Skip to content

Latest commit

 

History

History
76 lines (59 loc) · 2.24 KB

CHANGELOG.md

File metadata and controls

76 lines (59 loc) · 2.24 KB

Changelog

All notable changes to this project will be documented in this file.

Unreleased

Added

  • WMT2017 EN<->DE benchmark.
  • Load pretrained model.
  • Layer normalization for fflayer and LayerNormLSTMCell
  • Multi-GPU inference and evaluation (something wrong with training)
  • Pseudo Multi-GPU training with update_cycle option
  • Class LineReader

Changed

  • Default loss function.
  • Process beam results on GPU.
  • Replace multi-bleu.perl with an equivalent python version.
  • Replace model_analysis APIs with tf.profiler.
  • Move input_fields from class Dataset to class SequenceToSequence.
  • Tensorflow 1.6 at least

Removed

  • Configuration: multi_bleu_script and tokenize_scropt.
  • Class Dataset
  • loss_fns.py

Fixed

  • Bug when mode=EVAL and output attention with BPE.
  • All str.split(" ") => str.split().
  • Python3 compatibility of njunmt/tools/tokenizeChinese.py.

0.6.0 - 2018-01-26

Added

  • Evaluation entrance (bin/eval.py).
  • Cache for decoders (especially for TransformerDecoder).
  • Change Log.
  • Attention output of Transformer.

Changed

  • Tensorflow 1.4 requirement.
  • Attention file format when output_attention flag is enabled under mode=INFER/EVAL.
  • BPE in Vocab accepts an extra vocab file (also see --vocabulary in njunmt/tools/apply_bpe.py).
  • maximum_features_length and maximum_labels_length now indicate the length of encoded symbols (e.g. after BPE).
  • Attention interfaces.
  • BaseSeq2Seq -> SequenceToSequence.
  • More flexible and concise code structure of decoders and EnsembleModel.

Removed

  • input_prepose_processing_fn interface in decoders.
  • Redundant computation in TransformerDecoder (cache).

0.5.0 - 2018-01-09

Added

  • Transformer model.
  • Model ensemble.
  • Learning decaying functions.
  • Evaluation metrics with loss or BLEU score (multi-bleu.perl).
  • Beam search strategy with batch.
  • Tensorboard visualization.
  • Capability for BPE.
  • Shell script to fetch WMT2014 en-de data.

Changed

  • More flexible and concise code structure.

0.1.0 - 2017-11

Added

  • Sequence-to-sequence model with attention.