Etienne Monneret

According to the profiler, the time is not spent in the Python code, but in TensorFlow. Thus, the 30% gain doesn't come from a possible Python parallelisation effect.  ...

By "low-level batch translation", I understand that you mean: with a direct neural translation, without MMT on-the-fly tuning. This would be only the large (4) loop of my proposition above....

I read it quickly. As far as I understand, they do not really explain what can be obtained when looping the process. I think there is a lot to gain....

Suppose you want to train a DE-PT cooking dedicated model, you don't have any DE-PT well translated cooking data, but you have large monolingual cooking data. Start from 2 generic...

I know that using back-translation is improving the model. My questions was: 1) what occurs when looping the training? Is the model continuously improving? I don't know a work with...

Experimentation started here: http://www.lingua-et-machina.eu/transinfo/ Updated every hour with new published news. The models were not trained with news data, and they have quite special formulations and vocabulary. For the moment,...

Here is a very tiny code update (not sure it would really make a difference on any case, but cleaner): [TagProjector.java.V2b.zip](https://github.com/modernmt/modernmt/files/5179810/TagProjector.java.V2b.zip) This example shows that, if MMT is ok with...

My tool is supposed to be used by people not being technically capable to edit their SVG files. I will wait for your improvement. :)

Hi @R1j1t, perhaps later I will find the time to build such a PR. But, on the team side, if you have a direct access to edit, it's only few...

Try to create a virtual env using Python 3.6 (rather than 3.7)