Different predictions locally and in docker container

Hi all,

we are experiencing a very weird behaviour with a specific configuration in rasa. When using spacy locally all the intents are classified correctly and everything is fine. However, when deploying to a docker container (with image rasa/rasa-1.6.1-spacy-de) all the intents are messed up or predicted with such a low confidence that the conversation keeps ending up in the fallback case.

I personally would not expect any difference here.

config.yml (625 Bytes)

Has anybody experienced something similar?

Thanks in advance Best regards, Jes

Can you please share the logs of the docker container when doing the predictions (and also when the model is loaded?)

What version of RASA are you using locally?

I see you are using CountVectorFeaturizer with “char_wb” option. Might want to take a look at this: Loss of confidence in Rasa > 1.6.0 nlu (compared to 1.4.6) · Issue #5004 · RasaHQ/rasa · GitHub (still unfortunately unsolved at the moment).

1 Like

Problem solved. I used spaCy version 2.1.x on the docker and 2.2.x locally. Turns out the two versions are not compatible, or rather the models are incompatible. After downgrading my local spaCy version predictions were consistent again