Travis CI provides free builds for open source projects.

The .travis.yml could look like this.

language: scala
scala:
   - 2.10.2
   - 2.10.3
   - 2.10.4
   - 2.11.0

But first you have to add the project. Take a look at the Getting Started Guide

It's currently possible to serialize a liblinear model, but not a classifier that wraps around it (with associated featurizer, etc).

The current set of classes are spread around in different sub-packages w/o particular logic. Fix this. Simplify interfaces to models to make room for models trained using Liblinear.

We've battle tested v1.2.1-SNAPSHOT reasonably well. Can we cut a version before incorporating new changes? Cheers 🍻

https://github.com/utcompling/Scalabha/blob/master/src/main/scala/opennlp/scalabha/cluster/Kmeans.scala

As the title above, because I never used scala, and I just running java task on spark cluster, I want using nak's DBSCAN in my java task, thank you!

[info] Resolving org.scalanlp#nak;1.2.1-SNAPSHOT ...
[warn] module not found: org.scalanlp#nak;1.2.1-SNAPSHOT
[warn] ==== local: tried
[warn] /Users/scotthendrickson/.ivy2/local/org.scalanlp/nak/1.2.1-SNAPSHOT/ivys/ivy.xml
[warn] ==== opennlp sourceforge repo: tried
[warn] http://opennlp.sourceforge.net/maven2/org/scalanlp/nak/1.2.1-SNAPSHOT/nak-1.2.1-SNAPSHOT.pom
[warn] ==== Sonatype Snapshots: tried
[warn] https://oss.sonatype.org/content/repositories/snapshots/org/scalanlp/nak/1.2.1-SNAPSHOT/nak-1.2.1-SNAPSHOT.pom
[warn] ==== Sonatype Releases: tried
[warn] https://oss.sonatype.org/content/repositories/releases/org/scalanlp/nak/1.2.1-SNAPSHOT/nak-1.2.1-SNAPSHOT.pom
[warn] ==== public: tried
[warn] http://repo1.maven.org/maven2/org/scalanlp/nak/1.2.1-SNAPSHOT/nak-1.2.1-SNAPSHOT.pom
[info] Resolving gov.nist.math#jama;1.0.2 ...
[warn] ::::::::::::::::::::::::::::::::::::::::::::::
[warn] :: UNRESOLVED DEPENDENCIES ::
[warn] ::::::::::::::::::::::::::::::::::::::::::::::
[warn] :: org.scalanlp#nak;1.2.1-SNAPSHOT: not found
[warn] ::::::::::::::::::::::::::::::::::::::::::::::
sbt.ResolveException: unresolved dependency: org.scalanlp#nak;1.2.1-SNAPSHOT: not found

with maven, i import nak jar V1.2.1, but i cannot import nak.cluster.GDBSCAN._! How come?

The PCA code for dimensionality reduction for clustering requires using pca_transform-0.7.2.jar, which must be carried along as an explicit dependency. It would be great to have an implementation in Nak, and adding that would be a great student project.

Ensure that everything is working as the move toward getting the old opennlp maxent code out progresses.

Make it possible for liblinear models to be used as a drop-in replacement for the current models.

http://mvnrepository.com/artifact/org.scalanlp/nak/1.1.3

The package nak.regress is missing from the jar.

is it a intentional and moved somewhere else or a mistake ?

I see this is EOL'd.

These command-line parameters are not currently implemented, for usage see

https://github.com/utcompling/applied-nlp/wiki/Homework4

Hi,

can somebody please tell me if serializing nak.classify._ objects DataSerialization is even possible?

Also what is the state of the nak.classify._ namespace. It looks like there's a move toward using liblinear. Is the stuff pulled in from breeze-learn deprecated?

Thanks,
Richard

Hi,

I am a graduate student very new to NLP. I really enjoy the convenience of Scala so I decide to try out Breeze (and turns out it shared out it's NLP package). I need to use classifiers so I picked up Nak :)

The documentation is good and APIs are great. However, after spending 3 hours writing and reading the document (plus two examples), I can't help but notice they are both for LibLinear classifier. However I really want to try out NaiveBayes, Maximum Entropy (Logistic classifier), and Perceptron (well, I need to do my assignments as students right (: The NakContext has many utility functions like "trainClassifier" or "trainModel" that help building up classifiers but they are again, for Liblinear.

So what should I do if I want to use the three classifiers I mentioned above? I tried to just new them, but certainly I can't put unlabeled and unfeatured raw example in. Is there a way to use them??

I also posted this message on Google Group (at least the first half is the same..)

https://groups.google.com/forum/?fromgroups#!topic/scalanlp-discuss/KeFSy841NIE

But not consistently.

[info] - K-means SparseVector test *** FAILED ***
[info]   org.scalatest.exceptions.TestFailedException was thrown. (KMeansTest.scala:65)

Make sure the SNAPSHOT works with Chalk.

Use liblinear estimation instead, and make the API more coherent.

(David: I think clusters are getting emptied out sometimes--which I guess can happen. We need to decide what to do when that happens. Either just drop the cluster, or split another cluster in half.)

While performing K-means clustering, I got the following AssertionError:

Exception in thread "main" java.lang.AssertionError: assertion failed
    at scala.Predef$.assert(Predef.scala:165)
    at nak.cluster.Kmeans$$anonfun$5.apply(Kmeans.scala:112)
    at nak.cluster.Kmeans$$anonfun$5.apply(Kmeans.scala:108)
        ...

The assertion is at Kmeans.scala:112 but has to do with not finding a min in a list of doubles at Kmeans.scala:111. I suspect it's because sometimes double precision numbers don't behave correctly with == via a rounding error.

The code in breeze.learn should move to Nak, as described in the roadmap:

https://github.com/scalanlp/breeze/wiki/ScalaNLP-Roadmap

Nak is behind the now fast moving Breeze, need to catch up!

Is there a way to use Haversine distance instead of the euclidean distance in the DBSCAN algorithm?

Hi, in GDBScan, the last line of the expand method says "if new point is a cluster"... but in the "if clause" the point is evaluated with the current neighbourhood... not with the new neighbours.

    neighbours.foldLeft(neighbours) {
      case (neighbourhood, neighbour @ Point(row)) =>
        // if not visited yet, create a new neighbourhood
        val newNeighbours = if (!(visited contains neighbour)) {
          visited add neighbour
          getNeighbours(neighbour, points.filterNot(_.row == neighbour.row))
        } else {
          Seq.empty
        }
        // Add to cluster if neighbour point isn't assigned to a  cluster yet
        if (!(clustered contains neighbour)) {
          cluster add neighbour
          clustered add neighbour
        }
        // if the neighbour point is a cluster, join the neighbourhood
        if (isCorePoint(neighbour, neighbourhood)) neighbourhood ++ newNeighbours else neighbourhood
    }

I think the last line should be:

 // if the neighbour point is a cluster, join the neighbourhood
   if (isCorePoint(neighbour, newNeighbours)) neighbourhood ++ newNeighbours else neighbourhood

If you agree I can do a PR.

The current k-means implementation is something I did for homework assignments for teaching NLP courses at UT Austin. It can handle a fair amount, but it runs out of steam (in particular, memory) for larger datasets, especially if they have a lot of features. It currently uses dense vectors to represent the features for each data point, so it should be a fairly straightforward win to change this to use sparse vectors instead.