Merge branch 'main' of https://github.com/Byaidu/PDFMathTranslate

.gitignore

@@ -1,3 +1,5 @@
+pdf2zh_files
+gui/pdf2zh_files
 gradio_files
 tmp
 gui/gradio_files

Dockerfile.Demo

@@ -2,12 +2,14 @@ FROM python:3.12
 
 WORKDIR /app
 
+COPY . .
+
 ENV PYTHONUNBUFFERED=1
 
-RUN apt-get update && apt-get install -y libgl1 \
-    && rm -rf /var/lib/apt/lists/*
+RUN apt-get update && apt-get install -y libgl1
+
+RUN pip install .
 
-RUN pip install pdf2zh
 RUN mkdir -p /data
 RUN chmod 777 /data
 RUN mkdir -p /app
@@ -17,4 +19,4 @@ RUN chmod 777 /.cache
 RUN mkdir -p ./gradio_files
 RUN chmod 777 ./gradio_files
 
-CMD ["pdf2zh", "-i"]
+CMD ["pdf2zh", "-i"]

README.md

@@ -37,15 +37,15 @@ Feel free to provide feedback in [GitHub Issues](https://github.com/Byaidu/PDFMa
 
 <h2 id="updates">Updates</h2>
 
+- [Nov. 23 2024] [ONNX](https://github.com/onnx/onnx) support to reduce dependency sizes *(by [@Wybxc](https://github.com/Wybxc))*  
+- [Nov. 23 2024] 🌟 [Public Service](#demo)  online! *(by [@Byaidu](https://github.com/Byaidu))*  
+- [Nov. 23 2024] Non-PDF/A documents are now supported *(by [@reycn](https://github.com/reycn))*  
 - [Nov. 23 2024] Firewall for preventing web bots *(by [@Byaidu](https://github.com/Byaidu))*  
 - [Nov. 22 2024] GUI now supports Italian, and has been improved *(by [@Byaidu](https://github.com/Byaidu), [@reycn](https://github.com/reycn))*  
 - [Nov. 22 2024] You can now share your deployed service to others *(by [@Zxis233](https://github.com/Zxis233))*  
-- [Nov. 22 2024] Now supportsTencent Translation *(by [@hellofinch](https://github.com/hellofinch))*  
+- [Nov. 22 2024] Now supports Tencent Translation *(by [@hellofinch](https://github.com/hellofinch))*  
 - [Nov. 21 2024] GUI now supports downloading dual-document *(by [@reycn](https://github.com/reycn))*  
-- [Nov. 20 2024] GUI now supports specifying Ollama and OpenAI models *(by [@IuvenisSapiens](https://github.com/IuvenisSapiens), [@Byaidu](https://github.com/Byaidu))*  
 - [Nov. 20 2024] 🌟 [Demo](#demo)  online! *(by [@reycn](https://github.com/reycn))*  
-- [Nov. 20 2024] Supports [Docker](#docker) *(by [@Byaidu](https://github.com/Byaidu))*  
-- [Nov. 20 2024] Supports [multiple-threads translation](#threads) *(by [@Byaidu](https://github.com/Byaidu))*  
 
 <h2 id="preview">Preview</h2>
 
@@ -53,9 +53,15 @@ Feel free to provide feedback in [GitHub Issues](https://github.com/Byaidu/PDFMa
 <img src="./docs/images/preview.gif" width="80%"/>
 </div>
 
-<h2 id="demo">Demo 🌟</h2>
+<h2 id="demo">Public Service 🌟</h2>
 
-You can try [our demo on HuggingFace](https://huggingface.co/spaces/reycn/PDFMathTranslate-Docker) without installation.  
+### Free Service (<https://pdf2zh.com/>)
+
+You can try our [public service](https://pdf2zh.com/) online without installation.  
+
+### Hugging Face Demo
+
+You can try [our demo on HuggingFace](https://huggingface.co/spaces/reycn/PDFMathTranslate-Docker) without installation.
 Note that the computing resources of the demo are limited, so please avoid abusing them.
 
 <h2 id="install">Installation and Usage</h2>

README_zh-CN.md

@@ -37,16 +37,15 @@
 
 <h2 id="updates">近期更新</h2>
 
-
+- [Nov. 24 2024] 为降低依赖大小，提供 [ONNX](https://github.com/onnx/onnx) 支持 *(by [@Wybxc](https://github.com/Wybxc))*  
+- [Nov. 23 2024] 🌟 [免费公共服务](#demo) 上线! *(by [@Byaidu](https://github.com/Byaidu))*  
+- [Nov. 23 2024] 非 PDF/A 文档也能正常翻译了 *(by [@reycn](https://github.com/reycn))*  
 - [Nov. 23 2024] 防止网页爬虫的防火墙 *(by [@Byaidu](https://github.com/Byaidu))*  
 - [Nov. 22 2024] 图形用户界面现已支持意大利语，并获得了一些更新 *(by [@Byaidu](https://github.com/Byaidu), [@reycn](https://github.com/reycn))*  
 - [Nov. 22 2024] 现在你可以将自己部署的服务分享给朋友了 *(by [@Zxis233](https://github.com/Zxis233))*  
-- [Nov. 22 2024] Now supportsTencent Translation *(by [@hellofinch](https://github.com/hellofinch))*  
+- [Nov. 22 2024] 支持腾讯翻译 *(by [@hellofinch](https://github.com/hellofinch))*  
 - [Nov. 21 2024] 图形用户界面现在支持下载双语文档 *(by [@reycn](https://github.com/reycn))*  
-- [Nov. 20 2024] 图形用户界面现在支持指定 Ollama 和 OpenAI 的模型 *(by [@IuvenisSapiens](https://github.com/IuvenisSapiens), [@Byaidu](https://github.com/Byaidu))*  
 - [Nov. 20 2024] 🌟 提供了 [在线演示](#demo)！ *(by [@reycn](https://github.com/reycn))*  
-- [Nov. 20 2024] 支持 [容器化部署](#docker) *(by [@Byaidu](https://github.com/Byaidu))*  
-- [Nov. 20 2024] 支持速度更快的 [多线程翻译](#threads) *(by [@Byaidu](https://github.com/Byaidu))* 
 
 <h2 id="preview">效果预览</h2>
 
@@ -56,7 +55,13 @@
 
 <h2 id="demo">在线演示 🌟</h2>
 
-你可以立即尝试 [在 HuggingFace 上的在线演示](https://huggingface.co/spaces/reycn/PDFMathTranslate-Docker) 而无需安装.  
+### 免费服务 (<https://pdf2zh.com/>)
+
+你可以立即尝试 [免费公共服务](https://pdf2zh.com/) 而无需安装。
+
+### Hugging Face 在线演示
+
+你可以立即尝试 [在 HuggingFace 上的在线演示](https://huggingface.co/spaces/reycn/PDFMathTranslate-Docker) 而无需安装。
 请注意，演示的计算资源有限，因此请避免滥用。
 
 <h2 id="install">安装和使用</h2>

pdf2zh/doclayout.py

@@ -0,0 +1,213 @@
+import abc
+import cv2
+import numpy as np
+import contextlib
+from huggingface_hub import hf_hub_download
+
+
+class DocLayoutModel(abc.ABC):
+    @staticmethod
+    def load_torch():
+        model = TorchModel.from_pretrained(
+            repo_id="juliozhao/DocLayout-YOLO-DocStructBench",
+            filename="doclayout_yolo_docstructbench_imgsz1024.pt",
+        )
+        return model
+
+    @staticmethod
+    def load_onnx():
+        model = OnnxModel.from_pretrained(
+            repo_id="wybxc/DocLayout-YOLO-DocStructBench-onnx",
+            filename="doclayout_yolo_docstructbench_imgsz1024.onnx",
+        )
+        return model
+
+    @staticmethod
+    def load_available():
+        with contextlib.suppress(ImportError):
+            return DocLayoutModel.load_torch()
+
+        with contextlib.suppress(ImportError):
+            return DocLayoutModel.load_onnx()
+
+        raise ImportError(
+            "Please install the `torch` or `onnx` feature to use the DocLayout model."
+        )
+
+    @property
+    @abc.abstractmethod
+    def stride(self) -> int:
+        """Stride of the model input."""
+        pass
+
+    @abc.abstractmethod
+    def predict(self, image, imgsz=1024, **kwargs) -> list:
+        """
+        Predict the layout of a document page.
+
+        Args:
+            image: The image of the document page.
+            imgsz: Resize the image to this size. Must be a multiple of the stride.
+            **kwargs: Additional arguments.
+        """
+        pass
+
+
+class TorchModel(DocLayoutModel):
+    def __init__(self, model_path: str):
+        try:
+            import doclayout_yolo
+        except ImportError:
+            raise ImportError(
+                "Please install the `torch` feature to use the Torch model."
+            )
+
+        self.model_path = model_path
+        self.model = doclayout_yolo.YOLOv10(model_path)
+
+    @staticmethod
+    def from_pretrained(repo_id: str, filename: str):
+        pth = hf_hub_download(repo_id=repo_id, filename=filename)
+        return TorchModel(pth)
+
+    @property
+    def stride(self):
+        return 32
+
+    def predict(self, *args, **kwargs):
+        return self.model.predict(*args, **kwargs)
+
+
+class YoloResult:
+    """Helper class to store detection results from ONNX model."""
+
+    def __init__(self, boxes, names):
+        self.boxes = [YoloBox(data=d) for d in boxes]
+        self.boxes.sort(key=lambda x: x.conf, reverse=True)
+        self.names = names
+
+
+class YoloBox:
+    """Helper class to store detection results from ONNX model."""
+
+    def __init__(self, data):
+        self.xyxy = data[:4]
+        self.conf = data[-2]
+        self.cls = data[-1]
+
+
+class OnnxModel(DocLayoutModel):
+    def __init__(self, model_path: str):
+        import ast
+
+        try:
+
+            import onnx
+            import onnxruntime
+        except ImportError:
+            raise ImportError(
+                "Please install the `onnx` feature to use the ONNX model."
+            )
+
+        self.model_path = model_path
+
+        model = onnx.load(model_path)
+        metadata = {d.key: d.value for d in model.metadata_props}
+        self._stride = ast.literal_eval(metadata["stride"])
+        self._names = ast.literal_eval(metadata["names"])
+
+        self.model = onnxruntime.InferenceSession(model.SerializeToString())
+
+    @staticmethod
+    def from_pretrained(repo_id: str, filename: str):
+        pth = hf_hub_download(repo_id=repo_id, filename=filename)
+        return OnnxModel(pth)
+
+    @property
+    def stride(self):
+        return self._stride
+
+    def resize_and_pad_image(self, image, new_shape):
+        """
+        Resize and pad the image to the specified size, ensuring dimensions are multiples of stride.
+
+        Parameters:
+        - image: Input image
+        - new_shape: Target size (integer or (height, width) tuple)
+        - stride: Padding alignment stride, default 32
+
+        Returns:
+        - Processed image
+        """
+        if isinstance(new_shape, int):
+            new_shape = (new_shape, new_shape)
+
+        h, w = image.shape[:2]
+        new_h, new_w = new_shape
+
+        # Calculate scaling ratio
+        r = min(new_h / h, new_w / w)
+        resized_h, resized_w = int(round(h * r)), int(round(w * r))
+
+        # Resize image
+        image = cv2.resize(
+            image, (resized_w, resized_h), interpolation=cv2.INTER_LINEAR
+        )
+
+        # Calculate padding size and align to stride multiple
+        pad_w = (new_w - resized_w) % self.stride
+        pad_h = (new_h - resized_h) % self.stride
+        top, bottom = pad_h // 2, pad_h - pad_h // 2
+        left, right = pad_w // 2, pad_w - pad_w // 2
+
+        # Add padding
+        image = cv2.copyMakeBorder(
+            image, top, bottom, left, right, cv2.BORDER_CONSTANT, value=(114, 114, 114)
+        )
+
+        return image
+
+    def scale_boxes(self, img1_shape, boxes, img0_shape):
+        """
+        Rescales bounding boxes (in the format of xyxy by default) from the shape of the image they were originally
+        specified in (img1_shape) to the shape of a different image (img0_shape).
+
+        Args:
+            img1_shape (tuple): The shape of the image that the bounding boxes are for,
+                in the format of (height, width).
+            boxes (torch.Tensor): the bounding boxes of the objects in the image, in the format of (x1, y1, x2, y2)
+            img0_shape (tuple): the shape of the target image, in the format of (height, width).
+
+        Returns:
+            boxes (torch.Tensor): The scaled bounding boxes, in the format of (x1, y1, x2, y2)
+        """
+
+        # Calculate scaling ratio
+        gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1])
+
+        # Calculate padding size
+        pad_x = round((img1_shape[1] - img0_shape[1] * gain) / 2 - 0.1)
+        pad_y = round((img1_shape[0] - img0_shape[0] * gain) / 2 - 0.1)
+
+        # Remove padding and scale boxes
+        boxes[..., :4] = (boxes[..., :4] - [pad_x, pad_y, pad_x, pad_y]) / gain
+        return boxes
+
+    def predict(self, image, imgsz=1024, **kwargs):
+        # Preprocess input image
+        orig_h, orig_w = image.shape[:2]
+        pix = self.resize_and_pad_image(image, new_shape=imgsz)
+        pix = np.transpose(pix, (2, 0, 1))  # CHW
+        pix = np.expand_dims(pix, axis=0)  # BCHW
+        pix = pix.astype(np.float32) / 255.0  # Normalize to [0, 1]
+        new_h, new_w = pix.shape[2:]
+
+        # Run inference
+        preds = self.model.run(None, {"images": pix})[0]
+
+        # Postprocess predictions
+        preds = preds[preds[..., 4] > 0.25]
+        preds[..., :4] = self.scale_boxes(
+            (new_h, new_w), preds[..., :4], (orig_h, orig_w)
+        )
+        return [YoloResult(boxes=preds, names=self._names)]

pdf2zh/high_level.py

@@ -4,7 +4,6 @@ import logging
 import sys
 from io import StringIO
 from typing import Any, BinaryIO, Container, Iterator, Optional, cast
-import torch
 import numpy as np
 import tqdm
 from pymupdf import Document
@@ -22,7 +21,7 @@ from pdf2zh.pdfdevice import PDFDevice, TagExtractor
 from pdf2zh.pdfexceptions import PDFValueError
 from pdf2zh.pdfinterp import PDFPageInterpreter, PDFResourceManager
 from pdf2zh.pdfpage import PDFPage
-from pdf2zh.utils import AnyIO, FileOrName, open_filename
+from pdf2zh.utils import AnyIO, FileOrName, open_filename, get_device
 
 
 def extract_text_to_fp(
@@ -176,11 +175,7 @@ def extract_text_to_fp(
                 pix.height, pix.width, 3
             )[:, :, ::-1]
             page_layout = model.predict(
-                image,
-                imgsz=int(pix.height / 32) * 32,
-                device=(
-                    "cuda:0" if torch.cuda.is_available() else "cpu"
-                ),  # Auto-select GPU if available
+                image, imgsz=int(pix.height / 32) * 32, device=get_device()
             )[0]
             # kdtree 是不可能 kdtree 的，不如直接渲染成图片，用空间换时间
             box = np.ones((pix.height, pix.width))

pdf2zh/pdf2zh.py

@@ -14,7 +14,6 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Any, Container, Iterable, List, Optional
 
 import pymupdf
-from huggingface_hub import hf_hub_download
 
 from pdf2zh import __version__
 from pdf2zh.pdfexceptions import PDFValueError
@@ -27,10 +26,14 @@ OUTPUT_TYPES = ((".htm", "html"), (".html", "html"), (".xml", "xml"), (".tag", "
 
 
 def setup_log() -> None:
-    import doclayout_yolo
-
     logging.basicConfig()
-    doclayout_yolo.utils.LOGGER.setLevel(logging.WARNING)
+
+    try:
+        import doclayout_yolo
+
+        doclayout_yolo.utils.LOGGER.setLevel(logging.WARNING)
+    except ImportError:
+        pass
 
 
 def check_files(files: List[str]) -> List[str]:
@@ -73,8 +76,7 @@ def extract_text(
     output: str = "",
     **kwargs: Any,
 ) -> AnyIO:
-    import doclayout_yolo
-
+    from pdf2zh.doclayout import DocLayoutModel
     import pdf2zh.high_level
 
     if not files:
@@ -86,15 +88,7 @@ def extract_text(
                 output_type = alttype
 
     outfp: AnyIO = sys.stdout
-    # pth = os.path.join(tempfile.gettempdir(), 'doclayout_yolo_docstructbench_imgsz1024.pt')
-    # if not os.path.exists(pth):
-    #     print('Downloading...')
-    #     urllib.request.urlretrieve("http://huggingface.co/juliozhao/DocLayout-YOLO-DocStructBench/resolve/main/doclayout_yolo_docstructbench_imgsz1024.pt",pth)
-    pth = hf_hub_download(
-        repo_id="juliozhao/DocLayout-YOLO-DocStructBench",
-        filename="doclayout_yolo_docstructbench_imgsz1024.pt",
-    )
-    model = doclayout_yolo.YOLOv10(pth)
+    model = DocLayoutModel.load_available()
 
     for file in files:
         filename = os.path.splitext(os.path.basename(file))[0]

pdf2zh/utils.py

@@ -819,3 +819,16 @@ def format_int_alpha(value: int) -> str:
 
     result.reverse()
     return "".join(result)
+
+
+def get_device():
+    """Get the device to use for computation."""
+    try:
+        import torch
+
+        if torch.cuda.is_available():
+            return "cuda:0"
+    except ImportError:
+        pass
+
+    return "cpu"

pyproject.toml

@@ -5,7 +5,7 @@ description = "Latex PDF Translator"
 authors = [{ name = "Byaidu", email = "byaidux@gmail.com" }]
 license = "AGPL-3.0"
 readme = "README.md"
-requires-python = ">=3.8,<3.13"
+requires-python = ">=3.9,<3.13"
 classifiers = [
     "Programming Language :: Python :: 3",
     "Operating System :: OS Independent",
@@ -17,7 +17,6 @@ dependencies = [
     "pymupdf",
     "tqdm",
     "tenacity",
-    "doclayout-yolo",
     "numpy",
     "ollama",
     "deepl<1.19.1",
@@ -25,10 +24,16 @@ dependencies = [
     "azure-ai-translation-text<=1.0.1",
     "gradio",
     "huggingface_hub",
-    "torch",
+    "onnx",
+    "onnxruntime",
+    "opencv-python-headless",
 ]
 
 [project.optional-dependencies]
+torch = [
+    "doclayout-yolo",
+    "torch",
+]
 dev = [
     "black",
     "flake8",