Faced training problem using paddleOCR for new Language(Bengali or Bangla)

[Hasibur013]

I create a custom dataset using number plate image, text are bangla character and bangla number mixed, then I did rec_gt_train.txt, rec_gt_test.txt and word_dict_bn.txt file perfectly, then change yml file that need. After that I train this but got acc=0, tried different way still now acc=0. Did I any mistake? or any suggestions, please share your thought. If anyone did this for any new language please share your journey.

[GreatV]

You’re experiencing acc=0 when training PaddleOCR for Bangla number plates, despite preparing the dataset and config files. This is a common issue when training custom OCR for a new language. Let’s go through common mistakes and suggestions specific to your scenario:

1. Check Your Dictionary File

• Make sure your word_dict_bn.txt covers every Bangla character and number in your dataset.
• Each character or symbol (including spaces or special characters) present in your ground truth files must be included—one per line.

2. Ground Truth File Format

• Your rec_gt_train.txt and rec_gt_test.txt should follow PaddleOCR’s required format:

  path/to/image_1.jpg    corresponding_text_১
  path/to/image_2.jpg    কিছু_আরও_লেখা
  

• Ensure there are no tab or space issues; the separator is typically a tab or special separator.
• Check for broken text encoding. Bangla text must be UTF-8 encoded throughout.

3. YAML Configuration

• Verify you set the correct character_dict_path in the yaml config, pointing to your dictionary.
• Confirm max_text_length can accommodate the longest string in your dataset.

4. Dataset Images

• Make sure your images are clear and labeled correctly. Corrupted or empty labels can result in acc=0.
• Try training with a tiny subset (5–10 images) and validate it works, then scale up.

5. Start with a Pretrained Model

• Consider initializing training with a model pretrained on a similar script (like Hindi or Devanagari, or even default English), then fine-tune.

6. Increase Dataset Size

• If your dataset is too small or lacks variation, the model may not learn; try increasing dataset size and diversity.

7. PaddleOCR Issues & Discussions

Several users have faced similar issues when training on Bangla:

• Bangla OCR Discussion - PaddleOCR GitHub: Multiple users successfully prepared Bangla datasets and offered tips on dictionary setup and config tweaks. Reading this may offer extra insight.

8. Logging & Debugging

• Run initial epochs with print_batch_step: 1 in your config to observe what’s being fed into the model.
• Look for errors, warnings, or incorrectly formatted batches.

---

Summary Checklist:

• All unique characters in word_dict_bn.txt (check for missing or extra).
• Correct, UTF-8 encoded ground truth files formatted as required.
• YAML config paths and settings pointing to your files.
• Dataset images are readable and correctly labeled.
• Try starting with a small dataset and pretrained weights.

If you still see acc=0, consider sharing a snippet of your ground truth and dictionary file in the Bangla OCR Discussion for targeted help from other users who trained with new languages.

Good luck with your Bangla OCR project!

---

Response generated by 🤖 feifei-bot | gpt-4.1

⚠️ This is an AI-generated response. Please verify any critical information.

[Hasibur013]

I followed all but still 0 acc

[cuicheng01]

Feel free to submit your request directly in this issue:
👉 PaddleOCR Multilingual Support Request #15617