|

Optocycle PaddleOCR Fork

This fork includes configurations for fine-tuning PaddleOCR on license plate data, along with an extension for MLflow logging.

Fine-tuning a model

These instructions are based on the guide provided here.

1. Open the cloned repository and install the environment with poetry:

cd PaddleOCR/
poetry install

2. Download a pretrained model.

# Download the pre-trained model for en_PP-OCRv3 (the weights can also be found in the models.dev bucket on MinIO)
wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_train.tar
# Decompress model parameters
cd pretrain_models
tar -xf en_PP-OCRv3_rec_train.tar && rm -rf en_PP-OCRv3_rec_train.tar
cd ..

3. Ensure your dataset is formatted for PaddleOCR and move it to the datasets/ directory: mv /path/to/paddleocr_dataset ./datasets
4. Fine-tune the model:

poetry run python3 tools/train.py -c configs/rec/PP-OCRv3/slplates_finetuning.yml -o Global.pretrained_model=pretrain_models/en_PP-OCRv3_rec_train/best_accuracy

5. Export the fine-tuned model to an inference model and specify a custom destination directory (e.g., ./slplates_inference_model):

poetry run python3 tools/export_model.py -c configs/rec/PP-OCRv3/slplates_finetuning.yml -o Global.pretrained_model=output/normal_finetuned_slplates_paddleocr/best_model/model  Global.save_inference_dir=./slplates_inference_model

Enabling MLflow

To enable MLflow logging during fine-tuning, add the following to your configuration file:

mlflow:
  mlflow_log_every_n_iter: <n_iter>  # Log training metrics every <n_iter> iterations (evaluation metrics are always logged)
  name: <name_of_the_run>  # Name of the MLflow run (e.g., "lp_finetuned")
  log_only_best_model: <true/false>  # Whether to log only the best model based on its accuracy
  pbtxt_configs:  # Configuration for generating .pbtxt files to log models in Triton inference server format
    max_batch_size: <n_samples>  # Maximum batch size for inference
    dynamic_batching: <true/false>  # Enable or disable dynamic batching for Triton

Additionally, create a .env file by copying .env.example and filling in the correct values."

📄 License

This project is released under Apache License Version 2.0.

Changes

In addition to the modifications described in the header of the original files, the following changes have been made:

• The setup.py and MANIFEST.in files have been removed as the project has been switched to using Poetry for dependency management and packaging.
• Heavily revised README.md to reflect changes and usage for license plate fine-tuning.