OpenNMT-tf 2.0.1

OpenNMT-tf 2.0.1

Fixes and improvements

• Fix error when initializing language models
• Fix eval run when --features_file is not passed

OpenNMT-tf 2.0.0

OpenNMT-tf 2.0.0

OpenNMT-tf 2.0 is the first major update of the project. The goal of this release is to use the new features and practices introduced by TensorFlow 2.0.

Breaking changes

See the 2.0 Transition Guide for details about the following changes.

• TensorFlow 2.0 is now required
• Python 3.5 or greater is now required
• Checkpoints are no longer compatible as the code now uses object-based instead of name-based checkpointing (except Transformer checkpoints which are automatically upgraded when loaded)
• The onmt-main script now makes use of subparsers which require to move the run type and it specific options to the end of the command
• Some predefined models have been renamed or changed, see the transition guide
• Some parameters in the YAML configuration have been renamed or changed, see the transition guide
• A lot of public classes and functions have changed, see the API documentation for details
• TFRecord files generated with the opennmt.inputters.write_sequence_record function or the onmt-ark-to-records script are no longer compatible and should be re-generated

This version also changes the public API scope of the project:

• Only public symbols accessible from the top-level opennmt package and visible on the online documentation are now part of the public API and covered by backward compatibility guarantees
• The minimum required TensorFlow version is no longer part of the public API and can change in future minor versions

New features

• Object-based layers extending tf.keras.layers.Layer
• Many new reusable modules and layers, see the API documentation
• Replace tf.estimator by custom loops for more control and clearer execution path
• Multi-GPU training with tf.distribute
• Support early stopping based on any evaluation metrics
• Support GZIP compressed datasets
• eval run type accepts --features_file and --labels_file to evaluate files other than the ones defined in the YAML configuration
• Accept with_alignments: soft to output soft alignments during inference or scoring
• dropout can be configured in the YAML configuration to override the model values

Fixes and improvements

• Code and design simplification following TensorFlow 2.0 changes
• Improve logging during training
• Log level configuration also controls TensorFlow C++ logs
• All public classes and functions are now properly accessible from the root package opennmt
• Fix dtype error after updating the vocabulary of an averaged checkpoint
• When updating vocabularies, weights of new words are randomly initialized instead of zero initialized

Missing features

Some features available in OpenNMT-tf v1 were removed or are temporarily missing in this v2 release. If you relied on some of them, please open an issue to track future support or find workarounds.

• Asynchronous distributed training
• Horovod integration
• Adafactor optimizer
• Global parameter initialization strategy (a Glorot/Xavier uniform initialization is used by default)
• Automatic SavedModel export on evaluation
• Average attention network

OpenNMT-tf 1.25.1

OpenNMT-tf 1.25.1

Fixes and improvements

• Fix language model decoding from context to pass the correct decoding step to the model
• Fix WordDropout module when the input contains a single word
• When updating vocabularies, weights of new words are randomly initialized instead of zero initialized

OpenNMT-tf 1.25.0

OpenNMT-tf 1.25.0

New features

• Contrastive learning to reduce word omission errors

OpenNMT-tf 1.24.1

OpenNMT-tf 1.24.1

Fixes and improvements

• Fix NaN and inf values when using mixed precision and gradient clipping

OpenNMT-tf 1.24.0

OpenNMT-tf 1.24.0

New features

• Coverage penalty during beam search
• freeze_variables parameter to not update selected weights

Fixes and improvements

• Improve length penalty correctness

OpenNMT-tf 1.23.1

OpenNMT-tf 1.23.1

Fixes and improvements

• Fix invalid offset in alignment vectors
• Remove debug print in noise module

OpenNMT-tf 1.23.0

OpenNMT-tf 1.23.0

New features

• Support applying noise on the decoding output ("beam+noise" strategy from Edunov et al. 2018)

Fixes and improvements

• Fix crash on TensorFlow 1.14 due to an API change
• Fix unexpected number of hypotheses returned by exported models: by default only the best is returned and the number should be configured by either params/num_hypotheses or infer/n_best
• Do not require setting an external evaluator when saving evaluation predictions
• Write external evaluators summaries in a separate directory to avoid interfering with BestExporter

OpenNMT-tf 1.22.2

OpenNMT-tf 1.22.2

Fixes and improvements

• Update PyYAML to 5.1 and silence warnings
• Reduce default visible memory for batch size auto-tuning to handle larger memory usage variations during training

OpenNMT-tf 1.22.1

OpenNMT-tf 1.22.1

Fixes and improvements

• Fix usage of --checkpoint_path during training
• Fix embedding sharing with a ParallelInputter as input