magic-pdf 1.2.2__py3-none-any.whl → 1.3.0__py3-none-any.whl

magic_pdf/model/doc_analyze_by_custom_model.py

@@ -1,27 +1,19 @@
 import os
 import time
+
+import numpy as np
 import torch
 
 os.environ['FLAGS_npu_jit_compile'] = '0'  # 关闭paddle的jit编译
 os.environ['FLAGS_use_stride_kernel'] = '0'
 os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1'  # 让mps可以fallback
 os.environ['NO_ALBUMENTATIONS_UPDATE'] = '1'  # 禁止albumentations检查更新
-# 关闭paddle的信号处理
-import paddle
-paddle.disable_signal_handler()
+
 
 from loguru import logger
 
-from magic_pdf.model.batch_analyze import BatchAnalyze
 from magic_pdf.model.sub_modules.model_utils import get_vram
-
-try:
-    import torchtext
-    if torchtext.__version__ >= '0.18.0':
-        torchtext.disable_torchtext_deprecation_warning()
-except ImportError:
-    pass
-
+from magic_pdf.config.enums import SupportedPdfParseMethod
 import magic_pdf.model as model_config
 from magic_pdf.data.dataset import Dataset
 from magic_pdf.libs.clean_memory import clean_memory
@@ -30,8 +22,6 @@ from magic_pdf.libs.config_reader import (get_device, get_formula_config,
                                           get_local_models_dir,
                                           get_table_recog_config)
 from magic_pdf.model.model_list import MODEL
-from magic_pdf.operators.models import InferenceResult
-
 
 class ModelSingleton:
     _instance = None
@@ -72,9 +62,7 @@ def custom_model_init(
     formula_enable=None,
     table_enable=None,
 ):
-
     model = None
-
     if model_config.__model_mode__ == 'lite':
         logger.warning(
             'The Lite mode is provided for developers to conduct testing only, and the output quality is '
@@ -132,7 +120,6 @@ def custom_model_init(
 
     return custom_model
 
-
 def doc_analyze(
     dataset: Dataset,
     ocr: bool = False,
@@ -143,102 +130,160 @@ def doc_analyze(
     layout_model=None,
     formula_enable=None,
     table_enable=None,
-) -> InferenceResult:
-
+):
     end_page_id = (
         end_page_id
         if end_page_id is not None and end_page_id >= 0
         else len(dataset) - 1
     )
 
-    model_manager = ModelSingleton()
-    custom_model = model_manager.get_model(
-        ocr, show_log, lang, layout_model, formula_enable, table_enable
-    )
-
-    batch_analyze = False
-    batch_ratio = 1
-    device = get_device()
+    MIN_BATCH_INFERENCE_SIZE = int(os.environ.get('MINERU_MIN_BATCH_INFERENCE_SIZE', 200))
+    images = []
+    page_wh_list = []
+    for index in range(len(dataset)):
+        if start_page_id <= index <= end_page_id:
+            page_data = dataset.get_page(index)
+            img_dict = page_data.get_image()
+            images.append(img_dict['img'])
+            page_wh_list.append((img_dict['width'], img_dict['height']))
+    if lang is None or lang == 'auto':
+        images_with_extra_info = [(images[index], ocr, dataset._lang) for index in range(len(dataset))]
+    else:
+        images_with_extra_info = [(images[index], ocr, lang) for index in range(len(dataset))]
 
-    npu_support = False
-    if str(device).startswith("npu"):
-        import torch_npu
-        if torch_npu.npu.is_available():
-            npu_support = True
+    if len(images) >= MIN_BATCH_INFERENCE_SIZE:
+        batch_size = MIN_BATCH_INFERENCE_SIZE
+        batch_images = [images_with_extra_info[i:i+batch_size] for i in range(0, len(images_with_extra_info), batch_size)]
+    else:
+        batch_images = [images_with_extra_info]
 
-    if torch.cuda.is_available() and device != 'cpu' or npu_support:
-        gpu_memory = int(os.getenv("VIRTUAL_VRAM_SIZE", round(get_vram(device))))
-        if gpu_memory is not None and gpu_memory >= 8:
+    results = []
+    for sn, batch_image in enumerate(batch_images):
+        _, result = may_batch_image_analyze(batch_image, sn, ocr, show_log,layout_model, formula_enable, table_enable)
+        results.extend(result)
 
-            if gpu_memory >= 16:
-                batch_ratio = 8
-            elif gpu_memory >= 10:
-                batch_ratio = 4
-            else:
-                batch_ratio = 2
+    model_json = []
+    for index in range(len(dataset)):
+        if start_page_id <= index <= end_page_id:
+            result = results.pop(0)
+            page_width, page_height = page_wh_list.pop(0)
+        else:
+            result = []
+            page_height = 0
+            page_width = 0
 
-            logger.info(f'gpu_memory: {gpu_memory} GB, batch_ratio: {batch_ratio}')
-            batch_analyze = True
+        page_info = {'page_no': index, 'width': page_width, 'height': page_height}
+        page_dict = {'layout_dets': result, 'page_info': page_info}
+        model_json.append(page_dict)
 
-    model_json = []
-    doc_analyze_start = time.time()
+    from magic_pdf.operators.models import InferenceResult
+    return InferenceResult(model_json, dataset)
 
-    if batch_analyze:
-        # batch analyze
-        images = []
-        page_wh_list = []
-        for index in range(len(dataset)):
-            if start_page_id <= index <= end_page_id:
-                page_data = dataset.get_page(index)
-                img_dict = page_data.get_image()
-                images.append(img_dict['img'])
-                page_wh_list.append((img_dict['width'], img_dict['height']))
-        batch_model = BatchAnalyze(model=custom_model, batch_ratio=batch_ratio)
-        analyze_result = batch_model(images)
+def batch_doc_analyze(
+    datasets: list[Dataset],
+    parse_method: str,
+    show_log: bool = False,
+    lang=None,
+    layout_model=None,
+    formula_enable=None,
+    table_enable=None,
+):
+    MIN_BATCH_INFERENCE_SIZE = int(os.environ.get('MINERU_MIN_BATCH_INFERENCE_SIZE', 200))
+    batch_size = MIN_BATCH_INFERENCE_SIZE
+    images = []
+    page_wh_list = []
 
+    images_with_extra_info = []
+    for dataset in datasets:
         for index in range(len(dataset)):
-            if start_page_id <= index <= end_page_id:
-                result = analyze_result.pop(0)
-                page_width, page_height = page_wh_list.pop(0)
+            if lang is None or lang == 'auto':
+                _lang = dataset._lang
             else:
-                result = []
-                page_height = 0
-                page_width = 0
+                _lang = lang
 
-            page_info = {'page_no': index, 'width': page_width, 'height': page_height}
+            page_data = dataset.get_page(index)
+            img_dict = page_data.get_image()
+            images.append(img_dict['img'])
+            page_wh_list.append((img_dict['width'], img_dict['height']))
+            if parse_method == 'auto':
+                images_with_extra_info.append((images[-1], dataset.classify() == SupportedPdfParseMethod.OCR, _lang))
+            else:
+                images_with_extra_info.append((images[-1], parse_method == 'ocr', _lang))
+
+    batch_images = [images_with_extra_info[i:i+batch_size] for i in range(0, len(images_with_extra_info), batch_size)]
+    results = []
+    for sn, batch_image in enumerate(batch_images):
+        _, result = may_batch_image_analyze(batch_image, sn, True, show_log, layout_model, formula_enable, table_enable)
+        results.extend(result)
+
+    infer_results = []
+    from magic_pdf.operators.models import InferenceResult
+    for index in range(len(datasets)):
+        dataset = datasets[index]
+        model_json = []
+        for i in range(len(dataset)):
+            result = results.pop(0)
+            page_width, page_height = page_wh_list.pop(0)
+            page_info = {'page_no': i, 'width': page_width, 'height': page_height}
             page_dict = {'layout_dets': result, 'page_info': page_info}
             model_json.append(page_dict)
+        infer_results.append(InferenceResult(model_json, dataset))
+    return infer_results
 
-    else:
-        # single analyze
 
-        for index in range(len(dataset)):
-            page_data = dataset.get_page(index)
-            img_dict = page_data.get_image()
-            img = img_dict['img']
-            page_width = img_dict['width']
-            page_height = img_dict['height']
-            if start_page_id <= index <= end_page_id:
-                page_start = time.time()
-                result = custom_model(img)
-                logger.info(f'-----page_id : {index}, page total time: {round(time.time() - page_start, 2)}-----')
+def may_batch_image_analyze(
+        images_with_extra_info: list[(np.ndarray, bool, str)],
+        idx: int,
+        ocr: bool,
+        show_log: bool = False,
+        layout_model=None,
+        formula_enable=None,
+        table_enable=None):
+    # os.environ['CUDA_VISIBLE_DEVICES'] = str(idx)
+
+    from magic_pdf.model.batch_analyze import BatchAnalyze
+
+    model_manager = ModelSingleton()
+
+    # images = [image for image, _, _ in images_with_extra_info]
+    batch_ratio = 1
+    device = get_device()
+
+    if str(device).startswith('npu'):
+        import torch_npu
+        if torch_npu.npu.is_available():
+            torch.npu.set_compile_mode(jit_compile=False)
+
+    if str(device).startswith('npu') or str(device).startswith('cuda'):
+        gpu_memory = int(os.getenv('VIRTUAL_VRAM_SIZE', round(get_vram(device))))
+        if gpu_memory is not None:
+            if gpu_memory >= 16:
+                batch_ratio = 16
+            elif gpu_memory >= 12:
+                batch_ratio = 8
+            elif gpu_memory >= 8:
+                batch_ratio = 4
+            elif gpu_memory >= 6:
+                batch_ratio = 2
             else:
-                result = []
+                batch_ratio = 1
+            logger.info(f'gpu_memory: {gpu_memory} GB, batch_ratio: {batch_ratio}')
 
-            page_info = {'page_no': index, 'width': page_width, 'height': page_height}
-            page_dict = {'layout_dets': result, 'page_info': page_info}
-            model_json.append(page_dict)
 
-    gc_start = time.time()
-    clean_memory(get_device())
-    gc_time = round(time.time() - gc_start, 2)
-    logger.info(f'gc time: {gc_time}')
-
-    doc_analyze_time = round(time.time() - doc_analyze_start, 2)
-    doc_analyze_speed = round((end_page_id + 1 - start_page_id) / doc_analyze_time, 2)
-    logger.info(
-        f'doc analyze time: {round(time.time() - doc_analyze_start, 2)},'
-        f' speed: {doc_analyze_speed} pages/second'
-    )
+    # doc_analyze_start = time.time()
 
-    return InferenceResult(model_json, dataset)
+    batch_model = BatchAnalyze(model_manager, batch_ratio, show_log, layout_model, formula_enable, table_enable)
+    results = batch_model(images_with_extra_info)
+
+    # gc_start = time.time()
+    clean_memory(get_device())
+    # gc_time = round(time.time() - gc_start, 2)
+    # logger.debug(f'gc time: {gc_time}')
+
+    # doc_analyze_time = round(time.time() - doc_analyze_start, 2)
+    # doc_analyze_speed = round(len(images) / doc_analyze_time, 2)
+    # logger.debug(
+    #     f'doc analyze time: {round(time.time() - doc_analyze_start, 2)},'
+    #     f' speed: {doc_analyze_speed} pages/second'
+    # )
+    return idx, results

magic_pdf/model/pdf_extract_kit.py

@@ -3,28 +3,18 @@ import os
 import time
 
 import cv2
-import numpy as np
 import torch
 import yaml
 from loguru import logger
-from PIL import Image
 
 os.environ['NO_ALBUMENTATIONS_UPDATE'] = '1'  # 禁止albumentations检查更新
 
-try:
-    import torchtext
-
-    if torchtext.__version__ >= '0.18.0':
-        torchtext.disable_torchtext_deprecation_warning()
-except ImportError:
-    pass
-
 from magic_pdf.config.constants import *
 from magic_pdf.model.model_list import AtomicModel
 from magic_pdf.model.sub_modules.model_init import AtomModelSingleton
 from magic_pdf.model.sub_modules.model_utils import (
     clean_vram, crop_img, get_res_list_from_layout_res)
-from magic_pdf.model.sub_modules.ocr.paddleocr.ocr_utils import (
+from magic_pdf.model.sub_modules.ocr.paddleocr2pytorch.ocr_utils import (
     get_adjusted_mfdetrec_res, get_ocr_result_list)
 
 
@@ -120,7 +110,7 @@ class CustomPEKModel:
                 atom_model_name=AtomicModel.MFR,
                 mfr_weight_dir=mfr_weight_dir,
                 mfr_cfg_path=mfr_cfg_path,
-                device='cpu' if str(self.device).startswith("mps") else self.device,
+                device=self.device,
             )
 
         # 初始化layout模型
@@ -174,11 +164,6 @@ class CustomPEKModel:
         logger.info('DocAnalysis init done!')
 
     def __call__(self, image):
-
-        pil_img = Image.fromarray(image)
-        width, height = pil_img.size
-        # logger.info(f'width: {width}, height: {height}')
-
         # layout检测
         layout_start = time.time()
         layout_res = []
@@ -186,24 +171,6 @@ class CustomPEKModel:
             # layoutlmv3
             layout_res = self.layout_model(image, ignore_catids=[])
         elif self.layout_model_name == MODEL_NAME.DocLayout_YOLO:
-            # doclayout_yolo
-            # if height > width:
-            #     input_res = {"poly":[0,0,width,0,width,height,0,height]}
-            #     new_image, useful_list = crop_img(input_res, pil_img, crop_paste_x=width//2, crop_paste_y=0)
-            #     paste_x, paste_y, xmin, ymin, xmax, ymax, new_width, new_height = useful_list
-            #     layout_res = self.layout_model.predict(new_image)
-            #     for res in layout_res:
-            #         p1, p2, p3, p4, p5, p6, p7, p8 = res['poly']
-            #         p1 = p1 - paste_x + xmin
-            #         p2 = p2 - paste_y + ymin
-            #         p3 = p3 - paste_x + xmin
-            #         p4 = p4 - paste_y + ymin
-            #         p5 = p5 - paste_x + xmin
-            #         p6 = p6 - paste_y + ymin
-            #         p7 = p7 - paste_x + xmin
-            #         p8 = p8 - paste_y + ymin
-            #         res['poly'] = [p1, p2, p3, p4, p5, p6, p7, p8]
-            # else:
             layout_res = self.layout_model.predict(image)
 
         layout_cost = round(time.time() - layout_start, 2)
@@ -234,11 +201,11 @@ class CustomPEKModel:
         ocr_start = time.time()
         # Process each area that requires OCR processing
         for res in ocr_res_list:
-            new_image, useful_list = crop_img(res, pil_img, crop_paste_x=50, crop_paste_y=50)
+            new_image, useful_list = crop_img(res, image, crop_paste_x=50, crop_paste_y=50)
             adjusted_mfdetrec_res = get_adjusted_mfdetrec_res(single_page_mfdetrec_res, useful_list)
 
             # OCR recognition
-            new_image = cv2.cvtColor(np.asarray(new_image), cv2.COLOR_RGB2BGR)
+            new_image = cv2.cvtColor(new_image, cv2.COLOR_RGB2BGR)
 
             if self.apply_ocr:
                 ocr_res = self.ocr_model.ocr(new_image, mfd_res=adjusted_mfdetrec_res)[0]
@@ -260,7 +227,7 @@ class CustomPEKModel:
         if self.apply_table:
             table_start = time.time()
             for res in table_res_list:
-                new_image, _ = crop_img(res, pil_img)
+                new_image, _ = crop_img(res, image)
                 single_table_start_time = time.time()
                 html_code = None
                 if self.table_model_name == MODEL_NAME.STRUCT_EQTABLE:

magic_pdf/model/sub_modules/language_detection/utils.py

@@ -3,8 +3,6 @@ import os
 from pathlib import Path
 
 import yaml
-from PIL import Image
-
 os.environ['NO_ALBUMENTATIONS_UPDATE'] = '1'  # 禁止albumentations检查更新
 
 from magic_pdf.config.constants import MODEL_NAME
@@ -42,7 +40,7 @@ def get_text_images(simple_images):
     )
     text_images = []
     for simple_image in simple_images:
-        image = Image.fromarray(simple_image['img'])
+        image = simple_image['img']
         layout_res = temp_layout_model.predict(image)
         # 给textblock截图
         for res in layout_res:
@@ -51,7 +49,7 @@ def get_text_images(simple_images):
                 # 初步清洗（宽和高都小于100）
                 if x2 - x1 < 100 and y2 - y1 < 100:
                     continue
-                text_images.append(image.crop((x1, y1, x2, y2)))
+                text_images.append(image[y1:y2, x1:x2])
     return text_images
 
 

magic_pdf/model/sub_modules/language_detection/yolov11/YOLOv11.py

@@ -2,9 +2,9 @@
 import time
 from collections import Counter
 from uuid import uuid4
-
+import cv2
+import numpy as np
 import torch
-from PIL import Image
 from loguru import logger
 from ultralytics import YOLO
 
@@ -29,7 +29,7 @@ def split_images(image, result_images=None):
     if result_images is None:
         result_images = []
 
-    width, height = image.size
+    height, width = image.shape[:2]
     long_side = max(width, height)  # 获取较长边长度
 
     if long_side <= 400:
@@ -44,16 +44,14 @@ def split_images(image, result_images=None):
             # 判断裁剪区域是否超出图片范围，如果超出则不进行裁剪保存操作
             if x + new_long_side > width:
                 continue
-            box = (x, 0, x + new_long_side, height)
-            sub_image = image.crop(box)
+            sub_image = image[0:height, x:x + new_long_side]
             sub_images.append(sub_image)
     else:  # 如果高度是较长边
         for y in range(0, height, new_long_side):
             # 判断裁剪区域是否超出图片范围，如果超出则不进行裁剪保存操作
             if y + new_long_side > height:
                 continue
-            box = (0, y, width, y + new_long_side)
-            sub_image = image.crop(box)
+            sub_image = image[y:y + new_long_side, 0:width]
             sub_images.append(sub_image)
 
     for sub_image in sub_images:
@@ -64,24 +62,32 @@ def split_images(image, result_images=None):
 
 def resize_images_to_224(image):
     """
-    若分辨率小于224则用黑色背景补齐到224*224大小,若大于等于224则调整为224*224大小,并保存到输出文件夹中。
+    若分辨率小于224则用黑色背景补齐到224*224大小,若大于等于224则调整为224*224大小。
+    Works directly with NumPy arrays.
     """
     try:
-        width, height = image.size
+        height, width = image.shape[:2]
+
         if width < 224 or height < 224:
-            new_image = Image.new('RGB', (224, 224), (0, 0, 0))
-            paste_x = (224 - width) // 2
-            paste_y = (224 - height) // 2
-            new_image.paste(image, (paste_x, paste_y))
+            # Create black background
+            new_image = np.zeros((224, 224, 3), dtype=np.uint8)
+            # Calculate paste position (ensure they're not negative)
+            paste_x = max(0, (224 - width) // 2)
+            paste_y = max(0, (224 - height) // 2)
+            # Make sure we don't exceed the boundaries of new_image
+            paste_width = min(width, 224)
+            paste_height = min(height, 224)
+            # Paste original image onto black background
+            new_image[paste_y:paste_y + paste_height, paste_x:paste_x + paste_width] = image[:paste_height, :paste_width]
             image = new_image
         else:
-            image = image.resize((224, 224), Image.Resampling.LANCZOS)
+            # Resize using cv2
+            image = cv2.resize(image, (224, 224), interpolation=cv2.INTER_LANCZOS4)
 
-        # uuid = str(uuid4())
-        # image.save(f"/tmp/{uuid}.jpg")
         return image
     except Exception as e:
-        logger.exception(e)
+        logger.exception(f"Error in resize_images_to_224: {e}")
+        return None
 
 
 class YOLOv11LangDetModel(object):
@@ -96,8 +102,7 @@ class YOLOv11LangDetModel(object):
     def do_detect(self, images: list):
         all_images = []
         for image in images:
-            width, height = image.size
-            # logger.info(f"image size: {width} x {height}")
+            height, width = image.shape[:2]
             if width < 100 and height < 100:
                 continue
             temp_images = split_images(image)

magic_pdf/model/sub_modules/layout/doclayout_yolo/DocLayoutYOLO.py

@@ -1,4 +1,5 @@
 from doclayout_yolo import YOLOv10
+from tqdm import tqdm
 
 
 class DocLayoutYOLOModel(object):
@@ -31,7 +32,8 @@ class DocLayoutYOLOModel(object):
 
     def batch_predict(self, images: list, batch_size: int) -> list:
         images_layout_res = []
-        for index in range(0, len(images), batch_size):
+        # for index in range(0, len(images), batch_size):
+        for index in tqdm(range(0, len(images), batch_size), desc="Layout Predict"):
             doclayout_yolo_res = [
                 image_res.cpu()
                 for image_res in self.model.predict(

magic_pdf/model/sub_modules/mfd/yolov8/YOLOv8.py

@@ -1,3 +1,4 @@
+from tqdm import tqdm
 from ultralytics import YOLO
 
 
@@ -14,7 +15,8 @@ class YOLOv8MFDModel(object):
 
     def batch_predict(self, images: list, batch_size: int) -> list:
         images_mfd_res = []
-        for index in range(0, len(images), batch_size):
+        # for index in range(0, len(images), batch_size):
+        for index in tqdm(range(0, len(images), batch_size), desc="MFD Predict"):
             mfd_res = [
                 image_res.cpu()
                 for image_res in self.mfd_model.predict(