AttributeError when loading OCR model - kiri-ocr version incompatibility

[mrrtmob]

AttributeError: 'dict' object has no attribute 'UNK_TOKEN'

Error Details

Full Traceback:

Traceback (most recent call last):
  File "/home/user/app/app.py", line 49, in process_image
    ocr = load_ocr(device="cuda")
  File "/home/user/app/app.py", line 20, in load_ocr
    return OCR(
  File "/usr/local/lib/python3.10/site-packages/kiri_ocr/core.py", line 77, in __init__
    self._load_model(model_path, charset_path)
  File "/usr/local/lib/python3.10/site-packages/kiri_ocr/core.py", line 162, in _load_model
    self._load_transformer_model(checkpoint, model_path)
  File "/usr/local/lib/python3.10/site-packages/kiri_ocr/core.py", line 222, in _load_transformer_model
    self.transformer_tok = CharTokenizer(vocab_path, self.transformer_cfg)
  File "/usr/local/lib/python3.10/site-packages/kiri_ocr/model_transformer.py", line 76, in __init__
    if cfg.UNK_TOKEN not in vocab_raw:
AttributeError: 'dict' object has no attribute 'UNK_TOKEN'

Root Cause

The updated model on Hugging Face uses a newer configuration format that is incompatible with older versions of kiri-ocr. The CharTokenizer class expects cfg to have an UNK_TOKEN attribute, but receives a dictionary instead.

Solution

Upgrade kiri-ocr to version >= 0.2.0 which supports the updated model configuration format.

Steps to Reproduce

1. Use the updated model from Hugging Face
2. Attempt to load OCR with load_ocr(device="cuda")
3. Error occurs during CharTokenizer initialization

Fix

Update requirements.txt or your dependency file:

- kiri-ocr==0.1.x
+ kiri-ocr>=0.2.0

Then reinstall dependencies:

pip install --upgrade kiri-ocr

Environment

• Python: 3.10
• Current kiri-ocr version: < 0.2.0
• Required kiri-ocr version: >= 0.2.0

Additional Notes

This breaking change was introduced due to model updates on Hugging Face. All users should upgrade to ensure compatibility with the latest models.

[mrrtmob]

⚠️ UPDATE: Additional Step Required

If you still get the error after upgrading kiri-ocr, you need to clear your local Hugging Face model cache:

Linux/Mac:

rm -rf ~/.cache/huggingface/hub

Windows:

rmdir /s /q %USERPROFILE%\.cache\huggingface\hub

The old cached model files are incompatible with the new version and must be removed to download the updated model.