PaddleOCR/main/version3.x/pipeline_usage/table_recognition_v2

[giscus[bot]]

PaddleOCR/main/version3.x/pipeline_usage/table_recognition_v2

Awesome multilingual OCR toolkits based on PaddlePaddle (practical ultra lightweight OCR system, support 80+ languages recognition, provide data annotation and synthesis tools, support training and deployment among server, mobile, embedded and IoT devices)

https://paddlepaddle.github.io/PaddleOCR/main/version3.x/pipeline_usage/table_recognition_v2.html

[mr-abei]

通过参数指定本地模型路径:pipeline = TableRecognitionPipelineV2(wired_table_structure_recognition_model_dir="./your_model_path")显示AssertionError: Model name mismatch，please input the correct model dir. 本地配置路径是同级目录./models/official_models/PP-LCNet_x1_0_doc_ori，目录里面包含inference.json、inference.pdiparams、inference.yml请求是哪里有问题呢？

[liuhongen1234567]

您好，您的表格结构化模型并没有指定正确，PP-LCNet_x1_0_doc_ori是文本方向模型，表格结构化模型应该 是SLANeXt_wired

[tsh2018]

self.pipeline = TableRecognitionPipelineV2(
use_doc_orientation_classify= False,
use_doc_unwarping=False,
layout_detection_model_name='PP-DocLayout-M',
layout_detection_model_dir=os.path.join(models_base, 'PP-DocLayout-M/'),
table_classification_model_name='PP-LCNet_x1_0_table_cls',
table_classification_model_dir=os.path.join(models_base, 'PP-LCNet_x1_0_table_cls/'),
wired_table_structure_recognition_model_name='SLANet_plus',
wired_table_structure_recognition_model_dir=os.path.join(models_base, 'SLANet_plus/'),
wireless_table_structure_recognition_model_name='SLANet_plus',
wireless_table_structure_recognition_model_dir=os.path.join(models_base, 'SLANet_plus/'),
wired_table_cells_detection_model_name='RT-DETR-L_wired_table_cell_det',
wired_table_cells_detection_model_dir=os.path.join(models_base, 'RT-DETR-L_wired_table_cell_det/'),
wireless_table_cells_detection_model_name='RT-DETR-L_wireless_table_cell_det',
wireless_table_cells_detection_model_dir=os.path.join(models_base, 'RT-DETR-L_wireless_table_cell_det/'),
text_detection_model_name='PP-OCRv5_server_det',
text_detection_model_dir=os.path.join(models_base, 'PP-OCRv5_server_det/'),
text_recognition_model_name='PP-OCRv5_server_rec',
text_recognition_model_dir=os.path.join(models_base, 'PP-OCRv5_server_rec/'),
# 显式提供路径，即使不使用该功能
doc_orientation_classify_model_name='PP-LCNet_x1_0_doc_ori',
doc_orientation_classify_model_dir=os.path.join(models_base, 'PP-LCNet_x1_0_doc_ori/')
)

use_doc_orientation_classify设置为False，仍旧会在初始化完成后，调用predict推理时自动下载模型PP-LCNet_x1_0_doc_ori， 即使制定了模型路径并且确保了PP-LCNet_x1_0_doc_ori模型文件正常的情况下，仍旧会下载。

2025-09-03 06:31:19.159 | INFO | group_paddle_ocr_pdf:_initialize_pipeline:37 - 使用模型基础路径: /paddle/models
grep: warning: GREP_OPTIONS is deprecated; please use an alias or script
Creating model: ('PP-LCNet_x1_0_doc_ori', '/paddle/models/PP-LCNet_x1_0_doc_ori/')
Creating model: ('PP-DocLayout-M', '/paddle/models/PP-DocLayout-M/')
Creating model: ('PP-LCNet_x1_0_table_cls', '/paddle/models/PP-LCNet_x1_0_table_cls/')
Creating model: ('SLANet_plus', '/paddle/models/SLANet_plus/')
Creating model: ('SLANet_plus', '/paddle/models/SLANet_plus/')
Creating model: ('RT-DETR-L_wired_table_cell_det', '/paddle/models/RT-DETR-L_wired_table_cell_det/')
Creating model: ('RT-DETR-L_wireless_table_cell_det', '/paddle/models/RT-DETR-L_wireless_table_cell_det/')
Creating model: ('PP-OCRv5_server_det', '/paddle/models/PP-OCRv5_server_det/')
Creating model: ('PP-OCRv5_server_rec', '/paddle/models/PP-OCRv5_server_rec/')
2025-09-03 06:31:37.230 | INFO | group_paddle_ocr_pdf:_initialize_pipeline:61 - GroupTableRecognitionPipelineV2初始化完成
2025-09-03 14:31:42.425 | INFO | main::1786 - 启动API服务
INFO: Started server process [159707]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:9900 (Press CTRL+C to quit)
2025-09-03 14:31:53.412 | INFO | main:extract_text_from_file:1070 - extract_text_from_file - 开始提取文件内容，类型: pdf
Creating model: ('PP-LCNet_x1_0_doc_ori', None)
Using official model (PP-LCNet_x1_0_doc_ori), the model files will be automatically downloaded and saved in /root/.paddlex/official_models.
2025-09-03 14:32:17.952 | INFO | group_paddle_ocr_pdf:extract_char_coordinates_from_pdf:434 - PDF处理完成:

这是底层bug吗，use_doc_orientation_classify为True和False都会自动下载模型

[hyhwmt]

你好，我遇到了相同的问题，不管设置为true还是false，都会尝试联网下载模型，并且显式指定模型路径无效，请问你解决了吗

[tsh2018]

没有，大概率是bug，你可以升级3.2.0 ，这个版本虽然没法指定模型路径，但是会在默认路径下先找一下有没有模型，没有模型才会下载

[freebear666]

请问我使用了自定义数据集，在预训练ocrv5的基础上进行了微调，并将该模型导入了表格识别产线，但是识别出来的结果全是乱码，产线配置文件中也没有字符集指定地址参数，是哪个环节出了问题了吗

[liuhongen1234567]

您好，对应的字符集在导出模型的inference.yaml 里，推理时会自动加载，不需要额外指定。可以看看是不是训练集太小，识别模型性能变差导致的