OpenNMT-py v3.0.0
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)