Optimizing Rasa systems

For intent classification and entity extraction components based on machine learning, the learning rate and epoch settings are very important for the performance of this component. We can determine the adjustment direction (increase or decrease) of the learning rate and epoch by observing the loss curve.

In this example, we take the commonly used component in Rasa, DIETClassifier, as a way to illustrate how to adjust the settings to use TensorBoard, to visualize training and validation metrics. Here is an example of our Rasa configuration:

version: "2.0"
language: en
pipeline:
  - name: WhitespaceTokenizer
  - name: LanguageModelFeaturizer
    model_name: "bert"
    model_weights: "rasa/LaBSE"
  - name: RegexFeaturizer
  - name: DIETClassifier
    epochs: 100
    learning_rate: 0.001
    ...