Rasa 2.8 Training takes too much time


I am using Rasa 2.8.8 version

My training files are composed of 90 000 examples, 70 intents and 20 entities

I’m training only the NLU part

When I was doing my training, I had an OOM error. So I reduced the batch size but the training now takes more than 10 hours. I tried to run it on the GPU but the training time doesn’t reduce.

Do you have any ideas on how to achieve a shorter training time?

Here is my pipeline:

language: "fr"

- name: WhitespaceTokenizer
- name: LexicalSyntacticFeaturizer
- name: CountVectorsFeaturizer
    analyzer: "word"
- name: DIETClassifier
    batch_size: 4
    epochs: 50
- name: EntitySynonymMapper
- name: ResponseSelector
    epochs: 100

Salut Quentin!

You really have a LOT of examples, for an average of ~1300 examples per intent!

I would suggest having a maximum of 100 examples per intent, if not even 20.