Customized model for Thai language

[GoldenBigAnt]

I made Thai lang model. but paddleocr don't support the Thai lang yet. how can I use directly the trained thai model? please help me

[GreatV]

To use your trained Thai language model in PaddleOCR, even though Thai is not officially supported yet, you can follow these steps:

1. Update the YAML Configuration

• PaddleOCR uses YAML files to configure the training and inference process. Since you’ve already trained a Thai model, ensure that the YAML file you used during training is properly configured to include:

  • The Thai dataset path.
  • The Thai character set (character_dict_path), which should include all Thai characters you want to recognize.
  • Model architecture and hyperparameters suitable for Thai scripts.

• Example YAML changes:

  Global:
    character_dict_path: "path_to_your_thai_character_list.txt"

• If you’ve used a YAML configuration for English or another language, modify the character_dict_path to point to your Thai character list.

---

2. Ensure Thai Dataset Preprocessing

• If your Thai model was trained on a dataset, make sure the dataset was preprocessed correctly. PaddleOCR uses the tools/create_lmdb_dataset.py script to preprocess text datasets into LMDB format.
• For Thai, ensure the text annotations are encoded in Unicode to handle the complex script.

---

3. Inference with a Custom Model

• During inference, you need to specify the trained Thai model and its corresponding configuration YAML file. You can run the following command to perform inference:

  python tools/infer_rec.py --config_path="path_to_your_yaml_config.yaml" --image_dir="path_to_image" --rec_model_dir="path_to_trained_thai_model/"

• Replace path_to_your_yaml_config.yaml with your Thai YAML configuration file and path_to_trained_thai_model/ with the directory containing your trained Thai model.

---

4. Update the Dictionary for Thai

• PaddleOCR uses a character dictionary for recognition. Make sure your Thai character dictionary (character_dict_path) is correctly specified and matches the characters your model was trained on.
• Example Thai character list (save as thai_dict.txt):

  ก
  ข
  ฃ
  ค
  ...
  (all Thai characters)
  

---

5. Debugging Issues

If your model isn't working as expected or PaddleOCR defaults to a different language (like English), check the following:

• Ensure you are pointing to the correct .yaml file during inference.
• Validate that the trained weights (.pth or .pdparams file) are correctly linked in the YAML file.
• Ensure the character dictionary aligns with the model's training.

---

6. Multilingual Support

If you want PaddleOCR to support both Thai and other languages, you can create a combined multilingual character list that includes characters from all desired languages. Adjust the character_dict_path and retrain or fine-tune the model with this multilingual dataset.

Reference: There’s a related discussion here about fine-tuning a model for Indonesian. While it focuses on Indonesian, the process for Thai is similar:

• Fine-tune with a Thai-specific dataset.
• Adjust the YAML configuration for Thai.
• Use the proper dictionary for Thai characters.

---

By following these steps, you should be able to use your trained Thai language model in PaddleOCR. If you encounter specific errors, sharing the error messages can help diagnose the problem further.

---

Response generated by feifei-bot | chatgpt-4o-latest

[GoldenBigAnt]

thanks for your help. I followed above steps and made just best_accuracy model already.
how can I use the model in python?

[GoldenBigAnt]

[GoldenBigAnt]

please see, I can get this result from own trained model
but I can't know how can I use the model for other images. please help me

[jingsongliujing]

You can see this document：https://paddlepaddle.github.io/PaddleOCR/latest/en/ppocr/infer_deploy/python_infer.html#text-detection-model-inference

[GoldenBigAnt]

Happy new year.
I checked above link, but I can get good result only one line jpg file. but I have to ocr for mutiline jpg files.

like this file.
I used "python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/korean/1.jpg" --rec_model_dir="./your inference model" --rec_char_dict_path="ppocr/utils/dict/korean_dict.txt" --vis_font_path="doc/fonts/korean.ttf"" this command.
but it can get only one line.
please help me asap.