jmsquare / Giskard-deployment

Using Giskard's test in Github actions to deploy an AI model.

Giskard-deployment

Example of how to deploy the credit scoring once it passes Giskard's tests.

Workflow

The deployment happens according to three steps:

• Training: execution of train.py script every time a push occurs.
• Testing: execution of Giskard_test.py script after the training is done. During this step, the model produced by train.py is uploaded to the Giskard UI (where the user can define his custom test suite). Then all the tests are performed. If at least 50% of the tests pass (this can be tuned here in Giskard_test.py), the model is tagged as verified.
• Deployment: If the model is verified, it is committed to the repo under the folder deployed_model that contains:
  • a subfolder named as the date of the model training which itself contains:
    • logistic_regression_model.pkl: a pickle of the verified model.
    • test_data.zip: a pandas dataframe that was used to verify the model.
    • Giskard_tests.json: a summary of all the tests and metrics that were run by Giskard.