Trained model wrapper module structure

In order to make predictions using the trained ML model, the framework neds to have a Python mini-program capable of doing the actual work.

The name of the module should be stated in the settings.py file using the TRAINED_MODEL_MODULE_NAME entry. In the absence of this setting, the framework will locate the first file with .py extension in this directory (whose name isn't __init__.py or wrapper.py) and use it as the trained model module.

This module should implement the following functions (with mypy type hints!):

def init() -> None:

This method should deserialise the pickled model file that was trained by a data scientist using the ML toolkit (e.g. Scikit-learn, LightGBM, TensorFlow, Keras), plus any other files needed to transform the input data and return a prediction. This method is called by the framework during the initialisation phase.

def run(input_data: Iterable) -> Iterable:

This function does all the ML prediction heavy lifting. It will receive either a Python Iterable (most likely a dictionary) or a Pandas DataFrame as input, do all the necessary transformations, call the trained model's predict() method and return the result.

def sample() -> Dict:

This method should return a Python dictionary with sample values for each of the parameters required by the model. This is used in two contexts:

1. Initialisation of the Swagger documentation, listing the JSON parameters and their sample values
2. Validation of the JSON input when the model/predict method is invoked

The eagled-eyed reader will have noticed that the methods init() and run() are the same two that Microsoft prescribe when deploying a trained ML model to Azure Machine Learning. We just added some type hints, plus sample() to support more advanced functionality:

• Swagger documentation
• JSON input validation

An example file is provided: ml_trained_model.py