NJUNMT-tf is a general purpose sequence modeling tool in TensorFlow while neural machine translation is the main target task.

NJUNMT-tf builds NMT models almost from scratch without any high-level TensorFlow APIs which often hide details of many network components and lead to obscure code structure that is difficult to understand and manipulate. NJUNMT-tf only depends on basic TensorFlow modules, like array_ops, math_ops and nn_ops. Each operation in the code is under control.

NJUNMT-tf focuses on modularity and extensibility using standard TensorFlow modules and practices to support advanced modeling capability:

• arbitrarily complex encoder architectures, e.g. Bidirectional RNN encoder, Unidirectional RNN encoder and self-attention.
• arbitrarily complex decoder architectures, e.g. Conditional GRU/LSTM decoder, attention decoder and self-attention.
• hybrid encoder-decoder models, e.g. self-attention encoder and RNN decoder or vice versa.

and all of the above can be used simultaneously to train novel and complex architectures.

The code also supports:

• model ensemble.
• learning rate decaying according to loss on evaluation data.
• model validation on evaluation data with BLEU score and early stop strategy.
• monitoring with TensorBoard.
• capability for BPE

• tensorflow (>=1.6)
• pyyaml

Here is a minimal workflow to get you started in using NJUNMT-tf. This example uses a toy Chinese-English dataset for machine translation with a toy setting.

1. Build the word vocabularies:

python -m bin.generate_vocab testdata/toy.zh --max_vocab_size 100  > testdata/vocab.zh
python -m bin.generate_vocab testdata/toy.en0 --max_vocab_size 100  > testdata/vocab.en

2. Train with preset sequence-to-sequence parameters:

export CUDA_VISIBLE_DEVICES=
python -m bin.train --model_dir test_model \
    --config_paths "
        ./njunmt/example_configs/toy_seq2seq.yml,
        ./njunmt/example_configs/toy_training_options.yml,
        ./default_configs/default_optimizer.yml"

3. Translate a test file with the latest checkpoint:

export CUDA_VISIBLE_DEVICES=
python -m bin.infer --model_dir test_models \
  --infer "
    beam_size: 4
    source_words_vocabulary: testdata/vocab.zh
    target_words_vocabulary: testdata/vocab.en" \
  --infer_data "
    - features_file: testdata/toy.zh
      labels_file: testdata/toy.en
      output_file: toy.trans
      output_attention: false"

Note: do not expect any good translation results with this toy example. Consider training on larger parallel datasets instead.

As you can see, there are two ways to manipulate hyperparameters of the process:

• tf FLAGS
• yaml-style config file

For example, there is a config file specifying the datasets for training procedure.

# datasets.yml
data:
  train_features_file: testdata/toy.zh
  train_labels_file: testdata/toy.en0
  eval_features_file: testdata/toy.zh
  eval_labels_file: testdata/toy.en
  source_words_vocabulary: testdata/vocab.zh
  target_words_vocabulary: testdata/vocab.en

You can either use the command:

python -m bin.train --config_paths "datasets.yml" ...

or

python -m bin.train --data "
    train_features_file: testdata/toy.zh
    train_labels_file: testdata/toy.en0
    eval_features_file: testdata/toy.zh
    eval_labels_file: testdata/toy.en
    source_words_vocabulary: testdata/vocab.zh
    target_words_vocabulary: testdata/vocab.en" ...

They are of the same effect.

The available FLAGS (or the top levels of yaml configs) for bin.train are as follows:

• config_paths: the paths for config files
• model_dir: the directory for saving checkpoints
• train: training options, e.g. batch size, maximum length
• data: training data, evaluation data, vocabulary and (optional) BPE codes
• hooks: a list of training hooks (not provided, in the current version)
• metrics: a list of validation metrics on evaluation data
• model: the class name of the model
• model_params: parameters for the model
• optimizer_params: parameters for optimizer

The available FLAGS (or the top levels of yaml configs) for bin.infer are as follows:

• config_paths: the paths for config files
• model_dir: the checkpoint directory or directories separated by commas for model ensemble
• infer: inference options, e.g. beam size, length penalty rate
• infer_data: a list of data file to be translated
• weight_scheme: the weight scheme for model ensemble (only "average" available now)

Note that:

• each FLAG should be a string of yaml-style
• the hyperparameters provided by FLAGS will overwrite those presented in config files
• illegal parameters will interrupt the program, so see sample.yml of more detailed discription for each parameter.

The RNN benchmarks are performed on 1 GTX 1080Ti GPU with predefined configurations:

• default_configs/adam_loss_decay.yml
• default_configs/default_metrics.yml
• default_configs/default_training_options.yml
• default_configs/seq2seq_cgru.yml

The Transformer benchmarks are performed on 1 GTX 1080Ti GPU with predefined configurations:

• default_configs/transformer_base.yml
• default_configs/transformer_training_options.yml

Note that in Transformer model, we set batch_tokens_size=2500 with update_cycle=10 to realize pseudo parallel training.

The datasets are preprocessed using fetch_wmt2017_ende.sh and fetch_wmt2018_zhen.sh referring to Edinburgh’s Report. The BLEU scores are evaluated by the wrapper script run_mteval.sh.

The following features remain unimplemented:

• multi-gpu training
• schedule sampling
• minimum risk training

The implementation is inspired by the following:

• Neural Machine Translation by Jointly Learning to Align and Translate
• dl4mt-tutorial
• OpenNMT-tf
• Google's seq2seq
  Massive Exploration of Neural Machine Translation Architectures
• THUMT
• Google's tensor2tensor
  Attention is All You Need
• Stronger Baselines for Trustable Results in Neural Machine Translation

Any comments or suggestions are welcome.

Please email zhaocq.nlp@gmail.com.