OpenNMT-tf 1.7.0

OpenNMT-tf 1.7.0

New features

• Command line option --session_config to configure TensorFlow session parameters (see the "Configuration" documentation)
• share_embeddings argument to SequenceToSequence models to configure the level of embeddings sharing

Fixes and improvements

• Fix error when using --data_dir and parallel inputs in the data configuration
• Fix TensorFlow 1.9+ compatibility issue when using MultiplyReducer
• Better support of other filesystems (HDFS, S3, etc.) for the model directory and data files

OpenNMT-tf 1.6.2

OpenNMT-tf 1.6.2

Fixes and improvements

• Fix invalid scheduled sampling implementation with RNN decoders
• Fix possible "Data loss: Invalid size in bundle entry" error when loading an averaged checkpoint
• Fix PyramidalEncoder error when input lengths are smaller than the total reduction factor

OpenNMT-tf 1.6.1

OpenNMT-tf 1.6.1

Fixes and improvements

• Fix error when initialiazing the ROUGE evaluator
• Improve Transformer models performance:
  • fix performance regression of tf.layers.conv1d on TensorFlow 1.7+ (+20%)
  • better caching during decoding (+15%)

OpenNMT-tf 1.6.0

OpenNMT-tf 1.6.0

New features

• New model exporter types:
  • best to export a new model only if it achieves the best evaluation loss so far (requires TensorFlow 1.9+)
  • final to only export the model at the end of the training
• Script and API to map a checkpoint to new vocabularies while keeping the trained weights of common words

Fixes and improvements

• Fix error when reloading models with target pretrained embeddings
• Fix error message when the number of requested GPUs is incompatible with the number of visible devices
• Fix error when continuing the training from an averaged checkpoint
• Re-introduce rouge as a dependency since its installation is now fixed
• Make code forward compatible with future pyonmttok options

OpenNMT-tf 1.5.0

OpenNMT-tf 1.5.0

New features

• MultistepAdamOptimizer to simulate trainings with large batch size (credits to Tensor2Tensor, requires TensorFlow 1.6+)
• --log_prediction_time flag to summarize inference execution time:
  • total prediction time
  • average prediction time
  • tokens per second
• Training option average_last_checkpoints to automatically average checkpoints at the end of the training
• Command line options --{inter,intra}_op_parallelism_threads to control the level of CPU parallelism
• [experimental] Average attention network (Zhang et al. 2018) in the Transformer decoder

Fixes and improvements

• Fix possible error when training RNN models with in-graph replication (time dimension mismatch)
• Fix error when the set vocabulary file is in the model directory

OpenNMT-tf 1.4.1

OpenNMT-tf 1.4.1

Fixes and improvements

• Make rouge an optional dependency to avoid install error in a fresh environment

OpenNMT-tf 1.4.0

OpenNMT-tf 1.4.0

New features

• score run type to score existing predictions
• ROUGE external evaluator for summarization
• CharRNNEmbedder that runs a RNN layer over character embeddings
• High level APIs for efficient data pipelines (see utils.data.{training,inference}_pipeline)

Fixes and improvements

• Add more control over model export after evaluation with the exporters option
• Allow JoinReducer to be used on the ParallelEncoder output

OpenNMT-tf 1.3.0

OpenNMT-tf 1.3.0

New features

• RNMT+ encoder
• L1, L2, and L1 L2 regularization penalties (see regularization parameter)
• Support additional post processing layers in ParallelEncoder:
  • outputs_layer_fn applied on each encoder outputs
  • combined_output_layer_fn applied on the combined output

Fixes and improvements

• Fix SequenceClassifier "last" encoding for variable length sequences

OpenNMT-tf 1.2.0

OpenNMT-tf 1.2.0

New features

• Return alignment history when decoding from an AttentionalRNNDecoder (requires TensorFlow 1.8+ when decoding with beam search)
• Boolean parameter replace_unknown_target to replace unknown target tokens by the source token with the highest attention (requires a decoder that returns the alignment history)
• Support for arbitrary transition layers in SequentialEncoder

Fixes and improvements

• Fix sequence reduction when the maximum sequence length is not equal to the tensor time dimension (e.g. when splitting a batch for multi-GPU training)
• The number of prefetched batches is automatically tuned when prefetch_buffer_size is not set (for TensorFlow 1.8+)

OpenNMT-tf 1.1.0

OpenNMT-tf 1.1.0

New features

• Update the OpenNMT tokenizer to 1.3.0 and use its Python package instead of requiring a manual compilation (Linux only)
• Include a catalog of models in the library package and allow model selection with the --model_type command line option

Fixes and improvements

• Fix error when using FP16 and an AttentionMechanism module (for TensorFlow 1.5+)
• Manual export will remove default-valued attributes from the NodeDefs (for TensorFlow 1.6+)
• Silence some deprecation warnings with recent TensorFlow versions
• Training option sample_buffer_size now accepts special values:
  • 0 or null to disable shuffling
  • -1 to create a buffer with the same size as the training dataset