OpenNMT-tf 1.22.0

OpenNMT-tf 1.22.0

New features

• Unified dynamic decoding implementation
• Support random sampling during beam search

Fixes and improvements

• More than 3x faster beam search decoding
• Improve correctness of alignment vectors for the other hypotheses

OpenNMT-tf 1.21.7

OpenNMT-tf 1.21.7

Fixes and improvements

• Fix error when sharing target embedding and softmax weights
• Run checkpoint utilities in a separate graph to avoid possible variables collision

OpenNMT-tf 1.21.6

OpenNMT-tf 1.21.6

Fixes and improvements

• Fix inputter initialization from the configuration file: fields such as source_tokenization, target_embedding, etc. were ignored in 1.21.0.

OpenNMT-tf 1.21.5

OpenNMT-tf 1.21.5

Fixes and improvements

• Fix compatibility issue with legacy TensorFlow 1.4
• Fix inference of language models
• Fix inference error when using replace_unknown_target and the alignment vector was empty

OpenNMT-tf 1.21.4

OpenNMT-tf 1.21.4

Fixes and improvements

• Fix error during manual model export

OpenNMT-tf 1.21.3

OpenNMT-tf 1.21.3

Fixes and improvements

• Fix multi GPU training: some variables were not correctly reused when building the graph for other devices

OpenNMT-tf 1.21.2

OpenNMT-tf 1.21.2

Fixes and improvements

• Fix checkpoint restore during evaluation when using a PositionEmbedder or a DenseBridge in the decoder
• Fix inputters parameter sharing: shared variables were overriden when the layer was invoked (requires retraining)

OpenNMT-tf 1.21.1

OpenNMT-tf 1.21.1

Fixes and improvements

• Allow configuring tagging_scheme in the data configuration
• Fix dimension mismatch when using replace_unknown_target

OpenNMT-tf 1.21.0

OpenNMT-tf 1.21.0

New features

• New experimental model type LanguageModel to train generative language models (see the example GPT-2 configuration). The usage is the same as a sequence to sequence model except that "labels" data should not be set.
• cosine_annealing learning rate decay
• weight_decay parameter to apply decoupled weight decay regularization (as described in Loshchilov et al. 2017)
• sampling_temperature parameter to control the randomness of the generation

Fixes and improvements

• Improve correctness of MeanEncoder for variable lengths inputs (requires TensorFlow 1.13+)
• Internal refactoring and changes to prepare for 2.0 transition

OpenNMT-tf 1.20.1

OpenNMT-tf 1.20.1

Fixes and improvements

• Fix score run type that was broken after some internal refactoring
• Internal changes to improve TensorFlow 2.0 support