CHANGELOG.md

[Unreleased]

Breaking changes

New features

• '-' stands for stdin for inference tools (translate, lm, tag)

Fixes and improvements

• Fix empty translation returned by the REST translation server

v0.9.1 (2017-11-16)

Fixes and improvements

• Fix missing normalization during translation
• Fix normalization when the command contains pipes
• Fix incorrect TER normalization (#424)
• Fix error when the file to translate contains empty lines

v0.9.0 (2017-11-07)

Breaking changes

• Learning rate is also decayed when using Adam
• Fix some wrong tokenization rules (punctuation-numbers)
• -report_every option is renamed to -report_progress_every
• -EOT_marker option is renamed to -bpe_EOT_marker for tokenize.lua
• -BOT_marker option is renamed to -bpe_BOT_marker for tokenize.lua
• bit32 package is now required for LuaJIT users

New features

• Dynamic dataset to train on large and raw training data repository
• Convolutional encoder
• Shallow fusion of language model in decoder
• Lexically constrained beam search
• TER validation metric
• Protection blocks for tokenization - and implement placeholder
• Hook to call external normalization
• JSON log formatting when the log file suffix is .json
• Training option to save the validation translation to a file
• Training option to reset the optimizer states when the learning rate is decayed
• Training option to update the vocabularies during a retraining
• Translation option to save alignment history
• Translation translation option to mark replaced tokens with ｟unk:xxxxx｠
• Tokenization option to split numbers on each digit
• Multi-model rest server using yaml config file

Fixes and improvements

• Allow disabling gradients clipping with -max_grad_norm 0
• Allow disabling global parameters initialization with -param_init 0
• Introduce error estimation in scorer for all metrics
• Reduce memory footprint of Adam, Adadelta and Adagrad optimizers
• Make validation data optional for training
• Faster tokenization (up to x2 speedup)
• Fix missing final model with some values of -save_every_epochs
• Fix validation score delta that was applied in the incorrect direction
• Fix LuaJIT out of memory issues in learn_bpe.lua
• Fix documentation generation of embedded tokenization options
• Fix release of sequence tagger models

v0.8.0 (2017-06-28)

Breaking changes

• Models previously trained with -pdbrnn or -dbrnn are no more compatible
• -start_decay_ppl_delta option is renamed to -start_decay_score_delta
• -decay perplexity_only option is renamed to -decay score_only

Deprecations

• -brnn, -dbrnn and -pdbrnn options are replaced by -encoder_type <type> for future extensions
• -sample_tgt_vocab option is renamed -sample_vocab and is extended to language models

New features

• Implement inference for language models for scoring or sampling
• Support variational dropout and dropout on source sequence
• Support several validation metrics: loss, perplexity, BLEU and Damerau-Levenshtein edit ratio
• Add option in preprocessing to check that lengths of source and target are equal (e.g. for sequence tagging)
• Add -pdbrnn_merge option to define how to reduce the time dimension
• Add option to segment mixed cased words
• Add option to segment words of given alphabets or when switching alphabets
• Add Google's NMT encoder
• Add external scorer script for BLEU and Damerau-Levenshtein edit ratio
• Add script to average multiple models
• Add option to save the beam search as JSON

Fixes and improvements

• Support input vectors for sequence tagging
• Fix incorrect gradients when using variable length batches and bidirectional encoders

v0.7.1 (2017-05-29)

Fixes and improvements

• Fix backward compatibility with older models using target features
• Fix importance sampling when using multiple GPUs
• Fix language models training

v0.7.0 (2017-05-19)

Breaking changes

• -sample_w_ppl option is renamed -sample_type for future extensions

New features

• Support vectors as inputs using Kaldi input format
• Support parallel file alignment by index in addition to line-by-line
• Add script to generate pretrained word embeddings:
  • from Polyglot repository
  • from pretrained word2vec, GloVe or fastText files
• Add an option to only fix the pretrained part of word embeddings
• Add a bridge layer between the encoder and decoder to define how encoder states are passed to the decoder
• Add epoch_only decay strategy to only decay learning based on epochs
• Make epoch models save frequency configurable
• Optimize decoding and training with target vocabulary reduction (importance sampling)
• Introduce partition data sampling

Fixes and improvements

• Improve command line and configuration file parser
  • space-separated list of values
  • boolean arguments
  • disallow duplicate command line options
  • clearer error messages
• Improve correctness of DBiEncoder and PDBiEncoder implementation
• Improve unicode support for languages using combining marks like Hindi
• Improve logging during preprocessing with dataset statistics
• Fix translation error of models profiled during training
• Fix translation error of models trained without attention
• Fix error when using one-layer GRU
• Fix incorrect coverage normalization formula applied during the beam search

v0.6.0 (2017-04-07)

Breaking changes

• -fix_word_vecs options now requires 0 and 1 as argument for a better retraining experience

New features

• Add new encoders: deep bidirectional and pyramidal deep bidirectional
• Add attention variants: no attention and dot, general or concat global attention
• Add alternative learning rate decay strategy for SGD training
• Introduce dynamic parameter change for dropout and fixed word embeddings
• Add length and coverage normalization during the beam search
• Add translation option to dump input sentence encoding
• Add TensorBoard metrics visualisation with Crayon
• [experimental] Add sequence tagger model

Fixes and improvements

• Check consistency of option settings when training from checkpoints
• Save and restore random number generator states from checkpoints
• Output more dataset metrics during the preprocessing
• Improve error message on invalid options
• Fix missing n-best hypotheses list in the output file
• Fix individual losses that were always computed when using random sampling
• Fix duplicated logs in parallel mode

v0.5.3 (2017-03-30)

Fixes and improvements

• Fix data loading during training

v0.5.2 (2017-03-29)

Fixes and improvements

• Improve compatibility with older Torch versions missing the fmod implementation

v0.5.1 (2017-03-28)

Fixes and improvements

• Fix translation with FP16 precision
• Fix regression that make tds mandatory for translation

v0.5.0 (2017-03-06)

New features

• Training code is now part of the library
• Add -fallback_to_cpu option to continue execution on CPU if GPU can't be used
• Add standalone script to generate vocabularies
• Add script to extract word embeddings
• Add option to prune vocabularies by minimum word frequency
• New REST server
• [experimental] Add data sampling during training
• [experimental] Add half floating point (fp16) support (with cutorch@359ee80)

Fixes and improvements

• Make sure released model does not contain any serialized function
• Reduce size of released BRNN models (up to 2x smaller)
• Reported metrics are no longer averaged on the entire epoch
• Improve logging in asynchronous training
• Allow fixing word embeddings without providing pre-trained embeddings
• Fix pretrained word embeddings that were overriden by parameters initialization
• Fix error when using translation server with GPU model
• Fix gold data perplexity reporting during translation
• Fix wrong number of attention vectors returned by the translator

v0.4.1 (2017-02-16)

Fixes and improvements

• Fix translation server error when clients send escaped unicode sequences
• Fix compatibility issue with the :split() function

v0.4.0 (2017-02-10)

Breaking changes

• New translator API for better integration

New features

• Profiler option
• Support hypotheses filtering during the beam search
• Support individually setting features vocabulary and embedding size
• [experimental] Scripts to interact with the benchmark platform
• [experimental] Language modeling example

Fixes and improvements

• Improve beam search speed (up to 90% faster)
• Reduce released model size (up to 2x smaller)
• Fix tokenization of text containing the joiner marker character
• Fix -joiner_new option when using BPE
• Fix joiner marker generated without the option enabled
• Fix translation server crash on Lua errors
• Fix error when loading configuration files containing the gpuid option
• Fix BLEU drop when applying beam search on some models
• Fix error when using asynchronous parallel mode
• Fix non SGD model serialization after retraining
• Fix error when using -replace_unk with empty sentences in the batch
• Fix error when translating empty batch

v0.3.0 (2017-01-23)

Breaking changes

• Rename -epochs option to -end_epoch to clarify its behavior
• Remove -nparallel option and support a list of comma-separated identifiers on -gpuid
• Rename -sep_annotate option to -joiner_annotate

New features

• ZeroMQ translation server
• Advanced log management
• GRU cell
• Tokenization option to make the token separator an independent token
• Tokenization can run in parallel mode

Fixes and improvements

• Zero-Width Joiner unicode character (ZWJ) is now tokenizing but as a joiner
• Fix Hangul tokenization
• Fix duplicated tokens in aggressive tokenization
• Fix error when using BRNN and multiple source features
• Fix error when preprocessing empty lines and using additional features
• Fix error when translating empty sentences
• Fix error when retraining a BRNN model on multiple GPUs

v0.2.0 (2017-01-02)

Breaking changes

• -seq_length option is split into -src_seq_length and -tgt_seq_length

New features

• Asynchronous SGD
• Detokenization
• BPE support in tokenization

Fixes and improvements

• Smaller memory footprint during training
• Smaller released model size after a non-SGD training
• Fix out of memory errors in preprocessing
• Fix BRNN models serialization and release
• Fix error when retraining a model
• Fix error when using more than one feature

v0.1.0 (2016-12-19)

Initial release.