Guillaume Klein
Guillaume Klein
@plonibarploni Can you first try a clean reinstallation of the package? ``` pip install --ignore-installed --no-cache-dir argostranslate ```
Hi, The penalty is currently applied on repetitions of the same token, so it will not change the output of the example above. The penalty is applied when the value...
I think we can extend the penalty to be applied on all previously generated tokens. In that case it may help for the example shown above. I'll make this improvement...
Can you try again with CTranslate2 2.8.0? Now the penalty is applied to the complete history as described in the [source paper](https://arxiv.org/abs/1909.05858). You should start with `repetition_penalty=1` (no penalty) and...
Hi, Regarding memory, did you [release the model](http://opennmt.net/OpenNMT/translation/inference/) for inference? This could save quite a bit on disk and then memory. Regarding concurrency, `translation_server.lua` is very simple and acts more...
How large is the training dataset you are using? How did you prepare/tokenize it?
Thanks! @jsenellart assigning to you as you introduced the hook mechanism and may have some comments on the usage.
As far as I know, we did not see any gains.
When loading a class, the Torch loader expects it to be available. I would recommend to simply copy the release script in `nqg/paragraph` and run it from there to ensure...
How much GPU memory is available?