
 size 96388660

models/microsoft__resnet-50/preprocessor_config.json

@@ -4,5 +4,6 @@
   },
   "backbone_name_or_path": "microsoft/resnet-50",
   "image_processor_type": "BackboneMLPHead224ImageProcessor",
-  "is_training": false
+  "is_training": false,
+  "use_fast": false
 }

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

@@ -24,16 +24,18 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_111913",
+    "created_at": "20260210_113951",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/microsoft__swin-tiny-patch4-window7-224",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
     "processor_class": "BackboneMLPHead224ImageProcessor",
     "backbone_id": "microsoft/swin-tiny-patch4-window7-224",
+    "meta_type": "swin",
     "pretrained_injected": true,
     "policy_init_pretrained_forbidden": true,
     "policy_reload_no_pretrained_injection": true,
     "num_labels": 3,
-    "device": "mps"
+    "device": "mps",
+    "processor_use_fast": false
   }
 }

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

@@ -35,13 +35,22 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
     3) Runtime objects are rebuilt at init/load time based on backbone meta.
        런타임 객체는 backbone meta에 따라 init/load 시점에 재구성됩니다.
+
+    4) For reproducibility, use_fast must be explicitly persisted and honored on load.
+       재현성을 위해 use_fast는 명시적으로 저장되고, 로드시 반드시 반영되어야 합니다.
     """
 
     # HF vision models conventionally expect "pixel_values" as the primary input key.
     # HF vision 모델은 관례적으로 입력 키로 "pixel_values"를 기대합니다.
     model_input_names = ["pixel_values"]
 
-    def __init__(self, backbone_name_or_path: BackboneID, is_training: bool = False, **kwargs):
+    def __init__(
+        self,
+        backbone_name_or_path: BackboneID,
+        is_training: bool = False,
+        use_fast: bool = False,
+        **kwargs,
+    ):
         # ImageProcessingMixin stores extra kwargs and manages auto_map metadata.
         # ImageProcessingMixin은 추가 kwargs를 저장하고 auto_map 메타를 관리합니다.
         super().__init__(**kwargs)
@@ -59,6 +68,10 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         self.backbone_name_or_path = backbone_name_or_path
         self.is_training = bool(is_training)
 
+        # Reproducibility switch for transformers processors.
+        # transformers processor의 fast/slow 선택을 재현 가능하게 고정합니다.
+        self.use_fast = bool(use_fast)
+
         # Runtime-only fields: must never be serialized.
         # 런타임 전용 필드: 절대 직렬화되면 안 됩니다.
         self._meta = None
@@ -109,7 +122,13 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # Default: transformers backbone delegates to its official AutoImageProcessor.
         # 기본: transformers 백본은 공식 AutoImageProcessor에 위임합니다.
-        self._delegate = AutoImageProcessor.from_pretrained(self.backbone_name_or_path)
+        #
+        # IMPORTANT:
+        # - use_fast는 transformers 기본값 변경에 흔들리지 않도록 반드시 명시적으로 전달합니다.
+        self._delegate = AutoImageProcessor.from_pretrained(
+            self.backbone_name_or_path,
+            use_fast=self.use_fast,
+        )
 
     @staticmethod
     def _build_timm_transform(*, backbone_id: str, is_training: bool):
@@ -197,6 +216,7 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         d["image_processor_type"] = self.__class__.__name__
         d["backbone_name_or_path"] = self.backbone_name_or_path
         d["is_training"] = self.is_training
+        d["use_fast"] = self.use_fast
 
         # Remove any runtime-only fields defensively.
         # 런타임 전용 필드는 보수적으로 제거합니다.
@@ -216,7 +236,14 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
             raise ValueError("preprocessor_config.json missing key: backbone_name_or_path")
 
         is_training = bool(image_processor_dict.get("is_training", False))
-        return cls(backbone_name_or_path=backbone, is_training=is_training, **kwargs)
+        use_fast = bool(image_processor_dict.get("use_fast", False))
+
+        return cls(
+            backbone_name_or_path=backbone,
+            is_training=is_training,
+            use_fast=use_fast,
+            **kwargs,
+        )
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
@@ -237,7 +264,12 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # is_training is runtime-only and should default to False for inference/serving.
         # is_training은 런타임 전용이며 추론/서빙 기본값은 False가 맞습니다.
-        return cls(backbone_name_or_path=backbone, is_training=False)
+        #
+        # IMPORTANT:
+        # - use_fast는 kwargs로 전달될 수 있으므로, 있으면 반영합니다.
+        use_fast = bool(kwargs.pop("use_fast", False))
+
+        return cls(backbone_name_or_path=backbone, is_training=False, use_fast=use_fast)
 
     # ============================================================
     # Call interface

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

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:61d4b000d39bc1301116050226ee05a3cc3851359f4dc38febbb5f397fbb844f
+oid sha256:52509f874cadf1f0a3889ff8bd4e6de809877596a215138307c1c66ba6e8055e
 size 111128348

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

@@ -4,5 +4,6 @@
   },
   "backbone_name_or_path": "microsoft/swin-tiny-patch4-window7-224",
   "image_processor_type": "BackboneMLPHead224ImageProcessor",
-  "is_training": false
+  "is_training": false,
+  "use_fast": false
 }

models/timm__densenet121.tv_in1k/config.json

@@ -24,16 +24,18 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_111913",
+    "created_at": "20260210_113951",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/timm__densenet121.tv_in1k",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
     "processor_class": "BackboneMLPHead224ImageProcessor",
     "backbone_id": "timm/densenet121.tv_in1k",
+    "meta_type": "timm_densenet",
     "pretrained_injected": true,
     "policy_init_pretrained_forbidden": true,
     "policy_reload_no_pretrained_injection": true,
     "num_labels": 3,
-    "device": "mps"
+    "device": "mps",
+    "processor_use_fast": false
   }
 }

models/timm__densenet121.tv_in1k/ds_proc.py

@@ -35,13 +35,22 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
     3) Runtime objects are rebuilt at init/load time based on backbone meta.
        런타임 객체는 backbone meta에 따라 init/load 시점에 재구성됩니다.
+
+    4) For reproducibility, use_fast must be explicitly persisted and honored on load.
+       재현성을 위해 use_fast는 명시적으로 저장되고, 로드시 반드시 반영되어야 합니다.
     """
 
     # HF vision models conventionally expect "pixel_values" as the primary input key.
     # HF vision 모델은 관례적으로 입력 키로 "pixel_values"를 기대합니다.
     model_input_names = ["pixel_values"]
 
-    def __init__(self, backbone_name_or_path: BackboneID, is_training: bool = False, **kwargs):
+    def __init__(
+        self,
+        backbone_name_or_path: BackboneID,
+        is_training: bool = False,
+        use_fast: bool = False,
+        **kwargs,
+    ):
         # ImageProcessingMixin stores extra kwargs and manages auto_map metadata.
         # ImageProcessingMixin은 추가 kwargs를 저장하고 auto_map 메타를 관리합니다.
         super().__init__(**kwargs)
@@ -59,6 +68,10 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         self.backbone_name_or_path = backbone_name_or_path
         self.is_training = bool(is_training)
 
+        # Reproducibility switch for transformers processors.
+        # transformers processor의 fast/slow 선택을 재현 가능하게 고정합니다.
+        self.use_fast = bool(use_fast)
+
         # Runtime-only fields: must never be serialized.
         # 런타임 전용 필드: 절대 직렬화되면 안 됩니다.
         self._meta = None
@@ -109,7 +122,13 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # Default: transformers backbone delegates to its official AutoImageProcessor.
         # 기본: transformers 백본은 공식 AutoImageProcessor에 위임합니다.
-        self._delegate = AutoImageProcessor.from_pretrained(self.backbone_name_or_path)
+        #
+        # IMPORTANT:
+        # - use_fast는 transformers 기본값 변경에 흔들리지 않도록 반드시 명시적으로 전달합니다.
+        self._delegate = AutoImageProcessor.from_pretrained(
+            self.backbone_name_or_path,
+            use_fast=self.use_fast,
+        )
 
     @staticmethod
     def _build_timm_transform(*, backbone_id: str, is_training: bool):
@@ -197,6 +216,7 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         d["image_processor_type"] = self.__class__.__name__
         d["backbone_name_or_path"] = self.backbone_name_or_path
         d["is_training"] = self.is_training
+        d["use_fast"] = self.use_fast
 
         # Remove any runtime-only fields defensively.
         # 런타임 전용 필드는 보수적으로 제거합니다.
@@ -216,7 +236,14 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
             raise ValueError("preprocessor_config.json missing key: backbone_name_or_path")
 
         is_training = bool(image_processor_dict.get("is_training", False))
-        return cls(backbone_name_or_path=backbone, is_training=is_training, **kwargs)
+        use_fast = bool(image_processor_dict.get("use_fast", False))
+
+        return cls(
+            backbone_name_or_path=backbone,
+            is_training=is_training,
+            use_fast=use_fast,
+            **kwargs,
+        )
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
@@ -237,7 +264,12 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # is_training is runtime-only and should default to False for inference/serving.
         # is_training은 런타임 전용이며 추론/서빙 기본값은 False가 맞습니다.
-        return cls(backbone_name_or_path=backbone, is_training=False)
+        #
+        # IMPORTANT:
+        # - use_fast는 kwargs로 전달될 수 있으므로, 있으면 반영합니다.
+        use_fast = bool(kwargs.pop("use_fast", False))
+
+        return cls(backbone_name_or_path=backbone, is_training=False, use_fast=use_fast)
 
     # ============================================================
     # Call interface

models/timm__densenet121.tv_in1k/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:53c62717fc23ff4a775482d0f9ca8c730a2c04681a9eeefa01864847d713aa9f
+oid sha256:cbb8c09fbf696346234b700625e6649405512e1c8e6abbc22c15f5edf4da6212
 size 29293620

models/timm__densenet121.tv_in1k/preprocessor_config.json

@@ -4,5 +4,6 @@
   },
   "backbone_name_or_path": "timm/densenet121.tv_in1k",
   "image_processor_type": "BackboneMLPHead224ImageProcessor",
-  "is_training": false
+  "is_training": false,
+  "use_fast": false
 }

models/torchvision__densenet121/config.json

@@ -24,16 +24,18 @@
   "num_labels": 3,
   "transformers_version": "5.1.0",
   "ds_provenance": {
-    "created_at": "20260210_111913",
+    "created_at": "20260210_113951",
     "repo_id": "dsaint31/bb_mlp_224",
     "subdir": "models/torchvision__densenet121",
     "wrapper_class": "BackboneWithMLPHeadForImageClassification",
     "processor_class": "BackboneMLPHead224ImageProcessor",
     "backbone_id": "torchvision/densenet121",
+    "meta_type": "torchvision_densenet",
     "pretrained_injected": true,
     "policy_init_pretrained_forbidden": true,
     "policy_reload_no_pretrained_injection": true,
     "num_labels": 3,
-    "device": "mps"
+    "device": "mps",
+    "processor_use_fast": false
   }
 }

models/torchvision__densenet121/ds_proc.py

@@ -35,13 +35,22 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
     3) Runtime objects are rebuilt at init/load time based on backbone meta.
        런타임 객체는 backbone meta에 따라 init/load 시점에 재구성됩니다.
+
+    4) For reproducibility, use_fast must be explicitly persisted and honored on load.
+       재현성을 위해 use_fast는 명시적으로 저장되고, 로드시 반드시 반영되어야 합니다.
     """
 
     # HF vision models conventionally expect "pixel_values" as the primary input key.
     # HF vision 모델은 관례적으로 입력 키로 "pixel_values"를 기대합니다.
     model_input_names = ["pixel_values"]
 
-    def __init__(self, backbone_name_or_path: BackboneID, is_training: bool = False, **kwargs):
+    def __init__(
+        self,
+        backbone_name_or_path: BackboneID,
+        is_training: bool = False,
+        use_fast: bool = False,
+        **kwargs,
+    ):
         # ImageProcessingMixin stores extra kwargs and manages auto_map metadata.
         # ImageProcessingMixin은 추가 kwargs를 저장하고 auto_map 메타를 관리합니다.
         super().__init__(**kwargs)
@@ -59,6 +68,10 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         self.backbone_name_or_path = backbone_name_or_path
         self.is_training = bool(is_training)
 
+        # Reproducibility switch for transformers processors.
+        # transformers processor의 fast/slow 선택을 재현 가능하게 고정합니다.
+        self.use_fast = bool(use_fast)
+
         # Runtime-only fields: must never be serialized.
         # 런타임 전용 필드: 절대 직렬화되면 안 됩니다.
         self._meta = None
@@ -109,7 +122,13 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # Default: transformers backbone delegates to its official AutoImageProcessor.
         # 기본: transformers 백본은 공식 AutoImageProcessor에 위임합니다.
-        self._delegate = AutoImageProcessor.from_pretrained(self.backbone_name_or_path)
+        #
+        # IMPORTANT:
+        # - use_fast는 transformers 기본값 변경에 흔들리지 않도록 반드시 명시적으로 전달합니다.
+        self._delegate = AutoImageProcessor.from_pretrained(
+            self.backbone_name_or_path,
+            use_fast=self.use_fast,
+        )
 
     @staticmethod
     def _build_timm_transform(*, backbone_id: str, is_training: bool):
@@ -197,6 +216,7 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
         d["image_processor_type"] = self.__class__.__name__
         d["backbone_name_or_path"] = self.backbone_name_or_path
         d["is_training"] = self.is_training
+        d["use_fast"] = self.use_fast
 
         # Remove any runtime-only fields defensively.
         # 런타임 전용 필드는 보수적으로 제거합니다.
@@ -216,7 +236,14 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
             raise ValueError("preprocessor_config.json missing key: backbone_name_or_path")
 
         is_training = bool(image_processor_dict.get("is_training", False))
-        return cls(backbone_name_or_path=backbone, is_training=is_training, **kwargs)
+        use_fast = bool(image_processor_dict.get("use_fast", False))
+
+        return cls(
+            backbone_name_or_path=backbone,
+            is_training=is_training,
+            use_fast=use_fast,
+            **kwargs,
+        )
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
@@ -237,7 +264,12 @@ class BackboneMLPHead224ImageProcessor(ImageProcessingMixin):
 
         # is_training is runtime-only and should default to False for inference/serving.
         # is_training은 런타임 전용이며 추론/서빙 기본값은 False가 맞습니다.
-        return cls(backbone_name_or_path=backbone, is_training=False)
+        #
+        # IMPORTANT:
+        # - use_fast는 kwargs로 전달될 수 있으므로, 있으면 반영합니다.
+        use_fast = bool(kwargs.pop("use_fast", False))
+
+        return cls(backbone_name_or_path=backbone, is_training=False, use_fast=use_fast)
 
     # ============================================================
     # Call interface

models/torchvision__densenet121/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e8431398f3461657aea5362b326d351fcdd4c5c009bf126765d0dcd60f94aa72
+oid sha256:e5c4ad5ea10f956a0ead9687834bf22225e978714b4d059980dd3013b0a190ad
 size 33394052

models/torchvision__densenet121/preprocessor_config.json

@@ -4,5 +4,6 @@
   },
   "backbone_name_or_path": "torchvision/densenet121",
   "image_processor_type": "BackboneMLPHead224ImageProcessor",
-  "is_training": false
+  "is_training": false,
+  "use_fast": false
 }