v2.3

• introduced DistributedDataParallel.
• implemented language tags, see notebooks/torchhub.ipynb
• released a iwslt14 de-en-fr multilingual model (trained using DDP)
• special symbols definition refactoring
• configuration refactoring
• autocast refactoring
• bugfixes
• upgrade to python 3.11, torch 2.1.2
• documentation refactoring

v2.2

• compatibility with torch 1.13 tested
• torchhub introduced
• bugfixes, minor refactoring (#119, #198)

v2.1

• upgrade to python 3.10, torch 1.12
• replace Automated Mixed Precision from NVIDA's amp to Pytorch's amp package
• replace discord.py with pycord in the Discord Bot demo
• data Iterator refactoring (#189, #190, #191)
• migrate to pytorch's torch.testing.assert_close to check tensors in unittests
• add wmt14 ende / deen benchmark trained with joey v2 from scratch
• bugfixes

Joey NMT 2.0

Breaking changes:

• upgrade to python 3.9, torch 1.11
• torchtext.legacy dependencies are completely replaced by torch.utils.data
• joeynmt/tokenizers.py: handles tokenization internally (also supports bpe-dropout!)
• joeynmt/datasets.py: loads data from plaintext, tsv, and huggingface's datasets
• scripts/build_vocab.py: trains subwords, creates joint vocab
• enhancement in decoding
• scoring with hypotheses or references
• repetition penalty, ngram blocker
• attention plots for transformers
• yapf, isort, flake8 introduced
• bugfixes, minor refactoring

Requirements update

Six >= 1.12

Beam search & checkpointing improvements, dependency update

• upgrade to sacrebleu 2.0, python 3.7, torch 1.8
• bug fixes:
  • heaps in checkpoint maintenance #153
  • beam search stopping criterion #149
  • removing final BPE merge markers in hypotheses (dsfsi/masakhane-web#33)
  • keeping best and last ckpts #136
  • using utf encoding when opening files #150
• f-style formatting

n-best decoding, checkpointing, dependency updates

You can now retrieve the n-best outputs during inference (rather than just the one best translation) and track the latest checkpoint (for continuing training). We also added a colab for training a small translation model on the Tatoeba task. Now operating on Torch v1.8.0 and using deprecated Torchtext dataset implementations from v0.9.

1.0

Additions:

• Multi-GPU processing
• Data loading improvements
• Tokenizer integration
• Japanese benchmarks

Pre-release v0.9

Stable recurrent and Transformer models. Minor changes and refactoring might happen before v1.0.