DLMT, GNMT, MULTI_GNMT and their antonyms

[kadir-gunel]

Hello, I was wondering what are their "opened" names ? There was no explanation in the documentation.

B.R.

[emjotde]

Hi, these are already a bit outdated. With the next update those will be renamed, new names in parentheses:

• dl4mt (= amun): This signifies a model that is compatible with the original dl4mt model by Cho&Firat, also the same model that is available in basic Nematus. Works with the amun translation tool, but does not allow for extra features like multi-layer etc. The amun translator has been handwritten and optmized and adding features like multi-layer support is difficult.
• gnmt (= s2s): The old name came from the google architecture, but it is not that faithful a reimplementation. This is a more general model, that can be more or less easily extended, for instance it has multi-layer support. In the future it will be possible to define model architectures with a text file. These require the new s2s_translator binary (not yet in the amunmt repo, available in https://github.com/amunmt/marian). The translator is slower than amun, but uses the same architecture as the training tool. In theory definining a new model for training automatically means it can be used for translation with s2s_translator.
• multi_gnmt (= multi_s2s): the same as above, but can have two input sentences at the same time with separate attention models. Also to be used with the s2s_translator.

This stuff is sort of purposefully not documented yet, as I have to make this a bit more userfriendly. Working on it.

[kadir-gunel]

Thank you for information.