This project aims to provide an example on how to use MLFLOW within rasa in order to track config.yaml updates impact on the bot performance.
It serves to automatically track multiple experiments of rasa train on a defined search space using the hyperopt methodology and identifies the best combination.
First you need to create and setup a virtual environment to install dependencies needed to run the code by running within your venv the following commands in the project root directory.
python3.8 -m venv .mlflow
source .mlflow/bin/activate
pip install --upgrade pip
pip install mlflow
Second, you need to update the files of the sub-project you want to execute by adding your own test_data.yml and training_data.yml and config.yml files.
You can generate a test/train split in rasa using the command rasa data split nlu from you rasa project root directory. Than copy/paste them in files.
config.yml should be added as a template as in the example, where you replace the values with the their name.
The code is organized as a MLFLOW Project. It contains a complete workflow that spans a search space in search for best parameters and leverages multithreading for faster execution.
You can use this project in different ways:
- By running from root directory:
mlflow run. <pass you own params if needed>. - If you don't want to clone this project, you can run this command :
- without default params:
mlflow run git@github.com:YahyaGrb/mlflow_rasa_hp.git - with your own params/files:
mlflow run git@github.com:YahyaGrb/mlflow_rasa_hp.git -P <your params if needed>
- without default params:
- From a jupyter notebook:
-
import mlflow project_uri = "https://github.com/YahyaGrb/mlops_rasa/mlflow_rasa_hp" params = your_params # (as needed by the project) mlflow.run(project_uri, parameter=params)
-
You can track the results in Mlflow by running mlflow ui and opening http://127.0.0.1:5000 from your browser.