By default, OpenNMT-tf expects and generates tokenized text. The users are thus responsible to tokenize the input and detokenize the output with the tool of their choice.

However, OpenNMT-tf provides tokenization tools based on the C++ OpenNMT Tokenizer that can be used in 2 ways:

  • offline: use the provided scripts to manually tokenize the text files before the execution and detokenize the output for evaluation
  • online: configure the execution to apply tokenization and detokenization on-the-fly

Note : the pyonmttok package is only supported on Linux as of now.

Configuration files

YAML files are used to set the tokenizer options to ensure consistency during data preparation and training. For example, this configuration defines in simple word-based tokenization using the OpenNMT tokenizer:

mode: aggressive
joiner_annotate: true
segment_numbers: true
segment_alphabet_change: true

For a complete list of available options, see the Tokenizer documentation.

Offline usage

You can invoke the onmt-tokenize-text script directly and pass the tokenizer configuration:

$ echo "Hello world!" | onmt-tokenize-text --tokenizer_config config/tokenization/aggressive.yml
Hello world ■!

Online usage

A key feature is the possibility to tokenize the data on-the-fly during the training. This avoids the need of storing tokenized files and also increases the consistency of your preprocessing pipeline.

Here is an example workflow:

1. Build the vocabularies with the custom tokenizer, e.g.:

onmt-build-vocab --tokenizer_config config/tokenization/aggressive.yml --size 50000 --save_vocab data/enfr/en-vocab.txt data/enfr/en-train.txt
onmt-build-vocab --tokenizer_config config/tokenization/aggressive.yml --size 50000 --save_vocab data/enfr/fr-vocab.txt data/enfr/fr-train.txt

The text files are only given as examples and are not part of the repository.

2. Reference the tokenizer configurations in the data configuration, e.g.:

  source_tokenization: config/tokenization/aggressive.yml
  target_tokenization: config/tokenization/aggressive.yml


  • As of now, tokenizers are not part of the exported graph. However, all tokenization resources (configuration, subword models, etc.) are registered as additional assets in the SavedModel bundle in the assets.extra directory.
  • Predictions saved during inference or evaluation are detokenized. Consider using the “sacreBLEU” or “BLEU-detok” external evaluators.