OpenNMT-tf 2.25.0

New features

• Support TensorFlow 2.8
• Add training flag --continue_from_checkpoint to simplify continuing the training in another model directory (to be used in combination with --checkpoint_path)

Fixes and improvements

• Fix target unknowns replacement when the source has BOS or EOS tokens
• Update length constraints in Transformer automatic configuration to work with multiple sources
• Allow explicit configuration of the first argument of learning rate schedules (if not set, learning_rate is passed as the first argument)

OpenNMT-tf 2.24.0

New features

• Add experimental parameter mask_loss_outliers to mask high loss values considered as outliers (requires the tensorflow-probability module)

Fixes and improvements

• Fix TensorFlow Lite conversion for models using a PositionEmbedder layer
• Automatically pad the weights of linear layers to enable Tensor Cores in mixed precision training
• Correctly set the CTranslate2 options alignment_layer and alignment_heads when converting models using the attention reduction AVERAGE_LAST_LAYER
• Raise an error if a training dataset or annotation file has an unexpected size
• Warn about duplicated tokens when loading vocabularies

OpenNMT-tf 2.23.0

Changes

• Remove support for TensorFlow 2.3

New features

• Support TensorFlow 2.7
• Add CTranslate2 exporter with "int8_float16" quantization

Fixes and improvements

• Improve performance when applying the OpenNMT tokenization during training by vectorizing the dataset transformation
• Disable configuration merge for fields optimizer_params and decay_params
• Enable the CTranslate2 integration when installing OpenNMT-tf on Windows
• Include PyYAML 6 in supported versions

OpenNMT-tf 2.22.0

New features

• Support TensorFlow Lite conversion for Transformer models
• Make the options model_dir and auto_config available in both the command line and the configuration file
• Paths in the data configuration can now be relative to the model directory

Fixes and improvements

• Fix encoding when writing sentences detokenized by an in-graph tokenizer
• Always output the tokenized target in scoring even when a target tokenization is configured
• Enable the OpenNMT Tokenizer when installing OpenNMT-tf on Windows

OpenNMT-tf 2.21.0

New features

• Support TensorFlow 2.6
• Add tokenizer SentencePieceTokenizer, an in-graph SentencePiece tokenizer provided by tensorflow-text
• Add methods to facilitate training a Model instance:
  • model.compute_training_loss
  • model.compute_gradients
  • model.train
• Add --output_file argument to score command

Fixes and improvements

• Fix make_features method of inputters WordEmbedder and SequenceRecordInputter to work on a batch of elements
• Fix error when SelfAttentionDecoder is called without memory_sequence_length
• Fix ConvEncoder on variable-length inputs
• Support SacreBLEU 2.0

OpenNMT-tf 2.20.1

Fixes and improvements

• Fix missing environment variables in the child process when autotuning the batch size
• Fix error during evaluation when setting the inference parameter n_best > 1
• Fix error in Python serving example when using TensorFlow 2.5
• Log some information about the input layer after initialization (vocabulary size, special tokens, etc.)
• Update the minimum required pyonmttok version to 1.26.4 to include the latest fixes

OpenNMT-tf 2.20.0

New features

• Update the minimum required CTranslate2 version to 2.0

Fixes and improvements

• Set a timeout for each training attempt when autotuning the batch size
• Set keep_checkpoint_max to average_last_checkpoints if the later value is larger
• Update the minimum required pyonmttok version to include the latest fixes

OpenNMT-tf 2.19.0

New features

• Support TensorFlow 2.5

Fixes and improvements

• Fix dtype error in RNN decoder when enabling mixed precision
• Pass training flag to tokenizers to disable subword regularization in inference
• Update Sphinx from 2.3 to 3.5 to generate the documentation

OpenNMT-tf 2.18.1

Fixes and improvements

• Fix vocabulary update for models with shared embeddings
• Fix a compatibility issue with TensorFlow 2.5 for early users
• When all training attempts fail in batch size autotuning, log the error message of the last attempt

OpenNMT-tf 2.18.0

New features

• Add TransformerBaseSharedEmbeddings and TransformerBigSharedEmbeddings in the model catalog

Fixes and improvements

• Fix loss normalization when using sentence weighting
• Tune the automatic batch size selection to avoid some out of memory errors
• Harmonize training logs format when using onmt-main