LEKCut 1.0.0b1

LEKCut (เล็ก คัด) is a Thai tokenization library that ports the deep learning model to the onnx model.

LEKCut

LEKCut (เล็ก คัด) is a Thai tokenization library that ports the deep learning model to the onnx model.

Install

pip install lekcut

How to use

from lekcut import word_tokenize

# DeepCut model (default)
word_tokenize("ทดสอบการตัดคำ")
# output: ['ทดสอบ', 'การ', 'ตัด', 'คำ']

# AttaCut syllable + character model
word_tokenize("ทดสอบการตัดคำ", model="attacut-sc")
# output: ['ทดสอบ', 'การ', 'ตัด', 'คำ']

# AttaCut character-only model
word_tokenize("ทดสอบการตัดคำ", model="attacut-c")
# output: ['ทดสอบ', 'การ', 'ตัด', 'คำ']

# OSKut model
word_tokenize("เบียร์ยูไม่อร่อย", model="oskut")
# output: ['เบียร์', 'ยู', 'ไม่', 'อ', 'ร่อย']

# OSKut with a specific engine
word_tokenize("เบียร์ยูไม่อร่อย", model="oskut", engine="tnhc")
# output: ['เบียร์', 'ยู', 'ไม่', 'อร่อย']

API

word_tokenize(
    text: str,
    model: str = "deepcut",
    path: str = "default",
    providers: List[str] = None,
    engine: str = "ws",
    k: int = 1,
) -> List[str]

Parameters:

• text: Text to tokenize
• model: Model to use. Options: "deepcut" (default), "attacut-sc", "attacut-c", "oskut"
• path: Path to custom model file (default: "default", applies to deepcut and attacut-* models)
• providers: List of ONNX Runtime execution providers (default: None, which uses default CPU provider)
• engine: OSKut engine variant (applies to "oskut" model only). Options: "ws" (default), "ws-augment-60p", "tnhc", "scads", "tl-deepcut-ws", "tl-deepcut-tnhc", "deepcut"
• k: Percentage of characters to refine for OSKut (applies to "oskut" model only). The special default value of 1 is a sentinel that lets OSKut automatically select an appropriate percentage based on the engine. Pass any integer from 2 to 100 to override.

GPU Support

LEKCut supports GPU acceleration through ONNX Runtime execution providers. To use GPU acceleration:

1. Install ONNX Runtime with GPU support:

   pip install onnxruntime-gpu
   

2. Use the providers parameter to specify GPU execution:

   from lekcut import word_tokenize
   
   # Use CUDA GPU
   result = word_tokenize("ทดสอบการตัดคำ", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
   
   # Use TensorRT (if available)
   result = word_tokenize("ทดสอบการตัดคำ", providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'])
   

Available Execution Providers:

• CPUExecutionProvider - Default CPU execution
• CUDAExecutionProvider - NVIDIA CUDA GPU acceleration
• TensorrtExecutionProvider - NVIDIA TensorRT optimization
• DmlExecutionProvider - DirectML for Windows GPU
• And more (see ONNX Runtime documentation)

Note: The providers are tried in order, and the first available one will be used. Always include CPUExecutionProvider as a fallback.

Model

• deepcut - We ported deepcut model from tensorflow.keras to ONNX model. The model and code come from Deepcut's Github. The model is here.
• attacut-sc - We ported the AttaCut syllable + character model from PyTorch to ONNX. The model and code come from AttaCut's Github. Requires the ssg package for syllable tokenization.
• attacut-c - We ported the AttaCut character-only model from PyTorch to ONNX. The model and code come from AttaCut's Github.
• oskut - We ported the OSKut (Out-of-domain Stacked Cut) stacked ensemble models from TensorFlow/Keras to ONNX. The model and code come from OSKut's Github. Requires the pyahocorasick package. Supports multiple engines: ws (default), ws-augment-60p, tnhc, scads, tl-deepcut-ws, tl-deepcut-tnhc, deepcut.

Load custom model

If you have trained your custom model from deepcut or other that LEKCut support, You can load the custom model by path in word_tokenize after porting your model.

• How to train custom model with your dataset by deepcut - Notebook (Needs to update deepcut/train.py before train model)

How to porting model?

See notebooks/