PredictionIO Classification Engine Template with MultiLayerNetwork from Deeplearning4j

Engine Quickstart

Overview

This engine template integrates the MultiLayerNetwork implementation from the Deeplearning4j library into PredictionIO. In this template, we use PredictionIO to classify the widely-known IRIS flower dataset by constructing a deep-belief net.

Importing the IRIS Flower Dataset

The IRIS flower dataset contains 3 classes of flowers with 50 samples of each, for a total of 150 samples. An in-depth description can be found here and a copy of the dataset can be found here. Each sample consists of four measurements (sepal length, sepal width, petal length, and petal width), as well as the class name (Iris-setosa, Iris-versicolor, and Iris-virginica).

A Python script has been provided to import the data to the event server using PredictionIO's Python SDK. Run the following command, replacing the access_key parameter with the value of your access key.

python data/import_eventserver.py --access_key 4H5MigBfIvatrPKHNMkZSgKma58wRyy0Sfly20bLYLTSEZbrZ0uodTEBzqZukbcT

The data is stored in the event server with five properties per record. The measurements sepal-length, sepal-width, petal-length, and petal-width are stored as floats, and the species is stored as a string.

Building, Training, and Deploying the Engine

Build the engine using the following command.

pio build

Train the model with the following command. Note that the commons-math3 library included with Spark is incompatible with the one required by Deeplearning4j, so we must specify that PredictionIO use the version located at lib-commons-math3/commons-math3-3.3.jar.

pio train -- --driver-class-path lib-commons-math3/commons-math3-3.3.jar

Deploy the engine with the following command.

pio deploy

Querying the Engine

A query to the engine is a JSON consisting of the four flower measurements. The following is a sample JSON query.

{'sepal_length': 5.1, 'sepal_width': 3.5, 'petal_length': 1.4, 'petal_width': 0.2}

A Python script using the Python SDK has been provided to demonstrate a sample query. Execute the script as follows.

python data/send_query.py

The model will return in a JSON response the predicted species based on the flower measurements in the query. The following is a sample JSON response.