Rasa: The whole Core model is retrained even if only the templates section is changed

Created on 9 Aug 2019 · 12Comments · Source: RasaHQ/rasa

Description of Problem:
Having to retrain the Core ML model every time templates are changed is unnecessary and can be very expensive for a large project with lots of stories.

Overview of the Solution:
Create a separate, more granular fingerprint for the templates section in the domain and if only that has changed, simply replace that section in the new model and not retrain Core.

Examples (if relevant):

Blockers (if relevant):

Definition of Done:

[ ] Tests are added
[ ] Features described in the docs
[ ] Feature mentioned in the changelog
[ ] Any updates in Rasa X?

area type

Source

msamogh

👍1

Most helpful comment

This is the default behavior; the way to avoid retraining with template changes is to use an External CMS.

I don't think the choice of using an external CMS or not should penalize you with unnecessary training time. A change in the text or triggers of the template does not change the outcome of the core model training and is therefore unnecessary, don't you think?

nbeuchat on 15 Aug 2019

👍2

All 12 comments

This is the default behavior; the way to avoid retraining with template changes is to use an External CMS.

If there are no changes to the domain or the training data Rasa already prevents redundant training.

ortsaCniveK on 10 Aug 2019

This is the default behavior; the way to avoid retraining with template changes is to use an External CMS.

nbeuchat on 15 Aug 2019

👍2

@msamogh this should be in progress, shouldn't it?

wochinge on 19 Sep 2019

Yes. Thanks for noticing!

msamogh on 19 Sep 2019

@btotharye additionally to what it says in the issue description here, someone mentioned recently that sometimes NLU retrains if some parts of the domain are changed as well. So we should look into what's going on there