clj-vw

A Clojure client and wrapper for vowpal wabbit, a fast out-of-core learning system sponsored by Microsoft Research and (previously) Yahoo! Research.

Artifacts

clj-vw artifacts are released to Clojars.

If you are using Maven, add the following repository definition to your pom.xml:

<repository>
  <id>clojars.org</id>
  <url>http://clojars.org/repo</url>
</repository>

The Most Recent Release

With Leiningen:

[engagor/clj-vw "1.2.0"]

With Maven:

<dependency>
  <groupId>engagor</groupId>
  <artifactId>clj-vw</artifactId>
  <version>1.2.0</version>
</dependency>

Usage and documentation

Except when only using client related code for connecting to a remote vw daemon (see online.clj), this library requires that vowpal wabbit is installed locally. Basic knowledge of vowpal wabbit, its command line options and input format are recommended. See the vowpal wabbit tutorial for more information.

Codox documentation in html format can be generated by issuing the command lein doc. Documentation files will be put under .../clj-vw/doc/. There are three namespaces.

clj-vw.core

Generic core functionality for interacting with vowpal wabbit (input example formatting, writing data files, passing options and calling vw, ...)

All functions work on a settings argument, which is basically a map containing vowpal wabbit options, training examples, etc. Options are set by calling set-option. Finally, vowpal wabbit is called by calling the function vw on prepared settings.

Example:

(-> (set-option :data "foo/bar.dat"
                :final-regressor "foo/bar.model"
                :save-resume true
                :ngram 3
                :quadratic "ab"
                :learning-rate 0.3)
    (add-example {:label 1, :features ["a"]}
                 {:label -1, :features ["b"]})
    (write-data-file)
    (vw)
    (read-predictions))

Overview of available functions:

• (available-options)

  Print and return the set of available options (plus basic info).

• (set-option settings key val & more)

  Set one or more options. Can be chained, e.g.

  (def settings (-> {}
                    (set-option :data "foo/bar.dat")
                    (set-option :save-resume true)
                    (set-option :ngram 3)
                    (set-option :quadratic "ab")
                    (set-option :quadratic "ac")
                    (set-option :learning-rate 0.3)))
  

• (get-option settings key)

  Return option for key in settings.

• (remove-option settings key & more)

  Remove option key from settings.

• (add-example settings example & more)

  Add one ore more examples to settings.

• (format-example {:keys [label labels importance tag features])

  Turns an example given as a map into vowpal wabbit's string format (see https://github.com/JohnLangford/vowpal_wabbit/wiki/Input-format).

  (format-example {:label -1.231
                   :features ["f1"
                              {:name "f2" :value 3.5}
                              {:name "f3" :namespace "ns3"}]})
  => "-1.231 | f1 f2:3.5 |ns3:1.0 f3"
  

  See clj-vw/core_test.clj suite for more examples.

• (read-predictions settings)

  Read vowpal wabbit prediction file (get-option settings :predictions).

• (write-data-file settings & {:as writer-settings})

  Writes (:data settings), a collection of examples, to (get-option settings :data).

• (vw-command settings)

  Returns the vw command as a string as defined by settings. Useful for inspecting the command that will be issued when calling vw on settings.

• (vw settings)

  Calls vw as specified by settings. Puts output in settings under :output and returns updated settings.

  More info in clj-vw/core_test.clj.

clj-vw.offline

Higher level helper functions for interfacing to a local vowpal wabbit install.

• (predict settings)

  Use an existing vowpal wabbit model (as specified by (get-option settings :initial-regressor) or (get-option settings :final-regressor), in that order) to compute predictions for examples in a data file (as specified by (get-option settings :data)) or in memory (as specified by (:data settings)).

• (train settings)

  Train a vowpal wabbit model from a data file (as specified by (get-option settings :data)) or from a collection of in memory examples (as specified by (:data settings)).

clj-vw.online

Higher level helper functions for launching and connecting to a (local or remote) vowpal wabbit running in daemon mode.

• (daemon & {:as settings})

  Start a vw daemon. Port (and any other options) can be set via vw options, e.g. (daemon (set-option :port 8003)).

• (connect settings)

  Connect to a vw daemon.

  Host is determined as the value of either (get-in settings [:client :host]), (get-in settings [:daemon :host]) or "localhost", in this order.

  Port is determined as the value of either (get-in settings [:client :port]), (get-in settings [:daemon :port]), (get-option settings :port) or 26542, in this order.

  Example, to start a local daemon on port 8003 and connect to it, do:

  (-> (set-option :port 8003) 
      (daemon) 
      (connect))
  

  As of version 1.1.0, it is possible to specify a connection timeout (in milliseconds):

  (-> {}
      (set-connection-timeout 1000) 
      (connect))
  

  The default timeout is 1000. On timeout a java.net.SocketTimeoutException is thrown.

• (train settings)

  Send examples to a vw daemon connection (as returned by connect) for training. Examples in the return settings (under :data) are extended with a :prediction slot corresponding to vowpal wabbit's prediction before training.

• (save settings)

  Save daemon's model to (get-opt settings :final-regressor).

• (predict settings)

  Send examples to a vw daemon connection for prediction (without training). Predictions are put under :predictions in settings.

• (close settings)

  Close daemon and/or client and cleanup.

Testing

On the command line, issue lein test. Requires that vowpal wabbit is installed and that the directory "/tmp/" is read/writable.

License

Copyright © 2014 Engagor

Distributed under the BSD Clause-2 License as distributed in the file LICENSE at the root of this repository.