Neuro-comma

This library was developed with the idea to help us to create punctuation restoration models to memorize trained parameters, data, training visualization, etc. The Library doesn't use any high-level frameworks, such as PyTorch-lightning or Keras, to reduce the level entry threshold.

Feel free to fork this repo and edit model or dataset classes for your purposes.

Prerequirements

• Python 3.9 for training
• Docker for production

Why development environment is Python 3.9 if production environment in Dockerfile is 3.8?

• Our team always uses the latest version and features of Python. We started with Python 3.9, but realized, that there is no FastAPI image for Python 3.9. There is several PRs in image repositories, but no response from maintainers. So we decided to change code which we use in production to work with the 3.8 version of Python. In some functions we have 3.9 code, but we still use them, these functions are needed only for development purposes.

Installation

• Option 1:

  pip install -U pip wheel setuptools
  

  pip install -r requirements.txt
  

• Option 2:

  sh scripts/installation.sh
  

Python module usage

• Choose model from releases section
• Checkout to release tag!
• Download and unzip model
• Check quantization.ipynb notebook for example how use predictor.

Production usage

• Choose model from releases section
• Checkout to release tag!
• Download and unzip model
• Run docker-compose

  docker-compose up -d
  

• Stop container

  docker-compose down
  

Model training

Model training from scratch:

python src/train.py \
    --model-name repunct-model \
    --pretrained-model DeepPavlov/rubert-base-cased-sentence \
    --targets O COMMA PERIOD \
    --train-data data/repunct/train \
    --val-data data/repunct/test \
    --test-data data/repunct/test \
    --store-best-weights \
    --epoch 7 \
    --batch-size 4 \
    --augment-rate 0.15 \
    --labml \
    --seed 1 \
    --cuda 

Fine-tuning already trained model. Add --fine-tune argument, this will load params from repunct-model and apply them to training function. This will create new subdirectory with {model-name}_ft name in models/ directory. Source model will be untouched.

python src/train.py \
    --model-name repunct-model \
    --fine-tune \
    --targets O COMMA PERIOD \
    --train-data data/repunct/train \
    --val-data data/repunct/test \
    --test-data data/repunct/test \
    --store-best-weights \
    --epoch 3 \
    --batch-size 4 \
    --labml \
    --seed 1 \
    --cuda 

In some cases you want to resume training (computer crashed, light blinked, etc.). This will resume training from last model checkpoint (saved weight). Just add --resume argument.

python src/train.py \
    --model-name repunct-model \
    --resume \
    --pretrained-model DeepPavlov/rubert-base-cased-sentence \
    --targets O COMMA PERIOD \
    --train-data data/repunct/train \
    --val-data data/repunct/test \
    --test-data data/repunct/test \
    --store-best-weights \
    --epoch 4 \
    --batch-size 4 \
    --augment-rate 0.15 \
    --labml \
    --seed 1 \
    --cuda 

More examples here

How it works

Before inserting raw text into model it should be tokenized. Library handle it with BaseDataset.parse_tokens

Model architecture is pretty easy and straight forward:

• BERT layer - DeepPavlov/rubert-base-cased-sentence language model
• Bi-LSTM layer - to reduce demsions
• Linear layer - final layer to predict what symbol should go after token

Credits

Our article on habr.ru

This repository contains code (which was edited for production purposes) from xashru/punctuation-restoration.

Special thanks to @akvarats