Lightweight Weka server framework that can be deployed in a container and accessed as a micro-service via web-services endpoints. Currently supports serving predictions from Weka models (classifiers and clusterers).

Install any required Weka packages before executing. E.g.:

java -cp lib/weka.jar weka.core.WekaPackageManager -install-package SMOTE

Expects a wekaServer_<task_id>.props file to be present in ${user.home}/config. This file specifies parameters for the task to be executed - e.g. such as the task type, model file to load (from ${user.home}/models), any specific data preparer class to use, and various other options relating to input. See the iris example in config.

The default data peraparation class for model scoring expects incoming data to score in the JSON pandas-split format. This is converted directly to a Weka Instances object and passed to the model for prediction. Predictions are output/returned as JSON-formatted pandas-split.

To run a test from the command line:

1. copy/move config and models to ${user.home}
2. mvn clean compile assembly:single
3. java -cp target/weka-server-1.0-SNAPSHOT-jar-with-dependencies.jar weka.server.scorer.Test irisClassifier input_data/iris.arff

This loads the ARFF file, converts it to JSON and then invokes the WekaScoringModel; which, in-turn, loads the j48_iris.model from the models directory, converts the JSON back Weka Instances and passes it to the J48 model for prediction. Resulting probabilty distributions are returned in JSON pandas-split format.

To build:

``` sh
mvn clean compile assembly:single
```

Docker container build (assumes maven build complete):

``` sh
docker build -t server:test .
docker run -it -p 7000:7000 server:test
curl http://localhost:7000
```

Send iris data for scoring:

 ``` sh
 curl -X POST -H "Content-Type: application/json" -d @input_data/iris.json http://localhost:7000/invocations?taskid=irisClassifier
 ```