🔄 PaddleOCR Model Convert

 

Introduction

• This repository is mainly to convert Inference Model in PaddleOCR into ONNX format.
• Input: url or local tar path of inference model
• Output: converted ONNX model
• If it is a recognition model, you need to provide the original txt path of the corresponding dictionary (Open the txt file in github, click the path after raw in the upper right corner, similar to this), used to write the dictionary into the ONNX model
• ☆ It needs to be used with the relevant reasoning code in RapidOCR
• If you encounter a model that cannot be successfully converted, you can check which steps are wrong one by one according to the ideas in the figure below.

Overall framework

flowchart TD

A([PaddleOCR inference model]) --paddle2onnx--> B([ONNX])
B --> C([Change Dynamic Input]) --> D([Rec: save the character dict to onnx])
D --> E([Save])

Loading

Installation

pip install paddleocr_convert

Usage

Warning

Only support the reasoning model in the download address in link, if it is a training model, Manual conversion to inference format is required.

The slim quantized model in PaddleOCR does not support conversion.

Using the command line

• Usage:

  $ paddleocr_convert -h
  usage: paddleocr_convert [-h] [-p MODEL_PATH] [-o SAVE_DIR]
                          [-txt_path TXT_PATH]
  
  optional arguments:
  -h, --help show this help message and exit
  -p MODEL_PATH, --model_path MODEL_PATH
                          The inference model url or local path of paddleocr.
                          e.g. https://paddleocr.bj.bcebos.com/PP-
                          OCRv3/chinese/ch_PP-OCRv3_det_infer.tar or
                          models/ch_PP-OCRv3_det_infer.tar
  -o SAVE_DIR, --save_dir SAVE_DIR
                          The directory of saving the model.
  -txt_path TXT_PATH, --txt_path TXT_PATH
                          The raw txt url or local txt path, if the model is
                          recognition model.

• Example:

  #online
  $ paddleocr_convert -p https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar \
                      -o models
  
  $ paddleocr_convert -p https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar\
                      -o models\
                      -txt_path https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.6/ppocr/utils/ppocr_keys_v1.txt
  
  # offline
  $ paddleocr_convert -p models/ch_PP-OCRv3_det_infer.tar\
                      -o models
  
  $ paddleocr_convert -p models/ch_PP-OCRv3_rec_infer.tar\
                      -o models\
                      -txt_path models/ppocr_keys_v1.txt

Script use

• online mode

  from paddleocr_convert import PaddleOCRModelConvert
  
  converter = PaddleOCRModelConvert()
  save_dir = 'models'
  url = 'https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar'
  txt_url = 'https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/release/2.6/ppocr/utils/ppocr_keys_v1.txt'
  
  converter(url, save_dir, txt_path=txt_url)

• offline mode

  from paddleocr_convert import PaddleOCRModelConvert
  
  converter = PaddleOCRModelConvert()
  save_dir = 'models'
  model_path = 'models/ch_PP-OCRv3_rec_infer.tar'
  txt_path = 'models/ppocr_keys_v1.txt'
  converter(model_path, save_dir, txt_path=txt_path)

Use the model

Assuming that the model needs to be recognized in Japanese, and it has been converted, the path is local/models/japan.onnx

1. Install rapidocr_onnxruntime library

   pip install rapidocr_onnxruntime

2. Script use

   from rapidocr_onnxruntime import RapidOCR
   
   model_path = 'local/models/japan.onnx'
   engine = RapidOCR(rec_model_path=model_path)
   
   img = '1.jpg'
   result, elapse = engine(img)

3. CLI use

   rapidocr_onnxruntime -img 1.jpg --rec_model_path local/models/japan.onnx

Changelog

Click to expand

• 2023-09-22 v0.0.17 update:
  • Improve the log when meets the error.
• 2023-07-27 v0.0.16 update:
  • Added the online conversion version of ModelScope.
  • Change python version from python 3.6 ~ 3.11.
• 2023-04-13 update:
  • Add online conversion program link
• 2023-03-05 v0.0.4~7 update:
  • Support transliteration of local models and dictionaries
  • Optimize internal logic and error feedback
• 2023-02-28 v0.0.3 update:
  • Added setting to automatically change to dynamic input for models that are not dynamic input
• 2023-02-27 v0.0.2 update:
  • Encapsulate the conversion model code into a package, which is convenient for self-help model conversion
• 2022-08-15 v0.0.1 update:
  • Write the dictionary of the recognition model into the meta in the onnx model for subsequent distribution.