June 10, 2019

598 words 3 mins read

asyml/texar

asyml/texar

Toolkit for Machine Learning, Natural Language Processing, and Text Generation, in TensorFlow

repo name asyml/texar
repo link https://github.com/asyml/texar
homepage https://asyml.io
language Python
size (curr.) 13892 kB
stars (curr.) 1837
created 2017-07-22
license Apache License 2.0

pypi Build Status codecov Documentation Status License

Texar is a toolkit aiming to support a broad set of machine learning, especially natural language processing and text generation tasks. Texar provides a library of easy-to-use ML modules and functionalities for composing whatever models and algorithms. The tool is designed for both researchers and practitioners for fast prototyping and experimentation.

Key Features

  • Two Versions, (Mostly) Same Interfaces. Texar-TensorFlow (this repo) and Texar-PyTorch have mostly the same interfaces. Both further combine the best design of TF and PyTorch:
    • Interfaces and variable sharing in PyTorch convention
    • Excellent factorization and rich functionalities in TF convention.
  • Rich Pre-trained Models, Rich Usage with Uniform Interfaces. BERT, GPT2, XLNet, etc, for encoding, classification, generation, and composing complex models with other Texar components!
  • Fully Customizable at multiple abstraction level – both novice-friendly and expert-friendly.
    • Free to plug in whatever external modules, since Texar is fully compatible with the native TF/PyTorch APIs.
  • Versatile to support broad tasks, models, algorithms, data processing, evaluation, etc.
    • encoder(s) to decoder(s), sequential- and self-attentions, memory, hierarchical models, classifiers…
    • maximum likelihood learning, reinforcement learning, adversarial learning, probabilistic modeling, …
  • Modularized for maximal re-use and clean APIs, based on principled decomposition of Learning-Inference-Model Architecture.
  • Distributed model training with multiple GPUs.
  • Clean, detailed documentation and rich examples.

Library API Example

Builds an encoder-decoder model, with maximum likelihood learning:

import texar.tf as tx

# Data 
data = tx.data.PairedTextData(hparams=hparams_data) # a dict of hyperparameters 
iterator = tx.data.DataIterator(data)
batch = iterator.get_next()                         # get a data mini-batch

# Model architecture
embedder = tx.modules.WordEmbedder(data.target_vocab.size, hparams=hparams_emb)
encoder = tx.modules.TransformerEncoder(hparams=hparams_enc)
outputs_enc = encoder(inputs=embedder(batch['source_text_ids']),  # call as a function
                      sequence_length=batch['source_length'])
                      
decoder = tx.modules.TransformerDecoder(
    output_layer=tf.transpose(embedder.embedding) # tie input embedding w/ output layer
    hparams=hparams_decoder)
outputs, _, _ = decoder(memory=output_enc, 
                        memory_sequence_length=batch['source_length'],
                        inputs=embedder(batch['target_text_ids']),
                        sequence_length=batch['target_length']-1,
                        decoding_strategy='greedy_train')    # teacher-forcing decoding
                        
# Loss for maximum likelihood learning
loss = tx.losses.sequence_sparse_softmax_cross_entropy(
    labels=batch['target_text_ids'][:, 1:],
    logits=outputs.logits,
    sequence_length=batch['target_length']-1)  # automatic sequence masks

# Beam search decoding
outputs_bs, _, _ = tx.modules.beam_search_decode(
    decoder,
    embedding=embedder,
    start_tokens=[data.target_vocab.bos_token_id]*num_samples,
    end_token=data.target_vocab.eos_token_id)

The same model, but with adversarial learning:

helper = tx.modules.GumbelSoftmaxTraingHelper( # Gumbel-softmax decoding
    start_tokens=[BOS]*batch_size, end_token=EOS, embedding=embedder)
outputs, _ = decoder(helper=helper)            # automatic re-use of the decoder variables

discriminator = tx.modules.BertClassifier(hparams=hparams_bert)        # pre-trained model

G_loss, D_loss = tx.losses.binary_adversarial_losses(
    real_data=data['target_text_ids'][:, 1:],
    fake_data=outputs.sample_id,
    discriminator_fn=discriminator)

The same model, but with RL policy gradient learning:

agent = tx.agents.SeqPGAgent(samples=outputs.sample_id,
                             logits=outputs.logits,
                             sequence_length=batch['target_length']-1,
                             hparams=config_model.agent)

Many more examples are available here

Installation

(Note: Texar>0.2.3 requires Python 3.6 or 3.7. To use with older Python versions, please use Texar<=0.2.3)

Texar requires:

After tensorflow and tensorflow_probability are installed, install Texar from PyPI:

pip install texar

To use cutting-edge features or develop locally, install from source:

git clone https://github.com/asyml/texar.git
cd texar
pip install .

Getting Started

Reference

If you use Texar, please cite the tech report with the following BibTex entry:

Texar: A Modularized, Versatile, and Extensible Toolkit for Text Generation
Zhiting Hu, Haoran Shi, Bowen Tan, Wentao Wang, Zichao Yang, Tiancheng Zhao, Junxian He, Lianhui Qin, Di Wang, Xuezhe Ma, Zhengzhong Liu, Xiaodan Liang, Wanrong Zhu, Devendra Sachan and Eric Xing
ACL 2019

@inproceedings{hu2019texar,
  title={Texar: A Modularized, Versatile, and Extensible Toolkit for Text Generation},
  author={Hu, Zhiting and Shi, Haoran and Tan, Bowen and Wang, Wentao and Yang, Zichao and Zhao, Tiancheng and He, Junxian and Qin, Lianhui and Wang, Di and others},
  booktitle={ACL 2019, System Demonstrations},
  year={2019}
}

License

Apache License 2.0

comments powered by Disqus