initial push

.idea/CSATv2.iml

@@ -2,7 +2,7 @@
 <module type="PYTHON_MODULE" version="4">
   <component name="NewModuleRootManager">
     <content url="file://$MODULE_DIR$" />
-    <orderEntry type="jdk" jdkName="Python 3.6" jdkType="Python SDK" />
+    <orderEntry type="jdk" jdkName="alpha_evolve" jdkType="Python SDK" />
     <orderEntry type="sourceFolder" forTests="false" />
   </component>
   <component name="PyDocumentationSettings">

.idea/misc.xml

@@ -3,5 +3,5 @@
   <component name="Black">
     <option name="sdkName" value="Python 3.6" />
   </component>
-  <component name="ProjectRootManager" version="2" project-jdk-name="Python 3.6" project-jdk-type="Python SDK" />
+  <component name="ProjectRootManager" version="2" project-jdk-name="alpha_evolve" project-jdk-type="Python SDK" />
 </project>

CSAT_ImageNet.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81369c70ebd59a707a20d4e0c354f92a7056b1ac61ce41d9765d97fad6c36a37
+size 12383563

CSAT_RCKD.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4f44c33e0f30446bb28f41a0ef27d4f6a4018df9975a69795f598bdf0944036
+size 12381963

CSAT_v2_ImageNet.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:104253a41517faf75704853c207082c20b31264a7e9566a9a1ca22a9e088a729
+size 44535575

ResNet18_RCKD.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:243dc88aa8390386ef558f4f1202de86c6b2c5132722d080bac0380d9dcefa01
+size 46833867

config.json

@@ -0,0 +1,18 @@
+{
+  "_name_or_path": "Hyunil/CSATv2",
+  "model_type": "csatv2",
+  "architectures": ["CSATv2ForImageClassification"],
+  "image_size": 512,
+  "num_channels": 3,
+  "num_labels": 1000,
+  "drop_path_rate": 0.0,
+  "head_init_scale": 1.0,
+
+  "torch_dtype": "float32",
+
+  // Auto 클래스들이 이 레포 안의 어떤 코드를 써야 하는지 알려주는 부분
+  "auto_map": {
+    "AutoConfig": "modeling_csatv2.CSATv2Config",
+    "AutoModelForImageClassification": "modeling_csatv2.CSATv2ForImageClassification"
+  }
+}

image_processor.json

@@ -0,0 +1,15 @@
+{
+  "feature_extractor_type": "ImageFeatureExtraction",
+
+  "do_resize": true,
+  "size": 512,
+
+  "do_center_crop": false,
+
+  "do_rescale": true,
+  "rescale_factor": 0.00392156862745098,
+
+  "do_normalize": true,
+  "image_mean": [0.485, 0.456, 0.406],
+  "image_std": [0.229, 0.224, 0.225]
+}

modeling_csatv2.py

@@ -0,0 +1,113 @@
+# modeling_csatv2.py
+#
+# Hugging Face Transformers용 CSATv2 래퍼
+# - Config: CSATv2Config
+# - Model: CSATv2ForImageClassification
+#
+# 사용 예:
+#   from transformers import AutoImageProcessor, AutoModelForImageClassification
+#   model = AutoModelForImageClassification.from_pretrained(
+#       "Hyunil/CSATv2", trust_remote_code=True
+#   )
+
+from typing import Optional, Union, Tuple
+
+import torch
+import torch.nn as nn
+
+from transformers import PreTrainedModel, PretrainedConfig
+from transformers.modeling_outputs import ImageClassifierOutput
+
+from .CSATv2 import CSATv2  # 네가 올린 백본 클래스 사용
+
+
+class CSATv2Config(PretrainedConfig):
+    model_type = "csatv2"
+
+    def __init__(
+        self,
+        image_size: int = 224,
+        num_channels: int = 3,
+        num_labels: int = 1000,
+        drop_path_rate: float = 0.0,
+        head_init_scale: float = 1.0,
+        **kwargs,
+    ):
+        """
+        HF가 사용할 설정 값들.
+        """
+        super().__init__(num_labels=num_labels, **kwargs)
+        self.image_size = image_size
+        self.num_channels = num_channels
+        self.drop_path_rate = drop_path_rate
+        self.head_init_scale = head_init_scale
+
+        # label 매핑이 안 들어오면 기본값 생성
+        if self.id2label is None or self.label2id is None:
+            self.id2label = {i: f"LABEL_{i}" for i in range(num_labels)}
+            self.label2id = {v: k for k, v in self.id2label.items()}
+
+
+class CSATv2ForImageClassification(PreTrainedModel):
+    """
+    Hugging Face용 ImageNet 분류 모델 래퍼
+    - backbone: CSATv2 (네가 구현한 모델)
+    - forward(pixel_values, labels=None)
+    """
+
+    config_class = CSATv2Config
+
+    def __init__(self, config: CSATv2Config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        # 네가 만든 CSATv2 백본을 그대로 사용
+        self.backbone = CSATv2(
+            img_size=config.image_size,
+            num_classes=config.num_labels,
+            drop_path_rate=config.drop_path_rate,
+            head_init_scale=config.head_init_scale,
+        )
+
+        # transformers 권장: 내부 가중치 등록 후 post_init 호출
+        self.post_init()
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor = None,
+        labels: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[ImageClassifierOutput, Tuple]:
+        """
+        Args:
+            pixel_values: (batch, 3, H, W), ImageNet 정규화까지 된 이미지
+            labels: (batch,) 0~999 class index
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You must provide pixel_values")
+
+        # CSATv2는 이미 logits를 반환함
+        logits = self.backbone(pixel_values)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(
+                logits.view(-1, self.num_labels),
+                labels.view(-1),
+            )
+
+        if not return_dict:
+            output = (logits,)
+            return ((loss,) + output) if loss is not None else output
+
+        return ImageClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=None,
+            attentions=None,
+        )

tar2bin.py

@@ -0,0 +1,23 @@
+import torch
+from collections import OrderedDict
+
+
+ckpt_path = "./CSAT_RCKD.pth.tar"
+ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+
+# 1) state_dict 꺼내기
+#   - 보통 {'state_dict': ...} 형태니까 먼저 이걸 시도하고,
+#   - 아니면 그냥 ckpt 전체가 state_dict인 경우도 있어서 fallback
+state_dict = ckpt.get("state_dict", ckpt)
+# 2) DataParallel 썼으면 key 앞에 'module.' 붙어있을 수 있어서 제거
+new_state_dict = OrderedDict()
+for k, v in state_dict.items():
+    if k.startswith("module."):
+        new_k = k[len("module."):]
+    else:
+        new_k = k
+    new_state_dict[new_k] = v
+
+# 3) HuggingFace 관례대로 파일명 저장
+torch.save(new_state_dict, "CSAT_RCKD.bin")
+print("saved to pytorch_model.bin")