Merge remote-tracking branch 'origin/dev' into dev

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/__init__.py

+from .configuration_unimer_swin import UnimerSwinConfig
+from .modeling_unimer_swin import UnimerSwinModel
+from .image_processing_unimer_swin import UnimerSwinImageProcessor
+
+__all__ = [
+    "UnimerSwinConfig",
+    "UnimerSwinModel",
+    "UnimerSwinImageProcessor",
+]

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/configuration_unimer_swin.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Donut Swin Transformer model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class UnimerSwinConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`UnimerSwinModel`]. It is used to instantiate a
+    Donut model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Donut
+    [naver-clova-ix/donut-base](https://huggingface.co/naver-clova-ix/donut-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 4):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        embed_dim (`int`, *optional*, defaults to 96):
+            Dimensionality of patch embedding.
+        depths (`list(int)`, *optional*, defaults to `[2, 2, 6, 2]`):
+            Depth of each layer in the Transformer encoder.
+        num_heads (`list(int)`, *optional*, defaults to `[3, 6, 12, 24]`):
+            Number of attention heads in each layer of the Transformer encoder.
+        window_size (`int`, *optional*, defaults to 7):
+            Size of windows.
+        mlp_ratio (`float`, *optional*, defaults to 4.0):
+            Ratio of MLP hidden dimensionality to embedding dimensionality.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not a learnable bias should be added to the queries, keys and values.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings and encoder.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        drop_path_rate (`float`, *optional*, defaults to 0.1):
+            Stochastic depth rate.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder. If string, `"gelu"`, `"relu"`,
+            `"selu"` and `"gelu_new"` are supported.
+        use_absolute_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether or not to add absolute position embeddings to the patch embeddings.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import UnimerSwinConfig, UnimerSwinModel
+
+    >>> # Initializing a Donut naver-clova-ix/donut-base style configuration
+    >>> configuration = UnimerSwinConfig()
+
+    >>> # Randomly initializing a model from the naver-clova-ix/donut-base style configuration
+    >>> model = UnimerSwinModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "unimer-swin"
+
+    attribute_map = {
+        "num_attention_heads": "num_heads",
+        "num_hidden_layers": "num_layers",
+    }
+
+    def __init__(
+        self,
+        image_size=224,
+        patch_size=4,
+        num_channels=3,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        drop_path_rate=0.1,
+        hidden_act="gelu",
+        use_absolute_embeddings=False,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.num_layers = len(depths)
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.mlp_ratio = mlp_ratio
+        self.qkv_bias = qkv_bias
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.drop_path_rate = drop_path_rate
+        self.hidden_act = hidden_act
+        self.use_absolute_embeddings = use_absolute_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_range = initializer_range
+        # we set the hidden_size attribute in order to make Swin work with VisionEncoderDecoderModel
+        # this indicates the channel dimension after the last stage of the model
+        self.hidden_size = int(embed_dim * 2 ** (len(depths) - 1))

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/image_processing_unimer_swin.py

+from transformers.image_processing_utils import BaseImageProcessor
+import numpy as np
+import cv2
+import albumentations as alb
+from albumentations.pytorch import ToTensorV2
+
+
+# TODO: dereference cv2 if possible
+class UnimerSwinImageProcessor(BaseImageProcessor):
+    def __init__(
+            self,
+            image_size = (192, 672),
+        ):
+        self.input_size = [int(_) for _ in image_size]
+        assert len(self.input_size) == 2
+    
+        self.transform = alb.Compose(
+            [
+                alb.ToGray(always_apply=True),
+                alb.Normalize((0.7931, 0.7931, 0.7931), (0.1738, 0.1738, 0.1738)),
+                # alb.Sharpen()
+                ToTensorV2(),
+            ]
+        )
+
+    def __call__(self, item):
+        image = self.prepare_input(item)
+        return self.transform(image=image)['image'][:1]
+
+    @staticmethod
+    def crop_margin_numpy(img: np.ndarray) -> np.ndarray:
+        """Crop margins of image using NumPy operations"""
+        # Convert to grayscale if it's a color image
+        if len(img.shape) == 3 and img.shape[2] == 3:
+            gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
+        else:
+            gray = img.copy()
+
+        # Normalize and threshold
+        if gray.max() == gray.min():
+            return img
+
+        normalized = (((gray - gray.min()) / (gray.max() - gray.min())) * 255).astype(np.uint8)
+        binary = 255 * (normalized < 200).astype(np.uint8)
+
+        # Find bounding box
+        coords = cv2.findNonZero(binary)  # Find all non-zero points (text)
+        x, y, w, h = cv2.boundingRect(coords)  # Find minimum spanning bounding box
+
+        # Return cropped image
+        return img[y:y + h, x:x + w]
+
+    def prepare_input(self, img, random_padding: bool = False):
+        """
+        Convert PIL Image or numpy array to properly sized and padded image after:
+            - crop margins
+            - resize while maintaining aspect ratio
+            - pad to target size
+        """
+        if img is None:
+            return None
+
+        try:
+            img = self.crop_margin_numpy(img)
+        except Exception:
+            # might throw an error for broken files
+            return None
+
+        if img.shape[0] == 0 or img.shape[1] == 0:
+            return None
+
+        # Resize while preserving aspect ratio
+        h, w = img.shape[:2]
+        scale = min(self.input_size[0] / h, self.input_size[1] / w)
+        new_h, new_w = int(h * scale), int(w * scale)
+        resized_img = cv2.resize(img, (new_w, new_h), interpolation=cv2.INTER_AREA)
+
+        # Calculate padding
+        pad_width, pad_height = self._get_padding_values(new_w, new_h, random_padding)
+
+        # Create and apply padding
+        channels = 3 if len(img.shape) == 3 else 1
+        padded_img = np.full((self.input_size[0], self.input_size[1], channels), 255, dtype=np.uint8)
+        padded_img[pad_height:pad_height + new_h, pad_width:pad_width + new_w] = resized_img
+
+        return padded_img
+
+    def _calculate_padding(self, new_w, new_h, random_padding):
+        """Calculate padding values for PIL images"""
+        delta_width = self.input_size[1] - new_w
+        delta_height = self.input_size[0] - new_h
+
+        pad_width, pad_height = self._get_padding_values(new_w, new_h, random_padding)
+
+        return (
+            pad_width,
+            pad_height,
+            delta_width - pad_width,
+            delta_height - pad_height,
+        )
+
+    def _get_padding_values(self, new_w, new_h, random_padding):
+        """Get padding values based on image dimensions and padding strategy"""
+        delta_width = self.input_size[1] - new_w
+        delta_height = self.input_size[0] - new_h
+
+        if random_padding:
+            pad_width = np.random.randint(low=0, high=delta_width + 1)
+            pad_height = np.random.randint(low=0, high=delta_height + 1)
+        else:
+            pad_width = delta_width // 2
+            pad_height = delta_height // 2
+
+        return pad_width, pad_height

magic_pdf/model/sub_modules/model_init.py

@@ -5,12 +5,13 @@ from magic_pdf.config.constants import MODEL_NAME
 from magic_pdf.model.model_list import AtomicModel
 from magic_pdf.model.sub_modules.language_detection.yolov11.YOLOv11 import YOLOv11LangDetModel
 from magic_pdf.model.sub_modules.layout.doclayout_yolo.DocLayoutYOLO import DocLayoutYOLOModel
-from magic_pdf.model.sub_modules.layout.layoutlmv3.model_init import Layoutlmv3_Predictor
 from magic_pdf.model.sub_modules.mfd.yolov8.YOLOv8 import YOLOv8MFDModel
 from magic_pdf.model.sub_modules.mfr.unimernet.Unimernet import UnimernetModel
 
 try:
-    from magic_pdf_ascend_plugin.libs.license_verifier import load_license, LicenseFormatError, LicenseSignatureError, LicenseExpiredError
+    from magic_pdf_ascend_plugin.libs.license_verifier import (
+        LicenseExpiredError, LicenseFormatError, LicenseSignatureError,
+        load_license)
     from magic_pdf_ascend_plugin.model_plugin.ocr.paddleocr.ppocr_273_npu import ModifiedPaddleOCR
     from magic_pdf_ascend_plugin.model_plugin.table.rapidtable.rapid_table_npu import RapidTableModel
     license_key = load_license()
@@ -20,26 +21,26 @@ except Exception as e:
     if isinstance(e, ImportError):
         pass
     elif isinstance(e, LicenseFormatError):
-        logger.error("Ascend Plugin: Invalid license format. Please check the license file.")
+        logger.error('Ascend Plugin: Invalid license format. Please check the license file.')
     elif isinstance(e, LicenseSignatureError):
-        logger.error("Ascend Plugin: Invalid signature. The license may be tampered with.")
+        logger.error('Ascend Plugin: Invalid signature. The license may be tampered with.')
     elif isinstance(e, LicenseExpiredError):
-        logger.error("Ascend Plugin: License has expired. Please renew your license.")
+        logger.error('Ascend Plugin: License has expired. Please renew your license.')
     elif isinstance(e, FileNotFoundError):
-        logger.error("Ascend Plugin: Not found License file.")
+        logger.error('Ascend Plugin: Not found License file.')
     else:
-        logger.error(f"Ascend Plugin: {e}")
+        logger.error(f'Ascend Plugin: {e}')
     from magic_pdf.model.sub_modules.ocr.paddleocr.ppocr_273_mod import ModifiedPaddleOCR
     # from magic_pdf.model.sub_modules.ocr.paddleocr.ppocr_291_mod import ModifiedPaddleOCR
     from magic_pdf.model.sub_modules.table.rapidtable.rapid_table import RapidTableModel
 
-from magic_pdf.model.sub_modules.table.structeqtable.struct_eqtable import StructTableModel
-from magic_pdf.model.sub_modules.table.tablemaster.tablemaster_paddle import TableMasterPaddleModel
 
 def table_model_init(table_model_type, model_path, max_time, _device_='cpu', ocr_engine=None, table_sub_model_name=None):
     if table_model_type == MODEL_NAME.STRUCT_EQTABLE:
+        from magic_pdf.model.sub_modules.table.structeqtable.struct_eqtable import StructTableModel
         table_model = StructTableModel(model_path, max_new_tokens=2048, max_time=max_time)
     elif table_model_type == MODEL_NAME.TABLE_MASTER:
+        from magic_pdf.model.sub_modules.table.tablemaster.tablemaster_paddle import TableMasterPaddleModel
         config = {
             'model_dir': model_path,
             'device': _device_
@@ -55,7 +56,7 @@ def table_model_init(table_model_type, model_path, max_time, _device_='cpu', ocr
 
 
 def mfd_model_init(weight, device='cpu'):
-    if str(device).startswith("npu"):
+    if str(device).startswith('npu'):
         device = torch.device(device)
     mfd_model = YOLOv8MFDModel(weight, device)
     return mfd_model
@@ -67,19 +68,20 @@ def mfr_model_init(weight_dir, cfg_path, device='cpu'):
 
 
 def layout_model_init(weight, config_file, device):
+    from magic_pdf.model.sub_modules.layout.layoutlmv3.model_init import Layoutlmv3_Predictor
     model = Layoutlmv3_Predictor(weight, config_file, device)
     return model
 
 
 def doclayout_yolo_model_init(weight, device='cpu'):
-    if str(device).startswith("npu"):
+    if str(device).startswith('npu'):
         device = torch.device(device)
     model = DocLayoutYOLOModel(weight, device)
     return model
 
 
 def langdetect_model_init(langdetect_model_weight, device='cpu'):
-    if str(device).startswith("npu"):
+    if str(device).startswith('npu'):
         device = torch.device(device)
     model = YOLOv11LangDetModel(langdetect_model_weight, device)
     return model

magic_pdf/model/sub_modules/model_utils.py

 import time
-
 import torch
-from PIL import Image
 from loguru import logger
-
+import numpy as np
 from magic_pdf.libs.clean_memory import clean_memory
 
 
-def crop_img(input_res, input_pil_img, crop_paste_x=0, crop_paste_y=0):
+def crop_img(input_res, input_np_img, crop_paste_x=0, crop_paste_y=0):
+
     crop_xmin, crop_ymin = int(input_res['poly'][0]), int(input_res['poly'][1])
     crop_xmax, crop_ymax = int(input_res['poly'][4]), int(input_res['poly'][5])
-    # Create a white background with an additional width and height of 50
+
+    # Calculate new dimensions
     crop_new_width = crop_xmax - crop_xmin + crop_paste_x * 2
     crop_new_height = crop_ymax - crop_ymin + crop_paste_y * 2
-    return_image = Image.new('RGB', (crop_new_width, crop_new_height), 'white')
 
-    # Crop image
-    crop_box = (crop_xmin, crop_ymin, crop_xmax, crop_ymax)
-    cropped_img = input_pil_img.crop(crop_box)
-    return_image.paste(cropped_img, (crop_paste_x, crop_paste_y))
-    return_list = [crop_paste_x, crop_paste_y, crop_xmin, crop_ymin, crop_xmax, crop_ymax, crop_new_width, crop_new_height]
+    # Create a white background array
+    return_image = np.ones((crop_new_height, crop_new_width, 3), dtype=np.uint8) * 255
+
+    # Crop the original image using numpy slicing
+    cropped_img = input_np_img[crop_ymin:crop_ymax, crop_xmin:crop_xmax]
+
+    # Paste the cropped image onto the white background
+    return_image[crop_paste_y:crop_paste_y + (crop_ymax - crop_ymin),
+    crop_paste_x:crop_paste_x + (crop_xmax - crop_xmin)] = cropped_img
+
+    return_list = [crop_paste_x, crop_paste_y, crop_xmin, crop_ymin, crop_xmax, crop_ymax, crop_new_width,
+                   crop_new_height]
     return return_image, return_list
 
 

magic_pdf/model/sub_modules/ocr/paddleocr/ppocr_273_mod.py

@@ -5,6 +5,7 @@ import cv2
 import numpy as np
 import torch
 
+
 from paddleocr import PaddleOCR
 from ppocr.utils.logging import get_logger
 from ppocr.utils.utility import alpha_to_color, binarize_img

magic_pdf/model/sub_modules/table/tablemaster/tablemaster_paddle.py

@@ -2,6 +2,7 @@ import os
 
 import cv2
 import numpy as np
+from paddleocr import PaddleOCR
 from ppstructure.table.predict_table import TableSystem
 from ppstructure.utility import init_args
 from PIL import Image

magic_pdf/pdf_parse_union_core_v2.py

@@ -492,7 +492,7 @@ def insert_lines_into_block(block_bbox, line_height, page_w, page_h):
     else:
         return [[x0, y0, x1, y1]]
 
-# @measure_time
+
 def sort_lines_by_model(fix_blocks, page_w, page_h, line_height):
     page_line_list = []
 

magic_pdf/resources/model_config/model_configs.yaml

@@ -2,7 +2,7 @@ weights:
   layoutlmv3: Layout/LayoutLMv3/model_final.pth
   doclayout_yolo: Layout/YOLO/doclayout_yolo_docstructbench_imgsz1280_2501.pt
   yolo_v8_mfd: MFD/YOLO/yolo_v8_ft.pt
-  unimernet_small: MFR/unimernet_small_2501
+  unimernet_small: MFR/unimernet_hf_small_2503
   struct_eqtable: TabRec/StructEqTable
   tablemaster: TabRec/TableMaster
   rapid_table: TabRec/RapidTable
\ No newline at end of file

magic_pdf/tools/cli.py

 import os
 import shutil
 import tempfile
+from pathlib import Path
+
 import click
 import fitz
 from loguru import logger
-from pathlib import Path
 
 import magic_pdf.model as model_config
+from magic_pdf.data.batch_build_dataset import batch_build_dataset
 from magic_pdf.data.data_reader_writer import FileBasedDataReader
+from magic_pdf.data.dataset import Dataset
 from magic_pdf.libs.version import __version__
-from magic_pdf.tools.common import do_parse, parse_pdf_methods
+from magic_pdf.tools.common import batch_do_parse, do_parse, parse_pdf_methods
 from magic_pdf.utils.office_to_pdf import convert_file_to_pdf
 
 pdf_suffixes = ['.pdf']
@@ -94,30 +97,33 @@ def cli(path, output_dir, method, lang, debug_able, start_page_id, end_page_id):
     def read_fn(path: Path):
         if path.suffix in ms_office_suffixes:
             convert_file_to_pdf(str(path), temp_dir)
-            fn = os.path.join(temp_dir, f"{path.stem}.pdf")
+            fn = os.path.join(temp_dir, f'{path.stem}.pdf')
         elif path.suffix in image_suffixes:
             with open(str(path), 'rb') as f:
                 bits = f.read()
             pdf_bytes = fitz.open(stream=bits).convert_to_pdf()
-            fn = os.path.join(temp_dir, f"{path.stem}.pdf")
+            fn = os.path.join(temp_dir, f'{path.stem}.pdf')
             with open(fn, 'wb') as f:
                 f.write(pdf_bytes)
         elif path.suffix in pdf_suffixes:
             fn = str(path)
         else:
-            raise Exception(f"Unknown file suffix: {path.suffix}")
+            raise Exception(f'Unknown file suffix: {path.suffix}')
 
         disk_rw = FileBasedDataReader(os.path.dirname(fn))
         return disk_rw.read(os.path.basename(fn))
 
-    def parse_doc(doc_path: Path):
+    def parse_doc(doc_path: Path, dataset: Dataset | None = None):
         try:
             file_name = str(Path(doc_path).stem)
-            pdf_data = read_fn(doc_path)
+            if dataset is None:
+                pdf_data_or_dataset = read_fn(doc_path)
+            else:
+                pdf_data_or_dataset = dataset
             do_parse(
                 output_dir,
                 file_name,
-                pdf_data,
+                pdf_data_or_dataset,
                 [],
                 method,
                 debug_able,
@@ -130,9 +136,12 @@ def cli(path, output_dir, method, lang, debug_able, start_page_id, end_page_id):
             logger.exception(e)
 
     if os.path.isdir(path):
+        doc_paths = []
         for doc_path in Path(path).glob('*'):
             if doc_path.suffix in pdf_suffixes + image_suffixes + ms_office_suffixes:
-                parse_doc(doc_path)
+                doc_paths.append(doc_path)
+        datasets = batch_build_dataset(doc_paths, 4, lang)
+        batch_do_parse(output_dir, [str(doc_path.stem) for doc_path in doc_paths], datasets, method, debug_able, lang=lang)
     else:
         parse_doc(Path(path))
 

magic_pdf/tools/common.py

@@ -8,10 +8,10 @@ import magic_pdf.model as model_config
 from magic_pdf.config.enums import SupportedPdfParseMethod
 from magic_pdf.config.make_content_config import DropMode, MakeMode
 from magic_pdf.data.data_reader_writer import FileBasedDataWriter
-from magic_pdf.data.dataset import PymuDocDataset
+from magic_pdf.data.dataset import Dataset, PymuDocDataset
 from magic_pdf.libs.draw_bbox import draw_char_bbox
-from magic_pdf.model.doc_analyze_by_custom_model import doc_analyze
-from magic_pdf.operators.models import InferenceResult
+from magic_pdf.model.doc_analyze_by_custom_model import (batch_doc_analyze,
+                                                         doc_analyze)
 
 # from io import BytesIO
 # from pypdf import PdfReader, PdfWriter
@@ -67,10 +67,10 @@ def convert_pdf_bytes_to_bytes_by_pymupdf(pdf_bytes, start_page_id=0, end_page_i
     return output_bytes
 
 
-def do_parse(
+def _do_parse(
     output_dir,
     pdf_file_name,
-    pdf_bytes,
+    pdf_bytes_or_dataset,
     model_list,
     parse_method,
     debug_able,
@@ -92,16 +92,21 @@ def do_parse(
     formula_enable=None,
     table_enable=None,
 ):
+    from magic_pdf.operators.models import InferenceResult
     if debug_able:
         logger.warning('debug mode is on')
         f_draw_model_bbox = True
         f_draw_line_sort_bbox = True
         # f_draw_char_bbox = True
 
+    if isinstance(pdf_bytes_or_dataset, bytes):
         pdf_bytes = convert_pdf_bytes_to_bytes_by_pymupdf(
-        pdf_bytes, start_page_id, end_page_id
+            pdf_bytes_or_dataset, start_page_id, end_page_id
         )
-
+        ds = PymuDocDataset(pdf_bytes, lang=lang)
+    else:
+        ds = pdf_bytes_or_dataset
+    pdf_bytes = ds._raw_data
     local_image_dir, local_md_dir = prepare_env(output_dir, pdf_file_name, parse_method)
 
     image_writer, md_writer = FileBasedDataWriter(local_image_dir), FileBasedDataWriter(
@@ -109,8 +114,6 @@ def do_parse(
     )
     image_dir = str(os.path.basename(local_image_dir))
 
-    ds = PymuDocDataset(pdf_bytes, lang=lang)
-
     if len(model_list) == 0:
         if model_config.__use_inside_model__:
             if parse_method == 'auto':
@@ -241,5 +244,79 @@ def do_parse(
 
     logger.info(f'local output dir is {local_md_dir}')
 
+def do_parse(
+    output_dir,
+    pdf_file_name,
+    pdf_bytes_or_dataset,
+    model_list,
+    parse_method,
+    debug_able,
+    f_draw_span_bbox=True,
+    f_draw_layout_bbox=True,
+    f_dump_md=True,
+    f_dump_middle_json=True,
+    f_dump_model_json=True,
+    f_dump_orig_pdf=True,
+    f_dump_content_list=True,
+    f_make_md_mode=MakeMode.MM_MD,
+    f_draw_model_bbox=False,
+    f_draw_line_sort_bbox=False,
+    f_draw_char_bbox=False,
+    start_page_id=0,
+    end_page_id=None,
+    lang=None,
+    layout_model=None,
+    formula_enable=None,
+    table_enable=None,
+):
+    parallel_count = 1
+    if os.environ.get('MINERU_PARALLEL_INFERENCE_COUNT'):
+        parallel_count = int(os.environ['MINERU_PARALLEL_INFERENCE_COUNT'])
+
+    if parallel_count > 1:
+        if isinstance(pdf_bytes_or_dataset, bytes):
+            pdf_bytes = convert_pdf_bytes_to_bytes_by_pymupdf(
+                pdf_bytes_or_dataset, start_page_id, end_page_id
+            )
+            ds = PymuDocDataset(pdf_bytes, lang=lang)
+        else:
+            ds = pdf_bytes_or_dataset
+        batch_do_parse(output_dir, [pdf_file_name], [ds], parse_method, debug_able, f_draw_span_bbox=f_draw_span_bbox, f_draw_layout_bbox=f_draw_layout_bbox, f_dump_md=f_dump_md, f_dump_middle_json=f_dump_middle_json, f_dump_model_json=f_dump_model_json, f_dump_orig_pdf=f_dump_orig_pdf, f_dump_content_list=f_dump_content_list, f_make_md_mode=f_make_md_mode, f_draw_model_bbox=f_draw_model_bbox, f_draw_line_sort_bbox=f_draw_line_sort_bbox, f_draw_char_bbox=f_draw_char_bbox)
+    else:
+        _do_parse(output_dir, pdf_file_name, pdf_bytes_or_dataset, model_list, parse_method, debug_able, start_page_id=start_page_id, end_page_id=end_page_id, lang=lang, layout_model=layout_model, formula_enable=formula_enable, table_enable=table_enable,  f_draw_span_bbox=f_draw_span_bbox, f_draw_layout_bbox=f_draw_layout_bbox, f_dump_md=f_dump_md, f_dump_middle_json=f_dump_middle_json, f_dump_model_json=f_dump_model_json, f_dump_orig_pdf=f_dump_orig_pdf, f_dump_content_list=f_dump_content_list, f_make_md_mode=f_make_md_mode, f_draw_model_bbox=f_draw_model_bbox, f_draw_line_sort_bbox=f_draw_line_sort_bbox, f_draw_char_bbox=f_draw_char_bbox)
+
+
+def batch_do_parse(
+    output_dir,
+    pdf_file_names: list[str],
+    pdf_bytes_or_datasets: list[bytes | Dataset],
+    parse_method,
+    debug_able,
+    f_draw_span_bbox=True,
+    f_draw_layout_bbox=True,
+    f_dump_md=True,
+    f_dump_middle_json=True,
+    f_dump_model_json=True,
+    f_dump_orig_pdf=True,
+    f_dump_content_list=True,
+    f_make_md_mode=MakeMode.MM_MD,
+    f_draw_model_bbox=False,
+    f_draw_line_sort_bbox=False,
+    f_draw_char_bbox=False,
+    lang=None,
+    layout_model=None,
+    formula_enable=None,
+    table_enable=None,
+):
+    dss = []
+    for v in pdf_bytes_or_datasets:
+        if isinstance(v, bytes):
+            dss.append(PymuDocDataset(v, lang=lang))
+        else:
+            dss.append(v)
+    infer_results = batch_doc_analyze(dss, lang=lang, layout_model=layout_model, formula_enable=formula_enable, table_enable=table_enable)
+    for idx, infer_result in enumerate(infer_results):
+        _do_parse(output_dir, pdf_file_names[idx], dss[idx], infer_result.get_infer_res(), parse_method, debug_able, f_draw_span_bbox=f_draw_span_bbox, f_draw_layout_bbox=f_draw_layout_bbox, f_dump_md=f_dump_md, f_dump_middle_json=f_dump_middle_json, f_dump_model_json=f_dump_model_json, f_dump_orig_pdf=f_dump_orig_pdf, f_dump_content_list=f_dump_content_list, f_make_md_mode=f_make_md_mode, f_draw_model_bbox=f_draw_model_bbox, f_draw_line_sort_bbox=f_draw_line_sort_bbox, f_draw_char_bbox=f_draw_char_bbox)
+
 
 parse_pdf_methods = click.Choice(['ocr', 'txt', 'auto'])

projects/web_api/app.py

@@ -3,6 +3,7 @@ import os
 from base64 import b64encode
 from glob import glob
 from io import StringIO
+import tempfile
 from typing import Tuple, Union
 
 import uvicorn
@@ -10,11 +11,12 @@ from fastapi import FastAPI, HTTPException, UploadFile
 from fastapi.responses import JSONResponse
 from loguru import logger
 
+from magic_pdf.data.read_api import read_local_images, read_local_office
 import magic_pdf.model as model_config
 from magic_pdf.config.enums import SupportedPdfParseMethod
 from magic_pdf.data.data_reader_writer import DataWriter, FileBasedDataWriter
 from magic_pdf.data.data_reader_writer.s3 import S3DataReader, S3DataWriter
-from magic_pdf.data.dataset import PymuDocDataset
+from magic_pdf.data.dataset import ImageDataset, PymuDocDataset
 from magic_pdf.libs.config_reader import get_bucket_name, get_s3_config
 from magic_pdf.model.doc_analyze_by_custom_model import doc_analyze
 from magic_pdf.operators.models import InferenceResult
@@ -24,6 +26,9 @@ model_config.__use_inside_model__ = True
 
 app = FastAPI()
 
+pdf_extensions = [".pdf"]
+office_extensions = [".ppt", ".pptx", ".doc", ".docx"]
+image_extensions = [".png", ".jpg"]
 
 class MemoryDataWriter(DataWriter):
     def __init__(self):
@@ -46,8 +51,8 @@ class MemoryDataWriter(DataWriter):
 
 
 def init_writers(
-    pdf_path: str = None,
-    pdf_file: UploadFile = None,
+    file_path: str = None,
+    file: UploadFile = None,
     output_path: str = None,
     output_image_path: str = None,
 ) -> Tuple[
@@ -59,19 +64,19 @@ def init_writers(
     Initialize writers based on path type
 
     Args:
-        pdf_path: PDF file path (local path or S3 path)
-        pdf_file: Uploaded PDF file object
+        file_path: file path (local path or S3 path)
+        file: Uploaded file object
         output_path: Output directory path
         output_image_path: Image output directory path
 
     Returns:
-        Tuple[writer, image_writer, pdf_bytes]: Returns initialized writer tuple and PDF
-        file content
+        Tuple[writer, image_writer, file_bytes]: Returns initialized writer tuple and file content
     """
-    if pdf_path:
-        is_s3_path = pdf_path.startswith("s3://")
+    file_extension:str = None
+    if file_path:
+        is_s3_path = file_path.startswith("s3://")
         if is_s3_path:
-            bucket = get_bucket_name(pdf_path)
+            bucket = get_bucket_name(file_path)
             ak, sk, endpoint = get_s3_config(bucket)
 
             writer = S3DataWriter(
@@ -84,25 +89,29 @@ def init_writers(
             temp_reader = S3DataReader(
                 "", bucket=bucket, ak=ak, sk=sk, endpoint_url=endpoint
             )
-            pdf_bytes = temp_reader.read(pdf_path)
+            file_bytes = temp_reader.read(file_path)
+            file_extension = os.path.splitext(file_path)[1]
         else:
             writer = FileBasedDataWriter(output_path)
             image_writer = FileBasedDataWriter(output_image_path)
             os.makedirs(output_image_path, exist_ok=True)
-            with open(pdf_path, "rb") as f:
-                pdf_bytes = f.read()
+            with open(file_path, "rb") as f:
+                file_bytes = f.read()
+            file_extension = os.path.splitext(file_path)[1]
     else:
         # 处理上传的文件
-        pdf_bytes = pdf_file.file.read()
+        file_bytes = file.file.read()
+        file_extension = os.path.splitext(file.filename)[1]
         writer = FileBasedDataWriter(output_path)
         image_writer = FileBasedDataWriter(output_image_path)
         os.makedirs(output_image_path, exist_ok=True)
 
-    return writer, image_writer, pdf_bytes
+    return writer, image_writer, file_bytes, file_extension
 
 
-def process_pdf(
-    pdf_bytes: bytes,
+def process_file(
+    file_bytes: bytes,
+    file_extension: str,
     parse_method: str,
     image_writer: Union[S3DataWriter, FileBasedDataWriter],
 ) -> Tuple[InferenceResult, PipeResult]:
@@ -110,14 +119,30 @@ def process_pdf(
     Process PDF file content
 
     Args:
-        pdf_bytes: Binary content of PDF file
+        file_bytes: Binary content of file
+        file_extension: file extension
         parse_method: Parse method ('ocr', 'txt', 'auto')
         image_writer: Image writer
 
     Returns:
         Tuple[InferenceResult, PipeResult]: Returns inference result and pipeline result
     """
-    ds = PymuDocDataset(pdf_bytes)
+
+    ds = Union[PymuDocDataset, ImageDataset]
+    if file_extension in pdf_extensions:
+        ds = PymuDocDataset(file_bytes)
+    elif file_extension in office_extensions:
+        # 需要使用office解析
+        temp_dir = tempfile.mkdtemp()
+        with open(os.path.join(temp_dir, f"temp_file.{file_extension}"), "wb") as f:
+            f.write(file_bytes)
+        ds = read_local_office(temp_dir)[0]
+    elif file_extension in image_extensions:
+        # 需要使用ocr解析
+        temp_dir = tempfile.mkdtemp()
+        with open(os.path.join(temp_dir, f"temp_file.{file_extension}"), "wb") as f:
+            f.write(file_bytes)
+        ds = read_local_images(temp_dir)[0]
     infer_result: InferenceResult = None
     pipe_result: PipeResult = None
 
@@ -145,13 +170,13 @@ def encode_image(image_path: str) -> str:
 
 
 @app.post(
-    "/pdf_parse",
+    "/file_parse",
     tags=["projects"],
-    summary="Parse PDF files (supports local files and S3)",
+    summary="Parse files (supports local files and S3)",
 )
-async def pdf_parse(
-    pdf_file: UploadFile = None,
-    pdf_path: str = None,
+async def file_parse(
+    file: UploadFile = None,
+    file_path: str = None,
     parse_method: str = "auto",
     is_json_md_dump: bool = False,
     output_dir: str = "output",
@@ -165,10 +190,10 @@ async def pdf_parse(
     to the specified directory.
 
     Args:
-        pdf_file: The PDF file to be parsed. Must not be specified together with
-            `pdf_path`
-        pdf_path: The path to the PDF file to be parsed. Must not be specified together
-            with `pdf_file`
+        file: The PDF file to be parsed. Must not be specified together with
+            `file_path`
+        file_path: The path to the PDF file to be parsed. Must not be specified together
+            with `file`
         parse_method: Parsing method, can be auto, ocr, or txt. Default is auto. If
             results are not satisfactory, try ocr
         is_json_md_dump: Whether to write parsed data to .json and .md files. Default
@@ -181,31 +206,31 @@ async def pdf_parse(
         return_content_list: Whether to return parsed PDF content list. Default to False
     """
     try:
-        if (pdf_file is None and pdf_path is None) or (
-            pdf_file is not None and pdf_path is not None
+        if (file is None and file_path is None) or (
+            file is not None and file_path is not None
         ):
             return JSONResponse(
-                content={"error": "Must provide either pdf_file or pdf_path"},
+                content={"error": "Must provide either file or file_path"},
                 status_code=400,
             )
 
         # Get PDF filename
-        pdf_name = os.path.basename(pdf_path if pdf_path else pdf_file.filename).split(
+        file_name = os.path.basename(file_path if file_path else file.filename).split(
             "."
         )[0]
-        output_path = f"{output_dir}/{pdf_name}"
+        output_path = f"{output_dir}/{file_name}"
         output_image_path = f"{output_path}/images"
 
         # Initialize readers/writers and get PDF content
-        writer, image_writer, pdf_bytes = init_writers(
-            pdf_path=pdf_path,
-            pdf_file=pdf_file,
+        writer, image_writer, file_bytes, file_extension = init_writers(
+            file_path=file_path,
+            file=file,
             output_path=output_path,
             output_image_path=output_image_path,
         )
 
         # Process PDF
-        infer_result, pipe_result = process_pdf(pdf_bytes, parse_method, image_writer)
+        infer_result, pipe_result = process_file(file_bytes, file_extension, parse_method, image_writer)
 
         # Use MemoryDataWriter to get results
         content_list_writer = MemoryDataWriter()
@@ -226,23 +251,23 @@ async def pdf_parse(
         # If results need to be saved
         if is_json_md_dump:
             writer.write_string(
-                f"{pdf_name}_content_list.json", content_list_writer.get_value()
+                f"{file_name}_content_list.json", content_list_writer.get_value()
             )
-            writer.write_string(f"{pdf_name}.md", md_content)
+            writer.write_string(f"{file_name}.md", md_content)
             writer.write_string(
-                f"{pdf_name}_middle.json", middle_json_writer.get_value()
+                f"{file_name}_middle.json", middle_json_writer.get_value()
             )
             writer.write_string(
-                f"{pdf_name}_model.json",
+                f"{file_name}_model.json",
                 json.dumps(model_json, indent=4, ensure_ascii=False),
             )
             # Save visualization results
-            pipe_result.draw_layout(os.path.join(output_path, f"{pdf_name}_layout.pdf"))
-            pipe_result.draw_span(os.path.join(output_path, f"{pdf_name}_spans.pdf"))
+            pipe_result.draw_layout(os.path.join(output_path, f"{file_name}_layout.pdf"))
+            pipe_result.draw_span(os.path.join(output_path, f"{file_name}_spans.pdf"))
             pipe_result.draw_line_sort(
-                os.path.join(output_path, f"{pdf_name}_line_sort.pdf")
+                os.path.join(output_path, f"{file_name}_line_sort.pdf")
             )
-            infer_result.draw_model(os.path.join(output_path, f"{pdf_name}_model.pdf"))
+            infer_result.draw_model(os.path.join(output_path, f"{file_name}_model.pdf"))
 
         # Build return data
         data = {}

requirements.txt

@@ -7,7 +7,8 @@ numpy>=1.21.6,<2.0.0
 pydantic>=2.7.2
 PyMuPDF>=1.24.9,<=1.24.14
 scikit-learn>=1.0.2
-torch>=2.2.2
-transformers
+torch>=2.2.2,!=2.5.0,!=2.5.1,<=2.6.0
+torchvision
+transformers>=4.49.0
 pdfminer.six==20231228
 # The requirements.txt must ensure that only necessary external dependencies are introduced. If there are new dependencies to add, please contact the project administrator.

scripts/download_models.py

 import json
+import shutil
 import os
 
 import requests
@@ -16,7 +17,7 @@ def download_and_modify_json(url, local_filename, modifications):
     if os.path.exists(local_filename):
         data = json.load(open(local_filename))
         config_version = data.get('config_version', '0.0.0')
-        if config_version < '1.1.1':
+        if config_version < '1.2.0':
             data = download_json(url)
     else:
         data = download_json(url)
@@ -32,12 +33,13 @@ def download_and_modify_json(url, local_filename, modifications):
 
 if __name__ == '__main__':
     mineru_patterns = [
-        "models/Layout/LayoutLMv3/*",
+        # "models/Layout/LayoutLMv3/*",
         "models/Layout/YOLO/*",
         "models/MFD/YOLO/*",
-        "models/MFR/unimernet_small_2501/*",
-        "models/TabRec/TableMaster/*",
-        "models/TabRec/StructEqTable/*",
+        "models/MFR/unimernet_hf_small_2503/*",
+        "models/OCR/paddleocr/*",
+        # "models/TabRec/TableMaster/*",
+        # "models/TabRec/StructEqTable/*",
     ]
     model_dir = snapshot_download('opendatalab/PDF-Extract-Kit-1.0', allow_patterns=mineru_patterns)
     layoutreader_model_dir = snapshot_download('ppaanngggg/layoutreader')
@@ -45,6 +47,12 @@ if __name__ == '__main__':
     print(f'model_dir is: {model_dir}')
     print(f'layoutreader_model_dir is: {layoutreader_model_dir}')
 
+    paddleocr_model_dir = model_dir + '/OCR/paddleocr'
+    user_paddleocr_dir = os.path.expanduser('~/.paddleocr')
+    if os.path.exists(user_paddleocr_dir):
+        shutil.rmtree(user_paddleocr_dir)
+    shutil.copytree(paddleocr_model_dir, user_paddleocr_dir)
+
     json_url = 'https://gcore.jsdelivr.net/gh/opendatalab/MinerU@master/magic-pdf.template.json'
     config_file_name = 'magic-pdf.json'
     home_dir = os.path.expanduser('~')