Problems with deploying Rasa X on Server (with German bert-base embeddings model)

Hey there,

I have been facing an issue with deploying Rasa X on a server these past days. I am following the Quick Installation described on the Deploy to Server tutorial on a Google Cloud VM (using Ubuntu 16.04). Based on previously posted issues, I have used the installation for this (pinned version).

I am able to connect to the Rasa X UI through the provided public IP address and can upload models using the curl -k -F "model=@my_model.tar.gz" command.

After uploading the model and making it active, I can however not talk to my bot as it is not reypling and appears to be buffering on the response. Neither the “Mode: Talk”, “Mode: Interactive” or sharing with testers is working. I know that there appears to be an issue with Rasa X only showing responses only after refreshing the page (i.e. from this issue #3936). This is NOT the case with my bot. After refreshing the page, it completely restarts and my previous messages disappear, while no replies from the bot show up.

I have hypothesized that it might be a problem with the language model. The model is trained based on the recently released German bert-base embeddings (de_pytt_bertbasecased_lg). Is the language model “contained” in the trained model? Or is it necessary to previously download and link the language model, on which the model was trained, before being able to run the bot on the Rasa X server? If so, where would I need to download it to (modify a specific rasa docker-image or another solution?)?

Any help would be greatly appreciated as I’m quite new to this and have no experience with docker.

Thanks so much!

Hi @celinakirschner! A good way to see what’s going on here is by checking the docker logs of rasa-production, which handles conversations with the user. You can check by

cd /etc/rasa
sudo docker logs rasa-production

If you post the error message it’s getting here, we can help you debug it.

Hi @celinakirschner,

the transformer-based model is installed as a python library since in your architecture, most likely only the features are used to train the Rasa model. This results in the necessity to have the former de_pytt_bertbasecased_lg, now de_trf_bertbasecased_lg installed on the target system/container. Please keep in mind that those models can’t be linked, they won’t show up in e.g. python -m spacy validate hence can be used.

There are two “quick” ways to post-install the model. Since the Rasa-containers are normal Docker-containers, you can:

a. derive your own container from them in which you run your installation command b. execute the download command within the docker container by attaching yourself to its terminal

Tell me if you need help with that!

Kind regards Julian

1 Like