This page presents a minimal workflow to get you started in using OpenNMT-tf.
Step 0: Install OpenNMT-tf¶
We recommend using
virtualenv to setup and configure the environment for this quickstart:
virtualenv pyenv source pyenv/bin/activate pip install OpenNMT-tf[tensorflow_gpu]
Step 1: Prepare the data¶
To get started, we propose to download a toy English-German dataset for machine translation containing 10k tokenized sentences:
wget https://s3.amazonaws.com/opennmt-trainingdata/toy-ende.tar.gz tar xf toy-ende.tar.gz cd toy-ende
The first step is to build the source and target word vocabularies from the training files:
onmt-build-vocab --size 50000 --save_vocab src-vocab.txt src-train.txt onmt-build-vocab --size 50000 --save_vocab tgt-vocab.txt tgt-train.txt
Then, the data files should be declared in a YAML configuration file, let’s name it
model_dir: run/ data: train_features_file: src-train.txt train_labels_file: tgt-train.txt eval_features_file: src-val.txt eval_labels_file: tgt-val.txt source_words_vocabulary: src-vocab.txt target_words_vocabulary: tgt-vocab.txt
Step 2: Train the model¶
onmt-main train_and_eval --model_type NMTSmall --auto_config --config data.yml
This command will start the training and evaluation loop of a small RNN-based sequence to sequence model. The
--auto_config flag selects the best settings for this type of model.
The training will regularly produce checkpoints in the
run/ directory. To monitor the training progress, some logs are displayed in the console. However, to visually monitor the training we suggest using TensorBoard:
Step 3: Translate¶
onmt-main infer --auto_config --config data.yml --features_file src-test.txt
This command can be executed as soon as a checkpoint is saved by the training; the most recent checkpoint will be used by default. The predictions will be printed on the standard output.
That’s it! You successfully executed the 3 main steps to prepare, run, and evaluate an OpenNMT-tf training.
While this example gave you a quick overview of a typical OpenNMT-tf workflow, it will not produce state of the art results. The selected dataset and model are too small for this task.
To go further, here are some pointers: