Distilbert-punctuator

Introduction

Distilbert-punctuator is a python package provides a bert-based punctuator (fine-tuned model of pretrained huggingface DistilBertForTokenClassification) with following three components:

• data process: funcs for processing user's data to prepare for training. If user perfer to fine-tune the model with his/her own data.
• training: training pipeline and evaluation. User can fine-tune his/her own punctuator with the pipeline
• inference: easy-to-use interface for user to use trained punctuator.
• If user doesn't want to train a punctuator himself/herself, two pre-fined-tuned model from huggingface model hub
  • Qishuai/distilbert_punctuator_en 📎 Model details
  • Qishuai/distilbert_punctuator_zh 📎 Model details

• model examples in huggingface web page.

  • English model

    

  • Simplified Chinese model

    

Installation

• Installing the package from pypi: pip install distilbert-punctuator for directly usage of punctuator.
• Installing the package with option to do data processing pip install distilbert-punctuator[data_process].
• Installing the package with option to train and validate your own model pip install distilbert-punctuator[training]
• For development and contribution
  • clone the repo
  • make install

Data Process

Component for pre-processing the training data. To use this component, please install as pip install distilbert-punctuator[data_process]

The package is providing a simple pipeline for you to generate NER format training data.

Example

examples/data_sample.py

Train

Component for providing a training pipeline for fine-tuning a pretrained DistilBertForTokenClassification model from huggingface. The latest version has the implementation of R-Drop enhanced training. R-Drop github repo Paper of R-Drop

Example

examples/english_train_sample.py

Training_arguments:

Arguments required for the training pipeline.

• basic arguments

  • training_corpus(List[List[str]]): list of sequences for training, longest sequence should be no longer than pretrained LM # noqa: E501
  • validation_corpus(List[List[str]]): list of sequences for validation, longest sequence should be no longer than pretrained LM # noqa: E501
  • training_tags(List[List[int]]): tags(int) for training
  • validation_tags(List[List[int]]): tags(int) for validation
  • model_name_or_path(str): name or path of pre-trained model
  • tokenizer_name(str): name of pretrained tokenizer

• training arguments

  • epoch(int): number of epoch
  • batch_size(int): batch size
  • model_storage_dir(str): fine-tuned model storage path
  • label2id(Dict): the tags label and id mapping
  • early_stop_count(int): after how many epochs to early stop training if valid loss not become smaller. default 3 # noqa: E501
  • gpu_device(int): specific gpu card index, default is the CUDA_VISIBLE_DEVICES from environ
  • warm_up_steps(int): warm up steps.
  • r_drop(bool): whether to train with r-drop
  • r_alpha(int): alpha value for kl divengence in the loss, default is 0
  • plot_steps(int): record training status to tensorboard among how many steps
  • tensorboard_log_dir(Optional[str]): the tensorboard logs output directory, default is "runs"

• model arguments

  • addtional_model_config(Optional[Dict]): additional configuration for model

You can also train your own NER models with the trainer provided in this repo. The example can be found in notebooks/R-drop NER.ipynb

Evaluation

Validation of fine-tuned model

Example

examples/train_sample.py

Validation_arguments:

• evaluation_corpus(List[List[str]]): list of sequences for evaluation, longest sequence should be no longer than pretrained LM's max_position_embedding(512)
• evaluation_tags(List[List[int]]): tags(int) for evaluation (the GT)
• model_name_or_path(str): name or path of fine-tuned model
• tokenizer_name(str): name of tokenizer
• batch_size(int): batch size
• label2id(Optional[Dict]): label2id. Default one is from model config. Pass in this argument if your model doesn't have a label2id inside config
• gpu_device(int): specific gpu card index, default is the CUDA_VISIBLE_DEVICES from environ

Inference

Component for providing an inference interface for user to use punctuator.

Architecture

 +----------------------+              (child process)
 |   user application   |             +-------------------+
 +                      + <---------->| punctuator server |
 |   +inference object  |             +-------------------+
 +----------------------+

The punctuator will be deployed in a child process which communicates with main process through pipe connection. Therefore user can initialize an inference object and call its punctuation function when needed. The punctuator will never block the main process unless doing punctuation. There is a graceful shutdown methodology for the punctuator, hence user dosen't need to worry about the shutting-down.

Example

examples/inference_sample.py

Inference_arguments

Arguments required for the inference pipeline.

• model_name_or_path(str): name or path of pre-trained model
• tokenizer_name(str): name of pretrained tokenizer

• tag2punctuator(Dict[str, tuple]): tag to punctuator mapping. dbpunctuator.utils provides two default mappings for English and Chinese

  NORMAL_TOKEN_TAG = "O"
  DEFAULT_ENGLISH_TAG_PUNCTUATOR_MAP = {
     NORMAL_TOKEN_TAG: ("", False),
     "COMMA": (",", False),
     "PERIOD": (".", True),
     "QUESTIONMARK": ("?", True),
     "EXLAMATIONMARK": ("!", True),
  }
  
  DEFAULT_CHINESE_TAG_PUNCTUATOR_MAP = {
     NORMAL_TOKEN_TAG: ("", False),
     "C_COMMA": ("，", False),
     "C_PERIOD": ("。", True),
     "C_QUESTIONMARK": ("? ", True),
     "C_EXLAMATIONMARK": ("! ", True),
     "C_DUNHAO": ("、", False),
  }

  for own fine-tuned model with different tags, pass in your own mapping

• tag2id_storage_path(Optional[str]): tag2id storage path. Default one is from model config. Pass in this argument if your model doesn't have a tag2id inside config