Checkpoints averaging

The script onmt-average-checkpoints can be used to average the parameters of several checkpoints, usually increasing the model performance. For example:

onmt-average-checkpoints \
    --model_dir run/baseline-enfr \
    --output_dir run/baseline-enfr/avg \
    --max_count 5

will average the parameters of the 5 latest checkpoints in the run/baseline-enfr model directory and save a new checkpoint in the directory run/baseline-enfr/avg.

Then, execute the inference by setting the --checkpoint_path option, e.g.:

onmt-main infer \
    --config config/my_config.yml \
    --features_file newstest2014.en.tok \
    --predictions_file newstest2014.en.tok.out \
    --checkpoint_path run/baseline-enfr/avg/model.ckpt-200000

To control the saving of checkpoints during the training, configure the following options in your configuration file:

  # (optional) Save a checkpoint every this many steps.
  save_checkpoints_steps: 5000
  # (optional) How many checkpoints to keep on disk.
  keep_checkpoint_max: 10


Sampling predictions from the output distribution can be an effective decoding strategy for back-translation, as described by Edunov et al. 2018. To enable this feature, you should configure the parameter sampling_topk. Possible values are:

  • k, sample from the k most likely tokens
  • 0, sample from the full output distribution
  • 1, no sampling (default)

For example:

  beam_width: 1
  sampling_topk: 0

N-best list

A n-best list can be generated for models using beam search. You can configure it in your configuration file:

  n_best: 5

With this option, each input line will simply generate N consecutive lines in the output, ordered from best to worst.

Note that N can not be greater than the configured beam_width.


The main OpenNMT-tf script can also be used to score existing translations via the score run type. It requires 2 command line options to be set:

  • --features_file, the input labels;
  • --predictions_file, the translations to score.


onmt-main score \
    --config config/my_config.yml \
    --features_file newstest2014.en.tok \
    --predictions_file newstest2014.en.tok.out

The command will write on the standard output the score generated for each line in the following format:

<score> ||| <translation>

where <score> is the negative log likelihood of the provided translation.

Tip: combining the n-best list generation and the scoring can be used for reranking translations.