Frank Seide

It also implements target factors, which makes it a little tricky. I am also curious whether it will work for you. Get Outlook for iOS ________________________________ From: Stamenov Sent: Wednesday,...

I hope this is what you expect. Source factors are straight forward, but target factors impact decoding quite a bit. Unlike most, our implementation adds the factors I stead of...

We don't have good documentation yet for factors, unfortunately. The closest is the README of the factored-segmenter tool, which is the SentencePiece equivalent/wrapper that we use in Microsoft Translator: https://github.com/microsoft/factored-segmenter....

This is fully functional. It does, however, use a very specific design in case of tied vocabularies.

I agree with lazy initialization.

Q1: The embedding of the sentence-start (`BOS` or ``) context is hard-coded to be 0. It is not copied from the embedding matrix. I always felt that's a bug, but...

Actually no need to use `marian_vocab`. You just need these three entries, and by convention at the start of the vocab: ``` ``` I create my vocabs with something like...

To add to this, if you don't provide a vocabulary, Marian will create the file for you. The created file is the equivalent of this: ``` LC_AL=C cat data |...

That should solve it.

Are you looking for high-level bindings e.g. to load a model and use it for translation, or low-level bindings that allow you to define your models on operator level in...