runtime-prediction.md

layout	title
default	Runtime Prediction Algorithms

General Prediction Steps

setup server --> events --> feature extraction pipeline --> runtime scorer --> microservice scorer --> predictions

Runtime Prediction Algorithms

At runtime Seldon can score requests against prediction models. At present simple classification based Vowpal Wabbit models are supported inside the Seldon Server while the full range of Vowpal Wabbit models can be utilized via a micro-service API. The available runtime prediction algorithms are:

Runtime Prediction Algorithm	Model
vwClassifier	Vowpal Wabbit binary or one-against-all model
externalPredictionServer	Custom Model

These algorithms will be utilized when the prediction endpoints of the REST or Javascript API are used.

Vowpal Wabbit Classification

A runtime scorer for Vowpal Wabbit classification models. At present only a subset of VW models are handled, specifically the model needs to be:

• Saved as a readable model
• Can be only a binary classification model or "one-against-all", i.e., the only value of the "options:" line in the readable model can be an "--oaa" option.
• Does not use quadratic, cubic, ngram or other feature extensions of VW

If these requirements are too restrictive one can use an external prediction server accessed over the microservices REST API as described in pluggable prediction algorithms.

Algorithm : vwClassifier
Description : Utilizes a Vowpal Wabbit classification model

Example config to set in /all_clients/[client]/predict_algs:

{% highlight json %} { "algorithms":[ { "name":"vwClassifier", "config":[] } ], } {% endhighlight %}

The location of the readable model should be specified at /all_clients/[client]/vw. The vwClassifier will attempt to load a file model from this location. It will also attempt to load an optional file classes.txt which gives the mapping of target class ids to a human friendly name, for example:

{% highlight bash %} 1,Iris-setosa 2,Iris-versicolor 3,Iris-virginica {% endhighlight %}