How to train CRFEntityExtractor faster?

I have around 3000+ custom entities in my project. This exponentially increases the training data and rasa nlu trianing is taking upto two days especially the CRFEntityExtractor component, is there a way to multithread this component or improve training speed?

HI @pankti23, what version of rasa are you running? Have you tried playing around with any of the configuration values on the component or the featurizers?

1 Like

Hi @pankti23,

which pipeline components are you using currently?

If you are using spaCy, Iā€™d recommend to split the training process by re-training the underlying spaCy model with your custom entities and then let Rasa simply auto-fill them by specifying them in your config.yml.

Aside @erohmensing 's suggestion to play a little bit with the possible parameters, this at least would avoid retraining your model everytime this way.

Regards Julian

2 Likes

@erohmensing I am using the 1.4.3 version, and I did try tweaking the parameters however it still took a while. I have currently started using the tensorflow gpu for the training. Thank you, closing the issue for now. @JulianGerhard thank you, looks like that should work the next time