Serving

TensorFlow

Exporting a SavedModel

OpenNMT-tf can export SavedModel packages for inference in other environments, for example with TensorFlow Serving. A model export contains all information required for inference: the graph definition, the weights, and external assets such as vocabulary files. It typically looks like this on disk:

toy-ende/export/
├── assets
│   ├── src-vocab.txt
│   └── tgt-vocab.txt
├── saved_model.pb
└── variables
    ├── variables.data-00000-of-00001
    └── variables.index

Models can be manually exported using the export run type:

onmt-main --config my_config.yml --auto_config export --export_dir ~/my-models/ende

Automatic evaluation during the training can also export models, see Training to learn more.

Running a SavedModel

Once a SavedModel is exported, OpenNMT-tf is no longer needed to run it. However, you will need to know the input and output nodes of your model. You can use the saved_model_cli script provided by TensorFlow for inspection, e.g.:

saved_model_cli show --dir ~/my-models/ende \
    --tag_set serve --signature_def serving_default

Some examples using exported models are available in the examples/serving directory.

Input preprocessing and tokenization

TensorFlow Serving only runs TensorFlow operations. Preprocessing functions such as the tokenization is sometimes not implemented in terms of TensorFlow ops (see Tokenization for more details). In this case, these functions should be run outside of the TensorFlow runtime, either by the client or a proxy server.

CTranslate2

CTranslate2 is an optimized inference engine for OpenNMT models that is typically faster, lighter, and more customizable than the TensorFlow runtime.

Selected models can be exported to the CTranslate2 format directly from OpenNMT-tf. An additional export_format option should be configured to select this export type:

onmt-main [...] export --export_dir ~/my-models/ende --export_format ctranslate2

The same option is available in the eval block of the YAML configuration when exporting models during the training.