What is the best Rasa version to use for using bert models in the pipeline?
Since I check the Rasa github repo and others project and it seems there are some components are deprecated and some codes are changed.
Hi, can you specify what’s deprecated? If you can point us to something that seems out of date or stale, that would be very helpful, then we can make sure we update it!
You can use any of the more recent Rasa versions (1.8 and above, though I would strongly recommend using more recent versions). The difference is that before 2.1 you had to use:
- HFTransformersNLP
- LanguageModelTokenizer
- LanguageModelFeaturizer
If you use 2.1 or above, LanguageModelTokenizer and HFTransformersNLP are deprecated. Instead, you should use just any Tokenizer (for example, WhitespaceTokenizer) paired with LanguageModelFeaturizer in the pipeline:
- WhitespaceTokenizer
- LanguageModelFeaturizer
Does that make sense? Note that there is a slight increase in inference time for the LanguageModelFeaturizer component in versions 2.8.9 and above: this is related to the TensorFlow upgrade.
1 Like
Thanks for this great answer, it is very useful.
By deprecated I mean HFTransformersNLP & LanguageModelTokenizer components.
1 Like
Ah I see, well I’m glad we could get you sorted! 
1 Like