OpenNMT Py Versions Save

• fix major bug in 3.1.0 introduced with LoRa (3.1.0 not available)

• updated docs with Sphinx 6.4
• Restore source features to v3 (thanks @anderleich)
• add inline tags transform (thanks @panosk)
• add docify transform to allow doc-level training / inference
• fix NLLB training (decoder_start_token)
• New! LoRa adapters to finetune big models (egs: NLLB 3.3B)
• various bug fixes

• override_opts to override checkpoints opt when training from
• normalize transform based on (Sacre)Moses scripts
• uppercase transform for adhoc data augmentation
• suffix transform
• Fuzzy match transform
• WMT17 detailed example
• NLLB-200 (from Meta/FB) models support (after conversion)
• various bug fixes

• fix loss normalization when using accum or nb GPU > 1
• use native CrossEntropyLoss with Label Smoothing. reported loss/ppl impacted by LS
• fix long-time coverage loss bug thanks Sanghyuk-Choi
• fix detok at scoring / fix tokenization Subword_nmt + Sentencepiece
• various small bugs fixed

3.0.2 (2022-12-07)

• pyonmttok.Vocab is now pickable. dataloader switched to spawn. (MacOS/Windows compatible)
• fix scoring with specific metrics (BLEU, TER)
• fix tensorboard logging
• fix dedup in batch iterator (only for TRAIN, was happening at inference also)
• New: Change: tgt_prefix renamed to tgt_file_prefix
• New: tgt_prefix / src_prefix used for "prefix" Transform (onmt/transforms/misc.py)
• New: process transforms of buckets in batches (vs per example) / faster

• fix dynamic scoring
• reinstate apex.amp level O1/O2 for benchmarking
• New: LM distillation for NMT training
• New: bucket_size ramp-up to avoid slow start
• fix special tokens order
• remove Library and add link to Yasmin's Tuto

v3.0 !

• Removed completely torchtext. Use Vocab object of pyonmttok instead
• Dataloading changed accordingly with the use of pytorch Dataloader (num_workers)
• queue_size / pool_factor no longer needed. bucket_size optimal value > 64K
• options renamed: rnn_size => hidden_size (enc/dec_rnn_size => enc/dec_hid_size)
• new tools/convertv2_v3.py to upgrade v2 models.pt
• inference with length_penalty=avg is now the default
• add_qkvbias (default false, but true for old model)

New features

• BLEU/TER (& custom) scoring during training and validation (#2198)
• LM related tools (#2197)
• Allow encoder/decoder freezing (#2176)
• Dynamic data loading for inference (#2145)
• Sentence-level scores at inference (#2196)
• MBR and oracle reranking scoring tools (#2196)

Fixes and improvements

• Updated beam exit condition (#2190)
• Improve scores reporting (#2191)
• Fix dropout scheduling (#2194)
• Better catch CUDA ooms when training (#2195)
• Fix source features support in inference and REST server (#2109)
• Make REST server more flexible with dictionaries (#2104)
• Fix target prefixing in LM decoding (#2099)

New features

• Support source features (thanks @anderleich !)

Fixes and improvements

• Adaptations to relax torch version
• Customizable transform statistics (#2059)
• Adapt release code for ctranslate2 2.0

Fixes and improvements

• Fix update_vocab for LM (#2056)