Skip to content

Neural machine translation and sequence learning using TensorFlow

License

Notifications You must be signed in to change notification settings

OpenNMT/OpenNMT-tf

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CI codecov PyPI version Documentation Gitter Forum

OpenNMT-tf

OpenNMT-tf is a general purpose sequence learning toolkit using TensorFlow 2. While neural machine translation is the main target task, it has been designed to more generally support:

  • sequence to sequence mapping
  • sequence tagging
  • sequence classification
  • language modeling

The project is production-oriented and comes with backward compatibility guarantees.

Key features

Modular model architecture

Models are described with code to allow training custom architectures and overriding default behavior. For example, the following instance defines a sequence to sequence model with 2 concatenated input features, a self-attentional encoder, and an attentional RNN decoder sharing its input and output embeddings:

opennmt.models.SequenceToSequence(
    source_inputter=opennmt.inputters.ParallelInputter(
        [
            opennmt.inputters.WordEmbedder(embedding_size=256),
            opennmt.inputters.WordEmbedder(embedding_size=256),
        ],
        reducer=opennmt.layers.ConcatReducer(axis=-1),
    ),
    target_inputter=opennmt.inputters.WordEmbedder(embedding_size=512),
    encoder=opennmt.encoders.SelfAttentionEncoder(num_layers=6),
    decoder=opennmt.decoders.AttentionalRNNDecoder(
        num_layers=4,
        num_units=512,
        attention_mechanism_class=tfa.seq2seq.LuongAttention,
    ),
    share_embeddings=opennmt.models.EmbeddingsSharingLevel.TARGET,
)

The opennmt package exposes other building blocks that can be used to design:

Standard models such as the Transformer are defined in a model catalog and can be used without additional configuration.

Find more information about model configuration in the documentation.

Full TensorFlow 2 integration

OpenNMT-tf is fully integrated in the TensorFlow 2 ecosystem:

Compatibility with CTranslate2

CTranslate2 is an optimized inference engine for OpenNMT models featuring fast CPU and GPU execution, model quantization, parallel translations, dynamic memory usage, interactive decoding, and more! OpenNMT-tf can automatically export models to be used in CTranslate2.

Dynamic data pipeline

OpenNMT-tf does not require to compile the data before the training. Instead, it can directly read text files and preprocess the data when needed by the training. This allows on-the-fly tokenization and data augmentation by injecting random noise.

Model fine-tuning

OpenNMT-tf supports model fine-tuning workflows:

  • Model weights can be transferred to new word vocabularies, e.g. to inject domain terminology before fine-tuning on in-domain data
  • Contrastive learning to reduce word omission errors

Source-target alignment

Sequence to sequence models can be trained with guided alignment and alignment information are returned as part of the translation API.


OpenNMT-tf also implements most of the techniques commonly used to train and evaluate sequence models, such as:

  • automatic evaluation during the training
  • multiple decoding strategy: greedy search, beam search, random sampling
  • N-best rescoring
  • gradient accumulation
  • scheduled sampling
  • checkpoint averaging
  • ... and more!

See the documentation to learn how to use these features.

Usage

OpenNMT-tf requires:

  • Python 3.7 or above
  • TensorFlow 2.6, 2.7, 2.8, 2.9, 2.10, 2.11, 2.12, or 2.13

We recommend installing it with pip:

pip install --upgrade pip
pip install OpenNMT-tf

See the documentation for more information.

Command line

OpenNMT-tf comes with several command line utilities to prepare data, train, and evaluate models.

For all tasks involving a model execution, OpenNMT-tf uses a unique entrypoint: onmt-main. A typical OpenNMT-tf run consists of 3 elements:

  • the model type
  • the parameters described in a YAML file
  • the run type such as train, eval, infer, export, score, average_checkpoints, or update_vocab

that are passed to the main script:

onmt-main --model_type <model> --config <config_file.yml> --auto_config <run_type> <run_options>

For more information and examples on how to use OpenNMT-tf, please visit our documentation.

Library

OpenNMT-tf also exposes well-defined and stable APIs, from high-level training utilities to low-level model layers and dataset transformations.

For example, the Runner class can be used to train and evaluate models with few lines of code:

import opennmt

config = {
    "model_dir": "/data/wmt-ende/checkpoints/",
    "data": {
        "source_vocabulary": "/data/wmt-ende/joint-vocab.txt",
        "target_vocabulary": "/data/wmt-ende/joint-vocab.txt",
        "train_features_file": "/data/wmt-ende/train.en",
        "train_labels_file": "/data/wmt-ende/train.de",
        "eval_features_file": "/data/wmt-ende/valid.en",
        "eval_labels_file": "/data/wmt-ende/valid.de",
    }
}

model = opennmt.models.TransformerBase()
runner = opennmt.Runner(model, config, auto_config=True)
runner.train(num_devices=2, with_eval=True)

Here is another example using OpenNMT-tf to run efficient beam search with a self-attentional decoder:

decoder = opennmt.decoders.SelfAttentionDecoder(num_layers=6, vocab_size=32000)

initial_state = decoder.initial_state(
    memory=memory, memory_sequence_length=memory_sequence_length
)

batch_size = tf.shape(memory)[0]
start_ids = tf.fill([batch_size], opennmt.START_OF_SENTENCE_ID)

decoding_result = decoder.dynamic_decode(
    target_embedding,
    start_ids=start_ids,
    initial_state=initial_state,
    decoding_strategy=opennmt.utils.BeamSearch(4),
)

More examples using OpenNMT-tf as a library can be found online:

  • The directory examples/library contains additional examples that use OpenNMT-tf as a library
  • nmt-wizard-docker uses the high-level opennmt.Runner API to wrap OpenNMT-tf with a custom interface for training, translating, and serving

For a complete overview of the APIs, see the package documentation.

Additional resources