Davide Caroselli

Hi all! We have just released the support to CUDA 11 in the latest `master` commit! We tested ModernMT on the new NVIDIA A100 with CUDA 11, PyTorch 1.7.1, and...

This seems an interesting feature, we will definitely work on it.

Hi @dominikacecylia ModernMT pre-processing pipeline (including tokenization) is a mandatory step in the process. Unfortunately, the hint of @mzeidhassan won't work because the tokenizer is needed to identify words. If...

> Remark: it seems wiki pages have a problem. I see HTML code in place of the rendered viewing. A GitHub bug? Yes, very strange Github issue; I solved it...

Hi @AdlerKosma , most likely the problem could be related to the process requesting more RAM than what is installed. Can you please also paste here the content of the...

Actually, the problem is related to a missing library: `libicui18n`. We have never ever encountered this error, are you running the software on a standard Ubuntu 18 installation? You can...

Hi @EtienneAb3d , I suggest to try this option of the `mmt create` command: ``` --count-threshold. all tokens with a count less than this threshold will be used only for...

Hi @EtienneAb3d , we kow this trick and we sometimes use it too. However this is more an hack than a proper solution, this is why: 1. Translation throughput is...

Hi @mzeidhassan , I'm sorry but there are no plans to port new features to the open-source in the near feature, including the port of the new ModernMT version. The...

Hi @EtienneAb3d For public parallel corpora I suggest the [Opus Website](http://opus.nlpl.eu/), but probably you already knows it and there is no large FR-JA corpus (node: JP is not the right...