OpenNMT-tf 2.5.1

OpenNMT-tf 2.5.1

Fixes and improvements

• Fix first value of steps per second metric when continuing a training
• Fix reporting of learning rate values with some optimizers

OpenNMT-tf 2.5.0

OpenNMT-tf 2.5.0

New features

• Update to TensorFlow 2.1
  OpenNMT-tf now depends on the tensorflow pip package instead of tensorflow-gpu. The tensorflow package now includes GPU support by default. If you are upgrading an existing environment, we recommend uninstalling the tensorflow-gpu package before doing so. Also note that TensorFlow 2.1 requires CUDA 10.1.
• Update to TensorFlow Addons 0.7 with Windows support
• Data parameter export_vocabulary_assets to control whether vocabularies are exported as file assets or embedded in the graph itself

Fixes and improvements

• Fix error when reading loss returned by a sequence classifier model
• Fix error on sequences of length 1 in the sequence tagger with CRF
• Export vocabularies as file assets by default
• Remove an unnecessary synchronization when training multiple replicas
• Internal cleanup to fully remove Python 2 support

OpenNMT-tf 2.4.0

OpenNMT-tf 2.4.0

New features

• Transformer models with relative position representation: TransformerRelative and TransformerBigRelative

Fixes and improvements

• Fix invalid iteration count in checkpoint after a vocabulary update
• Fix possible NaN loss when retraining after a vocabulary update
• Fix checkpoint averaging for models with custom variable names
• Update opennmt.convert_to_v2_config to not fail on a V2 configuration
• Change default value of average_loss_in_time based on batch_type
• Reuse the same Python interpreter when running batch size auto-tuning

OpenNMT-tf 2.3.0

OpenNMT-tf 2.3.0

New features

• Predefined models NMTSmallV1, NMTMediumV1, and NMTBigV1 for compatibility with OpenNMT-tf v1
• Function opennmt.convert_to_v2_config to automatically upgrade a V1 configuration
• Function opennmt.utils.is_v1_checkpoint to detect OpenNMT-tf v1 checkpoints

Fixes and improvements

• Fix error when using auto_config with model LstmCnnCrfTagger
• Fix incomplete Model.create_variables after manually calling Model.build
• Increase LayerNorm default epsilon value to be closer to TensorFlow and PyTorch defaults

OpenNMT-tf 1.25.3

OpenNMT-tf 1.25.3

Fixes and improvements

• Fix compatibility with TensorFlow 1.15

OpenNMT-tf 2.2.1

OpenNMT-tf 2.2.1

Fixes and improvements

• Ensure that each training GPU receives a batch with the size configured by the user
• Fix error on the last partial batch when using multiple GPUs with single_pass enabled

OpenNMT-tf 2.2.0

OpenNMT-tf 2.2.0

New features

• Return detokenized predictions when using an in-graph tokenizer
• Injection of the special tokens <s> and </s> for language models can be configured with the data parameter sequence_controls

Fixes and improvements

• Fix the batch size in multi GPU training that was not scaled by the number of devices
• When updating vocabularies, mirror the existing embeddings distribution for newly created embeddings
• Fix error when running onmt-tokenize-text and onmt-detokenize-text scripts
• Transformer decoder now always returns the attention on the first source
• Calling model.initialize() also initializes the decoder (if any)

OpenNMT-tf 1.25.2

OpenNMT-tf 1.25.2

Fixes and improvements

• Revert "When updating vocabularies, weights of new words are randomly initialized instead of zero initialized" as the random distribution is possibly incompatible with the next trained layer

OpenNMT-tf 2.1.1

OpenNMT-tf 2.1.1

Fixes and improvements

• Force tokenizers and noisers to run on CPU to avoid errors when placing strings on GPU
• Do not apply noise modules on empty inputs
• Fix training of SequenceToSequence models with guided alignment
• Fix training of SequenceClassifier models

OpenNMT-tf 2.1.0

OpenNMT-tf 2.1.0

New features

• onmt-build-vocab script can now train a SentencePiece model and vocabulary from raw data
• Enable automatic model export during evaluation with export_on_best parameter
• Add perplexity in evaluation metrics
• Extend tokenization configuration to support in-graph tokenizers (currently SpaceTokenizer and CharacterTokenizer)
• Parameter decoder_subword_token_is_spacer to configure the type of decoder_subword_token
• [API] Support tf.RaggedTensor in tokenizer API

Fixes and improvements

• Fix early stopping logic
• Support spacer decoder_subword_token that is used as a suffix
• Improve errors when --model or --model_type options are invalid