criteolr.ssc

:silent
class xopts extends Learner.Options with FilesDS.Opts with GLM.Opts with ADAGrad.Opts with L1Regularizer.Opts
val opts = new xopts

val mdir = "d:/criteo/parts/"
opts.fnames = List(FilesDS.simpleEnum(mdir+"train%02d.smat.lz4", 1, 0),  
                   FilesDS.simpleEnum(mdir+"trainlabel%02d.fmat.lz4", 1, 0));
opts.nend = 67;
opts.lrate = 0.0001f;
opts.reg1weight = 0.001f;
opts.batchSize=2000;
opts.npasses=20;
//opts.npasses=1;
opts.eltsPerSample=60;
opts.links = icol(1);
//opts.pstep = 0.001f;
opts.lim = 8f;
opts.aopts = opts;

val ds = {
  implicit val ec = threadPool(4) // make sure there are enough threads (more than the lookahead count)
  new FilesDS(opts)               // the datasource
}


val nn = new Learner(                  // make a learner instance
    ds,                                // datasource
    new GLM(opts),                     // the model 
    Array(new L1Regularizer(opts)),    // list of mixins or regularizers
//    new ADAGrad(opts),                 // the optimization class to use
    null,
    opts)                              // pass the options to the learner as well


nn.train

val mm = FMat(nn.model.modelmats(0));

val filesize = 600000;
val tscores = dzeros(filesize*10, 1);
val tcats = dzeros(filesize*10, 1);
var len = 0;
for (i <- 0 until 10) {
  val a = loadSMat(mdir + "train%02d.smat.lz4" format i + 67);
  val c = loadFMat(mdir + "trainlabel%02d.fmat.lz4" format i + 67);
  val sc = mm * a;
  tscores(i * filesize -> (i * filesize + sc.length), 0) = DMat(sc.t);
  tcats(i * filesize -> (i * filesize + sc.length), 0) = DMat(c.t);
  len += c.length;
  print(".");
}
val scores = tscores(0->len,0);
val cats = tcats(0->len,0);

val rr = roc(scores, cats, 1-cats, 1000)

mean(rr)

val bounds = 7.0
val bscores = min(max(scores, drow(-bounds)), drow(bounds));
val probs = 1/(1 + exp(-bscores));
val lln = ((cats dot ln(probs)) + ((1-cats) dot ln(1-probs)))/probs.length

println("roc area = %5.4f, ll = %5.4f" format (mean(rr).v, lln.v))
:silent

// npasses  batchSize     lrate      l1reg       ll        time         AUC
//    2       1000        0.0003     0.0001    0.5176      3600
//   20      10000        0.0003     0.0001    0.5072      3732
//    5      10000        0.0003     0.0001    0.5242      941
//    5      10000        0.0003     0.00001   0.5241      936
//    5      10000        0.0003     0.001     0.5242      945
//    5      10000        0.0003     0.01      0.5245      944
//    5      10000        0.0003     0.1       0.5262      942
//    5      30000        0.0003     0.01      0.5364      364
//   15      30000        0.0003     0.01      0.5216      1096
//   50      30000        0.0003     0.01      0.5062      3614         0.7473
//  200      30000        0.0003     0.01      0.4922     14430         0.7592