OpenVino 无法运行ch_server模型

[ok-oldking]

问题描述 / Problem Description

使用ch_mobile, 或者使用onnx+ch_server都正常,
ch_server+openvino会报异常

 File "ok\\__init__.pyx", line 3128, in ok.OCR.rapid_ocr
    result = self.executor.ocr_lib(image, use_det=True, use_cls=False, use_rec=True)
  File "D:\projects\sanmou\.venv\Lib\site-packages\rapidocr\main.py", line 118, in __call__
    det_res = self.text_det(img)
              ^^^^^^^^^^^^^^^^^^
  File "D:\projects\sanmou\.venv\Lib\site-packages\rapidocr\ch_ppocr_det\main.py", line 59, in __call__
    preds = self.session(prepro_img)
            ^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\projects\sanmou\.venv\Lib\site-packages\rapidocr\inference_engine\openvino.py", line 53, in __call__
    raise OpenVIONError(error_info) from e
rapidocr.inference_engine.openvino.OpenVIONError: Traceback (most recent call last):
  File "D:\projects\sanmou\.venv\Lib\site-packages\rapidocr\inference_engine\openvino.py", line 50, in __call__
    return self.session.get_output_tensor().data
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Exception from src\inference\src\cpp\infer_request.cpp:219:
Check 'outputs.size() == 1' failed at src\inference\src\cpp\infer_request.cpp:219:
get_output_tensor() must be called on a function with exactly one parameter.

运行环境 / Runtime Environment

复现代码 / Reproduction Code

可能解决方案 / Possible solutions

[SWHL]

麻烦给出具体使用的环境，包括平台，版本，复现代码？

[ok-oldking]

windows11, rapidocr 2.0.6, openvino 2025.1.0

from rapidocr import RapidOCR

engine = RapidOCR(
                    params={"Global.use_cls": False,
                            'Global.with_openvino': True,
                            # "Global.max_side_len": 100000, "Global.min_side_len": 0,
                            # 'Global.with_onnx': True,
                            # "EngineConfig.onnxruntime.use_dml": True
                            'Global.lang_det': 'ch_server',
                            'Global.lang_rec': 'ch_server',
                            # "EngineConfig.openvino.inference_num_threads": 1
                            })


# {"Global.use_cls": False, "Global.max_side_len": 100000, "Global.min_side_len": 0,
#                           "EngineConfig.onnxruntime.use_dml": og.use_dml}
# img_url = "member.png"
# img_url = "land.png"
import cv2

img_url = "land_bw.png"
# img_url = "member_croped.png"
# img_url = "land.png"
img = cv2.imread(img_url)
result = engine(img_url, use_det=True, use_cls=False, use_rec=True)

[SWHL]

等我有空试试，最近openvino在macOS 15.4上各个版本都不能用了。。。。。。我这里测试不了还。

[SWHL]

经过排查是det server模型在OpenVINO上不太行。暂时无解。

import cv2
from rapidocr import RapidOCR

engine = RapidOCR(
                    params={"Global.use_cls": False,
                            'Global.with_openvino': True,
                            'Global.lang_det': 'ch_server',
                            # 'Global.lang_rec': 'ch_server',
                            })

img_url = "https://img1.baidu.com/it/u=3619974146,1266987475&fm=253&fmt=auto&app=138&f=JPEG?w=500&h=516"
result = engine(img_url, use_det=True, use_cls=False, use_rec=True)
result.vis()

[cole-dda]

这么修改就可以

因为返回2个tensor，所以必须指定参数，用哪一个

ps：为什么会返回2个，没有深入研究

[SWHL]

感谢感谢，有空我验证一下