Release custom MNIST model

__init__.py

@@ -18,7 +18,8 @@ AutoModelForImageClassification.register(
 try:
     AutoImageProcessor.register(
         MyMNISTConfig,
-        MyMNISTImageProcessor,
+        # MyMNISTImageProcessor, # To avoid FutureWarning
+        slow_image_processor_class=MyMNISTImageProcessor,
         exist_ok=True
     )
 except TypeError:

examples/eval.py

@@ -0,0 +1,197 @@
+# hf_custom_proj/examples/eval.py
+"""
+로컬(dist) 또는 Hugging Face Hub에서 모델/프로세서를 로드해
+MNIST test set을 평가(evaluate)하는 스크립트.
+
+사용법
+- 로컬(dist)에서 평가:
+  python -m examples.eval --source local --path dist/my-mnist-hf
+
+- Hub에서 평가(공개 repo):
+  python -m examples.eval --source hub --path YOUR_ID/my-mnist-hf
+
+- Hub에서 평가(private repo 또는 토큰 강제):
+  export HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxx
+  python -m examples.eval --source hub --path YOUR_ID/my-mnist-hf --use-token
+
+옵션
+- 일부만 빠르게:
+  python -m examples.eval --source local --path dist/my-mnist-hf --limit 1000
+"""
+
+from __future__ import annotations
+
+import argparse
+import os
+from pathlib import Path
+from typing import Tuple
+
+import torch
+from torch.utils.data import DataLoader
+from torchvision.datasets import MNIST
+from transformers import AutoImageProcessor, AutoModelForImageClassification
+
+
+def _resolve_token(use_token: bool) -> str | None:
+    if not use_token:
+        return None
+    token = os.environ.get("HF_TOKEN")
+    if not token:
+        raise RuntimeError(
+            "use_token=True 인데 환경변수 HF_TOKEN이 없습니다.\n"
+            "예) export HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxx"
+        )
+    return token
+
+
+def _validate_local_dir(path: str) -> None:
+    p = Path(path)
+    if not p.exists():
+        raise FileNotFoundError(f"Local path not found: {p.resolve()}")
+    if not p.is_dir():
+        raise NotADirectoryError(f"Local path is not a directory: {p.resolve()}")
+
+
+@torch.no_grad()
+def evaluate(
+    *,
+    source: str,
+    path: str,
+    data_dir: str,
+    batch_size: int,
+    num_workers: int,
+    limit: int,
+    use_token: bool,
+    device: str,
+) -> None:
+    """
+    MNIST test set 평가: 평균 loss + accuracy 출력
+    """
+    source = source.lower().strip()
+    if source not in ("local", "hub"):
+        raise ValueError("--source must be one of {'local', 'hub'}")
+
+    if source == "local":
+        _validate_local_dir(path)
+
+    token = _resolve_token(use_token) if source == "hub" else None
+
+    # processor / model 로드
+    processor = AutoImageProcessor.from_pretrained(
+        path,
+        trust_remote_code=True,
+        token=token,
+    )
+    model = AutoModelForImageClassification.from_pretrained(
+        path,
+        trust_remote_code=True,
+        token=token,
+    )
+    model.eval()
+
+    # device 설정
+    if device == "auto":
+        dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    else:
+        dev = torch.device(device)
+    model.to(dev)
+
+    # MNIST test set
+    ds = MNIST(root=data_dir, train=False, download=True)
+
+    # limit 적용(원하면 일부만)
+    if limit > 0:
+        ds = torch.utils.data.Subset(ds, range(min(limit, len(ds))))
+
+    # collate_fn: PIL 이미지들을 processor로 배치 변환 + label 텐서화
+    def collate_fn(batch) -> Tuple[dict, torch.Tensor]:
+        images, labels = zip(*batch)  # images: PIL list, labels: int list
+
+        # processor가 내부에서 resize/normalize 등 수행 (당신 커스텀 processor 구현대로)
+        out = processor(list(images), return_tensors="pt")
+
+        # labels는 CE loss용 long
+        y = torch.tensor(labels, dtype=torch.long)
+        return out, y
+
+    loader = DataLoader(
+        ds,
+        batch_size=batch_size,
+        shuffle=False,
+        num_workers=num_workers,
+        pin_memory=(dev.type == "cuda"),
+        collate_fn=collate_fn,
+    )
+
+    # metric 누적
+    total = 0
+    correct = 0
+    loss_sum = 0.0
+
+    for step, (batch, y) in enumerate(loader, start=1):
+        # batch(dict)와 y를 device로 이동
+        batch = {k: v.to(dev) for k, v in batch.items()}
+        y = y.to(dev)
+
+        out = model(**batch, labels=y)  # labels를 주면 loss 계산됨
+        loss = out.loss
+        logits = out.logits
+
+        pred = logits.argmax(dim=-1)
+        correct += (pred == y).sum().item()
+        total += y.numel()
+        loss_sum += loss.item() * y.size(0)  # 배치 평균 loss * 배치크기
+
+        if step % 50 == 0:
+            acc = correct / total if total else 0.0
+            avg_loss = loss_sum / total if total else 0.0
+            print(f"[step {step:4d}] running acc={acc:.4f}, avg_loss={avg_loss:.4f}")
+
+    acc = correct / total if total else 0.0
+    avg_loss = loss_sum / total if total else 0.0
+
+    print("=" * 60)
+    print("Evaluation Done")
+    print(f"source     : {source}")
+    print(f"path       : {path}")
+    print(f"data_dir   : {Path(data_dir).resolve()}")
+    print(f"device     : {dev}")
+    print(f"num_samples: {total}")
+    print(f"avg_loss   : {avg_loss:.6f}")
+    print(f"accuracy   : {acc:.6f}")
+    print("=" * 60)
+
+
+def build_argparser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(description="Evaluate MNIST test set for local(dist) or hub.")
+    p.add_argument("--source", choices=["local", "hub"], required=True)
+    p.add_argument("--path", required=True)
+    p.add_argument("--data-dir", default="data", help="MNIST 다운로드/캐시 폴더. 기본: data")
+    p.add_argument("--batch-size", type=int, default=256)
+    p.add_argument("--num-workers", type=int, default=2)
+    p.add_argument("--limit", type=int, default=0, help="0이면 전체 평가, 양수면 앞에서 limit개만 평가")
+    p.add_argument("--use-token", action="store_true")
+    p.add_argument(
+        "--device",
+        default="auto",
+        help="auto | cpu | cuda | cuda:0 같은 torch device string",
+    )
+    return p
+
+
+def main() -> None:
+    args = build_argparser().parse_args()
+    evaluate(
+        source=args.source,
+        path=args.path,
+        data_dir=args.data_dir,
+        batch_size=args.batch_size,
+        num_workers=args.num_workers,
+        limit=args.limit,
+        use_token=bool(args.use_token),
+        device=args.device,
+    )
+
+
+if __name__ == "__main__":
+    main()

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d8254addc5e867aedf73e99ef4998482b7b25c4dd2e97436aabf5baf264717a5
+oid sha256:a76fa806ff07f1193e9f4bd00a0c80e3711ea506920a5f73cda09fb9c1e2aafb
 size 1302056