Please retrain models to not depend on deprecated versions

[helpmefindaname]

Hello @redewiedergabe

as noted in flairNLP/flair#3382, some (or all?) models depend on deprecated embeddings and cannot be loaded in newer versions.

The issue is not only that it was trained on an older flair version, but also that it used the pytorch_transformers library that was deprecated in favour of the transformers library.
I tried to use some engineering tricks, to map the pytorch_transformers classes to the transformers objects, however I then ended up having a mismatch in the vocab size.

I suppose the easiest fix would be to retrain the models and host the new versions on your website.

[redewiedergabe]

Thank your very much for bringing this issue to our attention. We will be working on it and will update this thread when we are done.

[christofs]

I ran into the same problem just now. An updated model would be much appreciated!

[gnmarten]

haven't tested, but I am under the impression this pipeline is more compatible with current environments: https://github.com/cophi-wue/LLpro referencing https://huggingface.co/aehrm/redewiedergabe-direct