/langml

A Keras-based and TensorFlow-backend NLP Models Toolkit.

Primary LanguagePythonMIT LicenseMIT

LangML (Language ModeL) is a Keras-based and TensorFlow-backend language model toolkit, which provides mainstream pre-trained language models, e.g., BERT/RoBERTa/ALBERT, and their downstream application models.

pypi

Outline

Features

  • Common and widely-used Keras layers: CRF, Transformer, Attentions: Additive, ScaledDot, MultiHead, GatedAttentionUnit, and so on.
  • Pretrained Language Models: BERT, RoBERTa, ALBERT. Providing friendly designed interfaces and easy to implement downstream singleton, shared/unshared two-tower or multi-tower models.
  • Tokenizers: WPTokenizer (wordpiece), SPTokenizer (sentencepiece)
  • Baseline models: Text Classification, Named Entity Recognition, Contrastive Learning. It's no need to write any code, and just need to preprocess the data into a specific format and use the "langml-cli" to train various baseline models.
  • Prompt-Based Tuning: PTuning

Installation

You can install or upgrade langml/langml-cli via the following command:

pip install -U langml

Quick Start

Specify the Keras variant

  1. Use pure Keras (default setting)
export TF_KERAS=0
  1. Use TensorFlow Keras
export TF_KERAS=1

Load pretrained language models

from langml import WPTokenizer, SPTokenizer
from langml import load_bert, load_albert

# load bert / roberta plm
bert_model, bert = load_bert(config_path, checkpoint_path)
# load albert plm
albert_model, albert = load_albert(config_path, checkpoint_path)
# load wordpiece tokenizer
wp_tokenizer = WPTokenizer(vocab_path, lowercase)
# load sentencepiece tokenizer
sp_tokenizer = SPTokenizer(vocab_path, lowercase)

Finetune a model

from langml import keras, L
from langml import load_bert

config_path = '/path/to/bert_config.json'
ckpt_path = '/path/to/bert_model.ckpt'
vocab_path = '/path/to/vocab.txt'

bert_model, bert_instance = load_bert(config_path, ckpt_path)
# get CLS representation
cls_output = L.Lambda(lambda x: x[:, 0])(bert_model.output)
output = L.Dense(2, activation='softmax',
                 kernel_intializer=bert_instance.initializer)(cls_output)
train_model = keras.Model(bert_model.input, cls_output)
train_model.summary()
train_model.compile(loss='categorical_crossentropy', optimizer=keras.optimizer.Adam(1e-5))

Use langml-cli to train baseline models

  1. Text Classification
$ langml-cli baseline clf --help
Usage: langml baseline clf [OPTIONS] COMMAND [ARGS]...

  classification command line tools

Options:
  --help  Show this message and exit.

Commands:
  bert
  bilstm
  textcnn
  1. Named Entity Recognition
$ langml-cli baseline ner --help
Usage: langml baseline ner [OPTIONS] COMMAND [ARGS]...

  ner command line tools

Options:
  --help  Show this message and exit.

Commands:
  bert-crf
  lstm-crf
  1. Contrastive Learning
$ langml-cli baseline contrastive --help
Usage: langml baseline contrastive [OPTIONS] COMMAND [ARGS]...

  contrastive learning command line tools

Options:
  --help  Show this message and exit.

Commands:
  simcse
  1. Text Matching
$ langml-cli baseline matching --help
Usage: langml baseline matching [OPTIONS] COMMAND [ARGS]...

  text matching command line tools

Options:
  --help  Show this message and exit.

Commands:
  sbert

Documentation

Please visit the langml.readthedocs.io to check the latest documentation.

Reference

The implementation of pretrained language model is inspired by CyberZHG/keras-bert and bojone/bert4keras.