Any recommendation on how to use a trained OpenNMT model on iOS?

ironflood · October 24, 2018, 9:24am

Hi,

I trained a model on OpenNMT and it does really well on the task. For connectivity reason I need to have an embedded version on iOS, any pointers on how to achieve this and what is the current recommended route? I’ve already read about several approaches:

ONNX
tensorflow >= 1.5 to core ML model conversion
embedding python code into native iOS using https://github.com/pybee/Python-Apple-support

I wonder how much would need to be re-written, whether any solution can already handle like the attention layers and other specificities of the OpenNMT model. I guess beam search would need to be re-written, anything else?

Thanks a lot!

guillaumekln · October 24, 2018, 9:27am

Hello,

Which OpenNMT version did you use for your training?

ironflood · October 24, 2018, 9:39am

Hi Guillaume!

The version is from Dec 2017 (tensorflow), but I’m not at all against retraining on one of the latest version or even on a different language (pytorch, …).

guillaumekln · October 25, 2018, 10:06am

I think your best bet would be to train a model using the original OpenNMT and try to compile CTranslate for iOS.