florishoogenboom/restfull-scikit repository overview

⁠This project

This project is composed of two main parts:

• The source code for the RESTfull API
• A Dockerfile that wraps this code in an image that runs it using NGINX uWSGI.

The goal of this project is to provide a quick an simple way to deploy a Scikit-learn model to production

⁠Usage

To use this container two objects are needed (1) an fitted Scikit-learn model instance (2) a serializer (preferabbly based on marshmallow⁠, for else correct handling of errors is not guaranteed) that serializes a Python datastructure to something that can be used as input for your model. The model should be stored as a pickled file in the /objects folder while the serializer should be stored as a Python script in the /modules folder. Pickling of the Scikit-learn model should be done using the joblib library included in sklean.externals.

⁠1. The model and the schema

The model should be a fitted classifier that implements the Scikit-learn API (fit, transform etc.). This model should be stored as a pickeld file named model in the folder /objects. Note that you need to use the version of the joblib library included wich Scikit-learn to correctly pickle a model.

The schema should be an subclass of marshmallow.Schema and should be named ModelSchema. You can use the marshmallow.post_load decorator to describe how data should be deserialized into a format that is accepted by your model. For example:

from marshmallow import Schema, fields, post_load

class ModelSchema(Schema):
    param1 = fields.Str(required=True)
    param2 = fields.Int(required=True)

    @post_load
    def to_numpy_array(self, data):
        # your code here to transform the Dict data into 
        # e.g. a NumPy array
        return # Your desired datastructure.

⁠2. Building the Docker image

In the folder you have saved your schema and model files, create a file named Dockerfile with the following contents

FROM florishoogenboom/restfull-scikit-docker:latest

RUN {Install your dependencies here}

COPY ./model /data/model
COPY ./schema /data/schema

You can now build this image using docker build -t myname/my-model . . Note that you need to install the dependencies your model requires. This since unpickling requires the modules your model uses to be available.

⁠3. Running the image

To run this image simply use docker run -p 80:80 myname/my-model this binds the model to port 80 locally. Of course you can use Docker swam to spin up multiple instances and make use of Dockers load balancing features.