fbarrios / turtle Goto Github PK

# turtle - load readable_model from vowpal wabbit and serve it from java

## example train a model

$ vw -d /path/to/dataset/dataset --readable_model readable_model.txt

##  readable_model.txt looks like:
...
6163:0.0624969
7472:-0.12023
42847:-0.0421919
49960:-0.12023
51692:0.0624969
55432:-0.0421919
...

using https://gist.github.com/luoq/b4c374b5cbabe3ae76ffacdac22750af,
and some logic to produce the same buckets for the same feature combinations

## java code
add the dependency in pom.xml
<dependency>
    <groupId>bz.turtle</groupId>
    <artifactId>readable</artifactId>
    <version>2.0.4-BETA</version>
</dependency>

import bz.turtle.readable.*;

ReadableModel m = new ReadableModel("directory/of/readable_model.txt"); // .txt.gz also works


// this will load the model creating only float[] array with same size as 2**bits(-b from vw)
// it uses same algorithm to compute hash buckets as vw and then returns the inner product

m.predict(
          new Doc(
              new Namespace(
                  "your-namespace",
                  new Feature("a",1.0),
                  new Feature("b",1.0),
                  new Feature("c",1.0))));
// value 1.0 is default

## make sure it works
if you want to make sure your parameters are supported, in the repeatable_model.txt
add test.txt and predictions.txt (using -p from vw) and it will automatically
test if turtle gets the same predictions as the wabbit.

if you pass file to ReadableModel it will just try to load the file


check out resources/test for more examples


## known bugs

* -q ::
  the way vw does permutation is different, and it produces different results

## thanks

* Lucas Bernardi
* Tarek Sheasha
* Denis Bilenko

wouldn't have happened without the help of those guys

## todo

* more tests
* verify the checksum of the model
* support for --cubic
* support for ngrams and skips
* support for --lrq