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)
kwdchol
commented
3 days ago
🔎 Search before asking
🐛 Bug (问题描述)
同一张图，同一个模型，同一种配置，python只用1秒，c++要用5秒多。 其他图片都能够复现运行慢的问题。
🏃♂️ Environment (运行环境)
C++用的是最新的paddle引擎的release版本2.8.1，在VS2022下编译的。 python也是2.8.1 Name: paddleocr Version: 2.8.1 Summary: Awesome OCR toolkits based on PaddlePaddle(8.6M ultra-lightweight pre-trained model, support training and deployment among server, mobile, embedded and IoT devices) Home-page: https://github.com/PaddlePaddle/PaddleOCR Author: Author-email: PaddlePaddle Paddle-better@baidu.com License: Apache License 2.0 Location: C:\Users\Remy\AppData\Local\Programs\Python\Python312\Lib\site-packages Requires: beautifulsoup4, cython, fire, fonttools, imgaug, lmdb, numpy, opencv-contrib-python, opencv-python, Pillow, pyclipper, python-docx, pyyaml, rapidfuzz, requests, scikit-image, shapely, tqdm
针对于识别模型参数：
C++参数如下： this 0x000001ecde9646c0 {predictor_=empty usegpu=false gpuid=0 ...} PaddleOCR::CRNNRecognizer * predictor_ empty std::shared_ptr
usegpu false bool
gpuid 0 int
gpumem 4000 int
cpu_math_library_numthreads 10 int
usemkldnn false bool
labellist { size=6625 } std::vector<std::string,std::allocator>
mean { size=3 } std::vector<float,std::allocator>
scale { size=3 } std::vector<float,std::allocator>
isscale true bool
usetensorrt false bool
precision_ "fp16" std::string
rec_batchnum 6 int
rec_imgh 48 int
rec_imgw 320 int
rec_imageshape { size=3 } std::vector<int,std::allocator>
resizeop {...} PaddleOCR::CrnnResizeImg
normalizeop {...} PaddleOCR::Normalize
permuteop {...} PaddleOCR::PermuteBatch
cpu_math_library_num_threads 10 const int &
gpu_id 0 const int &
gpu_mem 4000 const int &
label_path "D:/Code/gitClone/PaddleOCR-2.8.1/deploy/cpp_infer/build/Release/ppocr_keys_v1.txt" const std::string &
model_dir "D:\softwarePake\ch_PP-OCRv3_rec_infer" const std::string &
precision "fp16" const std::string &
rec_batch_num 6 const int &
rec_image_shape { size=3 } std::vector<int,std::allocator>
rec_img_h 48 const int &
rec_img_w 320 const int &
use_gpu false const bool &
use_mkldnn false const bool &
use_tensorrt false const bool &
python参数： Namespace(help='==SUPPRESS==', use_gpu=False, use_xpu=False, use_npu=False, use_mlu=False, ir_optim=True, use_tensorrt=False, min_subgraph_size=15, precision='fp32', gpu_mem=500, gpu_id=0, image_dir=None, page_num=0, det_algorithm='DB', det_model_dir='C:\Users\44684/.paddleocr/whl\det\ch\ch_PP-OCRv4_det_infer', det_limit_side_len=960, det_limit_type='max', det_box_type='quad', det_db_thresh=0.3, det_db_box_thresh=0.6, det_db_unclip_ratio=1.5, max_batch_size=10, use_dilation=False, det_db_score_mode='fast', det_east_score_thresh=0.8, det_east_cover_thresh=0.1, det_east_nms_thresh=0.2, det_sast_score_thresh=0.5, det_sast_nms_thresh=0.2, det_pse_thresh=0, det_pse_box_thresh=0.85, det_pse_min_area=16, det_pse_scale=1, scales=[8, 16, 32], alpha=1.0, beta=1.0, fourier_degree=5, rec_algorithm='SVTR_LCNet', rec_model_dir='C:\Users\44684/.paddleocr/whl\rec\ch\ch_PP-OCRv4_rec_infer', rec_image_inverse=True, rec_image_shape='3, 48, 320', rec_batch_num=6, max_text_length=25, rec_char_dict_path='c:\Users\44684\AppData\Local\Programs\Python\Python312\Lib\site-packages\paddleocr\ppocr\utils\ppocr_keys_v1.txt', use_space_char=True, vis_font_path='./doc/fonts/simfang.ttf', drop_score=0.5, e2e_algorithm='PGNet', e2e_model_dir=None, e2e_limit_side_len=768, e2e_limit_type='max', e2e_pgnet_score_thresh=0.5, e2e_char_dict_path='./ppocr/utils/ic15_dict.txt', e2e_pgnet_valid_set='totaltext', e2e_pgnet_mode='fast', use_angle_cls=True, cls_model_dir='C:\Users\44684/.paddleocr/whl\cls\ch_ppocr_mobile_v2.0_cls_infer', cls_image_shape='3, 48, 192', label_list=['0', '180'], cls_batch_num=6, cls_thresh=0.9, enable_mkldnn=False, cpu_threads=10, use_pdserving=False, warmup=False, sr_model_dir=None, sr_image_shape='3, 32, 128', sr_batch_num=1, draw_img_save_dir='./inference_results', save_crop_res=False, crop_res_save_dir='./output', use_mp=False, total_process_num=1, process_id=0, benchmark=False, save_log_path='./log_output/', show_log=True, use_onnx=False, return_word_box=False, output='./output', table_max_len=488, table_algorithm='TableAttn', table_model_dir=None, merge_no_span_structure=True, table_char_dict_path=None, layout_model_dir=None, layout_dict_path=None, layout_score_threshold=0.5, layout_nms_threshold=0.5, kie_algorithm='LayoutXLM', ser_model_dir=None, re_model_dir=None, use_visual_backbone=True, ser_dict_path='../train_data/XFUND/class_list_xfun.txt', ocr_order_method=None, mode='structure', image_orientation=False, layout=True, table=True, ocr=True, recovery=False, use_pdf2docx_api=False, invert=False, binarize=False, alphacolor=(255, 255, 255), lang='ch', det=True, rec=True, type='ocr', savefile=False, ocr_version='PP-OCRv4', structure_version='PP-StructureV2')
🌰 Minimal Reproducible Example (最小可复现问题的Demo)
debug最后发现是rec模型最慢。 深入到源码层去看过，python和C++的没啥差异，唯独在运行推理的环节速度有明显差异。对应的C++是ocrrec.cpp的 this->predictor->Run() 这行。