This is a proof of concept to create a very simple component to serve scikit-learn machine learning models over HTTP. The idea is to use this component as a micro-service inside a more complex infrastructure.

It aims to provide a set of base classes, glue code and code patterns to make the deployment of these models easy and writing very few lines of code.

It's based on Flask because it's lightweight, easy to use and fairly well documented.

There are two base classes for models:

• ValuesModel: Expects a values input with all the model numeric inputs as comma separated values. Like values=1,3,7.2.
• KeyValueModel: Expects a list of key-value fields as input.

The following class-level attributes can be defined:

• name: Id of the model to be used in the API endpoint.
• pipeline: sklearn pipeline that accepts a 2d numpy array as input.
• schema: If using a KeyValueModel, a schema defining the POST fields expected as input.
• probability: if True, allow a /{model_id}/predict_proba/ endpoint to return a comma-separated list of probabilities to belong to each of the possible prediction classes. It is False by default.

Example:

class IrisModel(KeyValueModel):
    name = 'iris'
    probability = True
    pipeline = joblib.load('models/iris.pickle')
    schema = [
        {'name': 'sepal_length'},
        {'name': 'sepal_width'},
        {'name': 'petal_length'},
        {'name': 'petal_width'}
    ]

In this case, we are creating a model based on the pipeline loaded from a file, with the name iris that takes four inputs.

To serve this model, use the create_app function, accepting a list of model instances:

app = create_app([IrisModel()])

This will create a Flask app that serves the given model predictions under the /iris/predict/ endpoint, and /iris/predict_proba/ endpoints. If the input data provided is invalid, it will return 400 Bad Response. If it is valid, it will return the model predictions.

To serve this model, assuming you put your model in a iris.py file, run the following in your shell:

FLASK_APP=iris.py flask run

When using the KeyValueModel, the input is validated against a given schema definition. The schema is specified as a simple dictionary and it is based on the Cerberus validation library.

It takes the following parameters:

• 'name': field name to get from the input
• 'default': if present, will be used when no value is provided. Otherwise ValueError will be raised in this case.
• transform: callable accepting a single argument to transform the provided value

• Field validation failure return list of errors in 400 Response.
• Doctests.
• Travis.