feat: 初始化项目结构和依赖配置

.dockerignore

@@ -0,0 +1,32 @@
+# 忽略虛擬環境和 Python 緩存
+.venv/
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+
+# 忽略深度學習模型目錄 (根據你的目錄名修改，例如 models/ 或 weights/)
+models/
+weights/
+checkpoints/
+.onnx
+*.pth
+*.bin
+*.h5
+*.safetensors
+
+# 忽略數據集
+data/
+datasets/
+
+# 忽略版本控制和編輯器配置
+.git/
+.idea/
+.vscode/
+.gitignore
+.gitattributes
+
+# 忽略日誌和臨時文件
+logs/
+*.log
+tmp/

.gitignore

@@ -0,0 +1,48 @@
+/.venv/
+
+# Dependency directories
+node_modules/
+jspm_packages/
+
+# Build outputs
+dist/
+build/
+out/
+
+# Environment variables (Sensitive info)
+.env
+.env.local
+.env.development.local
+.env.test.local
+.env.production.local
+
+# Debug logs
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+logs
+*.log
+
+# Editor specific folders (Optional but recommended for JetBrains/VSCode)
+.idea/
+.vscode/
+*.swp
+*.bak
+
+# OS generated files
+.DS_Store
+Thumbs.db
+
+# 忽略深度學習模型目錄 (根據你的目錄名修改，例如 models/ 或 weights/)
+models/
+weights/
+checkpoints/
+.onnx
+*.pth
+*.bin
+*.h5
+*.safetensors
+
+# 忽略數據集
+data/
+datasets/

.python-version

@@ -0,0 +1 @@
+3.14

Dockerfile

@@ -0,0 +1,54 @@
+# 階段 1：構建環境
+FROM ghcr.io/astral-sh/uv:python3.14-bookworm-slim AS builder
+
+# 設置 uv 緩存和編譯環境變量
+ENV UV_COMPILE_BYTECODE=1 \
+    UV_LINK_MODE=copy
+
+WORKDIR /app
+
+# 利用 Docker 層緩存安裝依賴
+# 這裡使用 --mount 綁定 uv.lock 和 pyproject.toml，避免額外的 COPY 動作
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,source=uv.lock,target=uv.lock \
+    --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
+    uv sync --frozen --no-install-project --no-dev
+
+# 階段 2：運行環境
+FROM python:3.14-slim-bookworm
+
+# 設置 Python 環境變量
+# PYTHONUNBUFFERED=1: 確保日誌直接輸出而不被緩衝
+# PYTHONDONTWRITEBYTECODE=1: 不在容器內生成 .pyc 文件
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PATH="/app/.venv/bin:$PATH"
+
+WORKDIR /app
+
+# 創建非 root 用戶並設置權限
+RUN groupadd -r appuser && useradd -r -g appuser -u 1000 -m appuser && \
+    mkdir -p /app/models /app/.cache && \
+    chown -R appuser:appuser /app
+
+USER appuser
+
+# 從構建階段複製虛擬環境
+COPY --from=builder --chown=appuser:appuser /app/.venv /app/.venv
+
+# 下載模型
+RUN python -c "from huggingface_hub import snapshot_download; \
+    snapshot_download(repo_id='ynyg/InkErase', \
+    local_dir='/app/models/InkErase', \
+    ignore_patterns=['*.ckpt', '*.pth', '*.git*'])"
+
+# 複製應用代碼 (建議先複製代碼再啟動)
+# 注意：如果項目很大，建議在 .dockerignore 中排除 .venv, .git 等
+COPY --chown=appuser:appuser . .
+
+# 暴露 FastAPI 默認端口或 Hugging Face Spaces 要求的端口
+EXPOSE 7860
+
+# 使用虛擬環境中的 uvicorn 啟動
+# 增加 --proxy-headers 處理反向代理（如 Hugging Face 或 Nginx）
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", "--proxy-headers"]

app.py

@@ -0,0 +1,171 @@
+import io
+import json
+from contextlib import asynccontextmanager
+from pathlib import Path
+
+import albumentations as A
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+from anyio.to_thread import run_sync
+from fastapi import FastAPI, Request, UploadFile, File
+from fastapi.responses import Response
+from segmentation_models_pytorch import UnetPlusPlus
+
+# 模型路徑
+MODEL_PATH = "models/InkErase"
+# 設備
+device = "cuda" if torch.cuda.is_available() else "cpu"
+# 分块大小
+TRAIN_SIZE = 512
+
+
+def load_model() -> UnetPlusPlus:
+    """加載模型"""
+    # 模型路径
+    path = Path(MODEL_PATH)
+    # 读取配置文件
+    cfg = json.loads((path / "config.json").read_text(encoding="utf-8"))
+    # 加載模型
+    return UnetPlusPlus(
+        encoder_name=cfg.get("encoder_name", "resnet50"),
+        encoder_weights=None,
+        in_channels=int(cfg.get("in_channels", 3)),
+        classes=int(cfg.get("classes", 3)),
+        decoder_attention_type=cfg.get("decoder_attention_type"),
+        activation=cfg.get("activation", "sigmoid"),
+    )
+
+
+def get_preprocessing() -> A.Compose:
+    """获取Albumentations 預處理 pipeline"""
+    return A.Compose([
+        A.Normalize(mean=(0, 0, 0), std=(1, 1, 1), max_pixel_value=255.0),
+        A.ToTensorV2()
+    ])
+
+
+@asynccontextmanager
+async def lifespan(instance: FastAPI):
+    """
+    FastAPI 應用程序的生命周期管理器。
+    :param instance: FastAPI 應用程序實例
+    """
+    # 加載模型
+    instance.state.model = load_model()
+    # 初始化預處理函數
+    instance.state.preprocess_fn = get_preprocessing()
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
+
+
+@app.post("/predict")
+async def predict(request: Request, file: UploadFile = File(...)):
+    """
+    笔迹擦除
+    :param request: 请求对象
+    :param file: 待处理的图片
+    :return: 預測結果，包括文本、預測類別和置信度
+    """
+    # 1. 使用 OpenCV 直接從內存讀取圖片
+    content = await file.read()
+    # 將 bytes 轉換為 numpy array
+    nparr = np.frombuffer(content, np.uint8)
+    # 解碼圖片 (默認 BGR)
+    original_image = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
+    # 转换为 RGB
+    original_image = cv2.cvtColor(original_image, cv2.COLOR_BGR2RGB)
+
+    # 获取图片尺寸
+    orig_h, orig_w = original_image.shape[:2]
+    # 获取模型和处理流
+    model = request.app.state.model
+    preprocess_fn = request.app.state.preprocess_fn
+
+    def _inference_logic():
+        with torch.no_grad():
+            # ==============================
+            # 情況 A: 圖片大於 512，進行切塊處理
+            # ==============================
+            if orig_w > TRAIN_SIZE or orig_h > TRAIN_SIZE:
+                # 1. 計算新的寬高（補齊為 512 的倍數）
+                new_w = (orig_w // TRAIN_SIZE + (1 if orig_w % TRAIN_SIZE != 0 else 0)) * TRAIN_SIZE
+                new_h = (orig_h // TRAIN_SIZE + (1 if orig_h % TRAIN_SIZE != 0 else 0)) * TRAIN_SIZE
+
+                # 2. Padding 原圖
+                padded_img = Image.new("RGB", (new_w, new_h), (0, 0, 0))
+                padded_img.paste(original_image, (0, 0))
+
+                result_mask = Image.new("L", (new_w, new_h))
+
+                # 3. 循環切割
+                for y in range(0, new_h, TRAIN_SIZE):
+                    for x in range(0, new_w, TRAIN_SIZE):
+                        box = (x, y, x + TRAIN_SIZE, y + TRAIN_SIZE)
+                        patch = padded_img.crop(box)
+
+                        # --- 修改部分: Albumentations 處理 ---
+                        patch_np = np.array(patch)  # 轉換為 numpy
+                        transformed = preprocess_fn(image=patch_np)
+                        input_tensor = transformed["image"].unsqueeze(0).to(device)  # [1, C, H, W]
+                        # -----------------------------------
+
+                        output = model(input_tensor)
+
+                        pred_mask = (output > 0.5).float().squeeze().cpu().numpy()
+                        pred_mask = (pred_mask * 255).astype(np.uint8)
+
+                        patch_mask_img = Image.fromarray(pred_mask)
+                        result_mask.paste(patch_mask_img, (x, y))
+
+                final_image = result_mask.crop((0, 0, orig_w, orig_h))
+
+            # ==============================
+            # 情況 B: 圖片小於等於 512
+            # ==============================
+            else:
+                pad_w = TRAIN_SIZE
+                pad_h = TRAIN_SIZE
+
+                padded_img = Image.new("RGB", (pad_w, pad_h), (0, 0, 0))
+                padded_img.paste(original_image, (0, 0))
+
+                # --- 修改部分: Albumentations 處理 ---
+                patch_np = np.array(padded_img)  # 轉換為 numpy
+                transformed = preprocess_fn(image=patch_np)
+                input_tensor = transformed["image"].unsqueeze(0).to(device)
+                # -----------------------------------
+
+                output = model(input_tensor)
+
+                pred_mask = (output > 0.5).float().squeeze().cpu().numpy()
+                pred_mask = (pred_mask * 255).astype(np.uint8)
+
+                final_image = Image.fromarray(pred_mask).crop((0, 0, orig_w, orig_h))
+
+        return final_image
+
+    # 執行推理
+    result_image = await run_sync(_inference_logic)
+
+    # 返回圖片流
+    img_byte_arr = io.BytesIO()
+    result_image.save(img_byte_arr, format='PNG')
+    return Response(content=img_byte_arr.getvalue(), media_type="image/png")
+
+
+@app.get("/")
+def greet_json():
+    """
+    返回一個 JSON 格式的歡迎訊息。
+    """
+    return {"Hello": "World!"}
+
+
+if __name__ == '__main__':
+    import uvicorn
+
+    uvicorn.run("app:app", host="0.0.0.0", port=8000)

main.py

@@ -0,0 +1,6 @@
+def main():
+    print("Hello from inkerase!")
+
+
+if __name__ == "__main__":
+    main()

pyproject.toml

@@ -0,0 +1,14 @@
+[project]
+name = "inkerase"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.14"
+dependencies = [
+    "albumentations>=2.0.8",
+    "fastapi[all]>=0.128.0",
+    "safetensors>=0.7.0",
+    "segmentation-models-pytorch>=0.5.0",
+    "torch>=2.10.0",
+    "torchvision>=0.25.0",
+]