search

BIDData / BIDMach

CPU and GPU-accelerated Machine Learning Library
BSD 3-Clause "New" or "Revised" License
916 stars 168 forks source link

DNN learnerX/dlayers won't run with sparse targets #48

Closed coryschillaci closed 8 years ago

coryschillaci commented 9 years ago 
val (nn,opts)=DNN.learnerX(loadSMat(indir+"trainData000.smat.lz4"),FMat(loadSMat(indir+"trainLabels000.smat.lz4")));

opts.aopts = opts;
opts.featType = 1;               // (1) feature type, 0=binary, 1=linear
opts.addConstFeat = false;        // add a constant feature (effectively adds a $\beta_0$ term to $X\beta$)
opts.batchSize=500;
opts.reg1weight = 0.0001;
opts.lrate = 0.2f;
opts.texp = 0.4f;
opts.npasses = 5;
opts.links = iones(132,1);

DNN.dlayers(3,100,0.25f,132,opts,2);

nn.train;

runs just fine, but if I don't cast the targets as an FMat, i.e., 

val (nn,opts)=DNN.learnerX(loadSMat(indir+"trainData000.smat.lz4"),loadSMat(indir+"trainLabels000.smat.lz4"));

opts.aopts = opts;
opts.featType = 1;               // (1) feature type, 0=binary, 1=linear
opts.addConstFeat = false;        // add a constant feature (effectively adds a $\beta_0$ term to $X\beta$)
opts.batchSize=500;
opts.reg1weight = 0.0001;
opts.lrate = 0.2f;
opts.texp = 0.4f;
opts.npasses = 5;
opts.links = iones(132,1);

DNN.dlayers(3,100,0.25f,132,opts,2);

nn.train;

gives

scala> nn.train;
pass= 0
scala.MatchError: (           1     0.59643           1     0.54870     0.96072           1           1     0.97798...
           1  0.00054444  4.6897e-12     0.93011     0.99481     0.98594  4.0819e-37     0.99795...
     0.99987  1.1189e-14   0.0054939     0.96049     0.96663  0.00032491           0    0.010586...
  1.2195e-06  0.00018742  3.1474e-09     0.28522     0.54711  5.9957e-10           0     0.92651...
  0.00010903           1           1     0.31402     0.10718    0.016666           1     0.68628...
           1  9.1819e-18     0.99999    0.032690     0.99763      1.0000           1     0.94152...
          ..          ..          ..          ..          ..          ..          ..          ..
,(  14,   0)   1
(   6,   1)   1
(  40,   2)   1
(  59,   3)   1
(  46,   4)   1
(   1,   5)   1
(  94,   6)   1
(  60,   7)   1
  ...  ...  ...
,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1) (of class scala.Tuple3)
  at BIDMach.models.GLM$.derivs(GLM.scala:475)
  at BIDMach.networks.DNN$GLMLayer.backward(DNN.scala:347)
  at BIDMach.networks.DNN$Layer.backward(DNN.scala:230)
  at BIDMach.networks.DNN.dobatch(DNN.scala:161)
  at BIDMach.models.Model.dobatchg(Model.scala:101)
  at BIDMach.Learner.retrain(Learner.scala:87)
  at BIDMach.Learner.train(Learner.scala:53)
  ... 33 elided
jcanny commented 8 years ago

Fixed now. You can do this for any classifier by applying the "full()" function to its target matrix, which is normally gmats(1).