PaddleOCR Training Model Reference

This article is about training with NVidia GPUs. CPU training is for reference only; for any differences, please refer to the official website.

Official Links:

• https://github.com/PaddlePaddle/PaddleOCR
• https://gitee.com/paddlepaddle/PaddleOCR

Prerequisites

• Python 3.9 (3.10): Initial tests had ongoing issues, so the author switched to 3.9. If needed, verify with 3.10 on your own: https://www.python.org/

• Python: This is the keyword to execute scripts. Environment variables need to be configured. Many of the following components also require environment variable configuration; refer to online resources for details.

• pip: The author is not a Python developer; this is understood as an installation plugin. It can be used to install third‑party libraries. If pip3 doesn't work, try pip instead for unknown reasons.

• pip network issues: You can add the parameter -i when using pip, e.g., https://pypi.tuna.tsinghua.edu.cn/simple

For example: pip install -i https://pypi.tuna.tsinghua.edu.cn/simple pyspider will install the pyspider library from the Tsinghua mirror.

• CUDA

https://developer.nvidia.com/cuda-downloads

The author installed version 10.2.

• CUDNN

https://developer.nvidia.com/cudnn

After downloading, extract and copy the files to the CUDA directory.

• PaddleOCR

https://github.com/PaddlePaddle/PaddleOCR

Clone the project locally.

• cd PaddleOCR

pip3 install -r requirements.txt

Install the required Python libraries for OCR.

• PPOCRLabel

This is an annotation tool for creating training data – not mandatory, but very convenient.

cd ./PPOCRLabel # Switch to the PPOCRLabel folder
pip install pyqt5 # Install Qt5 runtime environment
pip3 install -r requirements.txt
python PPOCRLabel.py --lang ch # Launch the tool; if nothing happens, the environment is incomplete.

• ch_ppocr_mobile_v2.0_rec

Pre‑trained model (other models can be found at: models_list.md)

ch_ppocr_mobile_v2.0_rec_pre.tar

• Training parameter documentation

config.md

Local config file path: PaddleOCR-release-2.4\configs\rec\ch_ppocr_v2.0\rec_chinese_lite_train_v2.0.yml

• Values to modify:

epoch_num: 1000 # Number of epochs
data_dir: ./train_data/ # Training data directory
label_file_list: ["./train_data/train_list.txt"] # Label file for training data
batch_size_per_card: 128 # Batch size per GPU card (reduce if it fails to start)

• Explanation of training directories

PaddleOCR-release-2.4\train_data
PaddleOCR-release-2.4\train_data\crop_img # Place cropped images generated by the annotation tool here
PaddleOCR-release-2.4\train_data\train_list.txt # Training label file
PaddleOCR-release-2.4\train_data\val_list.txt   # Validation label file (currently the author uses the same content as the training file); content example below

PaddleOCR-release-2.4\pretrain_models  # Place pre‑trained models downloaded from the official site here
PaddleOCR-release-2.4\output # Training output directory
PaddleOCR-release-2.4\output\inference # Final exported model

• Training scripts

// Train the model
python tools/train.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.pretrained_model=./pretrain_models/best_accuracy

// Export the model
python tools/export_model.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.checkpoints=output/rec_chinese_lite_v2.0/latest Global.save_inference_dir=output/inference

// Predict using the trained model (folder of images)
python tools/infer_rec.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.checkpoints=output/rec_chinese_lite_v2.0/latest Global.load_static_weights=false Global.infer_img=trainTest/

// Predict using the trained model (single image)
python tools/infer_rec.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.checkpoints=output/rec_chinese_lite_v2.0/latest Global.load_static_weights=false Global.infer_img=trainTest/1000.jpg

// Predict using the exported model
python tools/infer/predict_rec.py --image_dir="./trainTest/" --det_model_dir="./ch_PP-OCRv2_det_infer/"  --rec_model_dir="./output/inference/" --cls_model_dir="./ch_ppocr_mobile_v2.0_cls_infer/"

Author: Dream.Machine

Website: www.dmskin.com