Training rasa nlu produces different results on different machines

Rasa Open Source
1

Hi guys,

We are using rasa nlu v1.4.6 with the supervised_embeddings pipeline and random seed specified for the EmbeddingIntentClassifier. We have found that training on different machines is producing different models. All of our machines are running the same versions of RASA (and python). Each machine consistently produce the same model (we are using running “rasa test nlu” to measure the performance of the model) but the results differ between machines.

The machines we are using are running Fadora 31, Ubuntu 18.04, Manjaro 18 and Arch.

Any comments or insights are appreciated.

3

I came across the same issue while deploying RASA on OpenShift Container platform. My theory was that there weren’t enough resources for the container, and training underperformed as a consequence. To solve the issue I assigned a minimum of 4 GiB of Memory to the RASA container on the platform.

Hope this helps you.