I'm looking for pretrained neural translation models, preferably for pytorch or tensorflow.
Pretrained models should be well documented. Ideally, the training should be reproducable with both data and hyperparameters being available.
Models should be reasonably easy to deploy (assuming you have knowledge in pytorch or tensorflow).
So far I've found:
- fairseq, based on pytorch. As of April 2020 it is still under active development. Many pretrained models are available: https://github.com/pytorch/fairseq/tree/master/examples/translation
- tensor2tensor, based on tensorflow 1. As of April 2020, it is recommended to use the successor trax. Some pretrained models are available: https://github.com/tensorflow/tensor2tensor
- trax, based on tensorflow 2(?). As of April 2020, pretrained models don't seem to be available: https://github.com/tensorflow/tensor2tensor
While there are some pretrained models in these repositories, they seem to be mostly research projects.
As such, there are only pretrained models for a few select languages (which are used for competitive benchmarking).
I'm looking for an integrated collection of translation models across languages.
Unfortunately, I don't think that such a ressource exists. Maybe we can start collecting language pairs here. But please only answer if your pretrained model will be available for a long time (hosting needs to be secure) and if it is reproducible.