OpenNMT / CTranslate

Lightweight C++ translator for OpenNMT Torch models (deprecated)
https://opennmt.net/
MIT License
79 stars 50 forks source link

CTranslate does not work with Deep bidirectional encoders #43

Open fdalvi opened 5 years ago

fdalvi commented 5 years ago

CTranslate silently quits if the model being loaded was trained with the -encoder_type dbrnn option. I tried to run cli/translate under gdb, but it did not give additional information, other than loading the model and silently exiting. A model trained with similar data and options, but -encoder_type brnn works fine.

Is this expected since Deep bidirectional encoders were introduced later in the lifetime of OpenNMT, and can one expect this to be implemented in CTranslate any time soon?

guillaumekln commented 5 years ago

The supported features are documented in the README:

https://github.com/OpenNMT/CTranslate#supported-features

and dbrnn is not one of them.

There is no plan to support this encoder. Did you get better results with dbrnn compared to brnn?

fdalvi commented 5 years ago

Ah, sorry I missed that!

As for results, yes, we do consistently get better results with dbrnn compared to brnn (we train models involving Arabic, English & German)