onmt-main train -h
onmt-main --model_type Transformer --config data.yml --auto_config train --with_eval
OpenNMT-tf uses TensorBoard to log information during the training. Simply start
tensorboard by setting the active log directory, e.g.:
then open the URL displayed in the shell to monitor and visualize several data, including:
training and evaluation loss
decoder sampling probability
Evaluation can be run automatically when using the
onmt-main [...] train --with_eval
This minimally requires you to set some evaluation files in your data configuration, e.g.:
data: eval_features_file: ... eval_labels_file: ...
By default, it will run every 5000 training steps and report evaluation results in the console output and on TensorBoard.
Export model on best metric¶
Automatic evaluation can also export an inference model when a metric reaches its best value so far. For example, the following configuration will make the training exports a model each time the evaluation scores the best BLEU score so far:
eval: external_evaluators: bleu export_on_best: bleu
These models are saved in the model directory under
export/<step>. See also Serving for more information about exported models.
Early stopping is useful to stop the training automatically when the model performance is not improving anymore.
For example, the following configuration stops the training when the BLEU score does not improve by more than 0.2 points in the last 4 evaluations:
eval: external_evaluators: bleu early_stopping: metric: bleu min_improvement: 0.2 steps: 4
OpenNMT-tf training can make use of multiple GPUs. In this mode, the graph is replicated over multiple devices and batches are processed in parallel. The resulting graph is equivalent to train with batches
N times larger, where
N is the number of used GPUs.
For example, if your machine has 4 GPUs, simply add the
onmt-main [...] train --num_gpus 4
Note that evaluation and inference will run on a single device.
Distributed training with Horovod¶
--horovod training flag enables distributed training with Horovod:
onmt-main [...] train --horovod
Similar to multi-GPU training, this is equivalent to train with batches
N times larger, where
N is the total number of Horovod processes.
See the Horovod documentation for more information about installation and usage.
Mixed precision training¶
Mixed precision can be enabled with the
onmt-main --model_type Transformer --auto_config --config data.yml --mixed_precision train
It uses TensorFlow’s
auto_mixed_precision optimization which converts selected nodes in the graph to
float16. During training, dynamic loss scaling is used.
Maximizing the FP16 performance¶
Some extra steps may be required to ensure good FP16 performance:
Mixed precision training requires a Volta GPU or above
Tensor Cores require the input dimensions to be a multiple of 8. You may need to tune your vocabulary size using
onmt-build-vocabwhich will ensure that
(vocab_size + 1) % 8 == 0(+ 1 is the
<unk>token that is automatically added during the training).
Continuing from a stopped training¶
This is the most common case of retrainings: the training was interrupted but should run longer. In that case, simply launch the same command that you used for the initial training, e.g.:
# Start the training. onmt-main --model_type Transformer --config data.yml --auto_config train # ... the training is interrupted or stopped ... # Continue from the latest checkpoint. onmt-main --model_type Transformer --config data.yml --auto_config train
Note: If the train was stopped because
max_step was reached, you should first increase this value before continuing.
Fine-tune an existing model¶
Retraining can also be useful to fine-tune an existing model. For example in machine translation, it is faster to adapt a generic model to a specific domain compared to starting a training from scratch.
For a basic domain adaptation, we recommend tokenizing your data with a subword segmentation model (such as SentencePiece) and simply continue the training on new data.
However, OpenNMT-tf also offers some features that could be useful in more advanced workflows:
The run type
update_vocabcan be used to change the word vocabularies contained in a checkpoint while keeping learned weights of shared words (e.g. to add a domain terminology)
The command line argument
--checkpoint_pathcan be used to load the weights of an existing checkpoint while starting from a fresh training state (i.e. with new learning rate schedule and optimizer variables)
freeze_layersin the YAML configuration can be used to prevent the adapation of specific layers in the model