Online machine learning with Vowpal Wabbit as Node.js streams.
- Stream Abstraction: Treat a Vowpal Wabbit learner process as a writable & readable stream. You write training or testing examples, and read predictions from the stream.
- Sparse Example Format: Represent your training or testing examples and observations naturally, as sparse JavaScript objects.
- Namespace Mapping: Use descriptive feature namespace names, which are mapped under-the-hood onto Vowpal Wabbit's preferred single-character namespace codes.
- Live Model Persistence: Retrieve the Vowpal Wabbit binary model ("regressor") during training or testing. You never need to shut-down the Vowpal Wabbit process!
- Child, Not Daemon: Uses a spawned child process for Vowpal Wabbit, for simpler interaction and IPC management.
From the Vowpal Wabbit tutorial:
var vw = new VowpalWabbitStream({ learningRate: 10 });
vw.on('data', function(obj) {
console.log("Prediction & Average Loss:", obj.ex, obj.pred, vw.getAverageLoss());
});
var exs = [
{ resp: 0, featMap: { price: 0.23, sqft: 0.25, age: 0.05, yr2006: 1.0 } },
{ resp: 1, imp: 2.0, featMap: { price: 0.18, sqft: 0.15, age: 0.35, yr1976: 1.0 } },
{ resp: 0, initPred: 0.5, featMap: { price: 0.53, sqft: 0.32, age: 0.87, yr1924: 1.0 } }
];
for (var pass=0; pass < 25; pass++) {
for (var i=0; i < exs.length; i++) {
vw.write(exs[i]);
}
}
vw.end();
/*
[20150604@15:30:24.893] DEBUG -- VW(STDERR): finished run
[20150604@15:30:24.893] DEBUG -- VW(STDERR): number of examples per pass = 75
...
[20150604@15:30:24.894] DEBUG -- VW(STDERR): average loss = 0.057188
*/
- To driffer85 for his earlier wrapper.