Training Data with supervised_embeddings is slow, How can i improve training

i am training my RASA NLU with supervised_embeddings, the speed is very slow, How can i improve training speed?

How slow is it, and how much NLU training data do you have?

i have almost 25 intents without any entities or any slots , utterances would be around 15000. it’s taking around 7-8 minutes for the same.Rasa version is 0.15

With 15000 utterances and 25 intents, 7 minutes is not that slow. You can improve the training time by tweaking the hyperparameters of supervised_embeddings, but your model will probably be less accurate as a result.

Thanks @Remy, can i do any incremental training. Because even after adding one intent or only few utterances , i am forced to re-train it for all

Not that I know of. You have to remember that if you add some utterances the model needs to relate those words to the other sentences it knows. It can’t do that without retraining everything because it has to adjust the word vectors for the -already trained- intents.

1 Like