@Override
public String toJavaString() {
double[] ary = expand();
if (ary == null || ary.length == 0) return "\"null\"";
SB sb = new SB().p('{');
for (int i = 0; i < ary.length - 1; ++i) sb.p(ary[i]).p(',');
return sb.p('}').toString();
}
@Override
protected void toJavaUnifyPreds(SB bodySb) {
if (isClassifier()) {
bodySb.i().p("float sum = 0;").nl();
bodySb.i().p("for(int i=1; i<preds.length; i++) sum += preds[i];").nl();
bodySb.i().p("if (sum>0) for(int i=1; i<preds.length; i++) preds[i] /= sum;").nl();
} else bodySb.i().p("preds[1] = preds[1]/NTREES;").nl();
}
@Override
protected SB toJavaInit(SB sb, SB fileContextSB) {
sb = super.toJavaInit(sb, fileContextSB);
sb.ip("public boolean isSupervised() { return " + isSupervised() + "; }").nl();
sb.ip("public int nfeatures() { return " + _output.nfeatures() + "; }").nl();
sb.ip("public int nclasses() { return " + _parms._k + "; }").nl();
if (_output._nnums > 0) {
JCodeGen.toStaticVar(
sb,
"NORMMUL",
_output._normMul,
"Standardization/Normalization scaling factor for numerical variables.");
JCodeGen.toStaticVar(
sb,
"NORMSUB",
_output._normSub,
"Standardization/Normalization offset for numerical variables.");
}
JCodeGen.toStaticVar(sb, "CATOFFS", _output._catOffsets, "Categorical column offsets.");
JCodeGen.toStaticVar(sb, "PERMUTE", _output._permutation, "Permutation index vector.");
JCodeGen.toStaticVar(sb, "EIGVECS", _output._eigenvectors_raw, "Eigenvector matrix.");
return sb;
}
@Override
public String str() {
SB sb = new SB().p('[');
for (int i = 0; i < _bases.length; i++) {
sb.p(_bases[i]);
if (_cnts[i] != 1) {
sb.p(':').p(_bases[i] + _cnts[i] * _strides[i]);
if (_strides[i] != 1 || ((long) _bases[i]) != _bases[i]) sb.p(':').p(_strides[i]);
}
if (i < _bases.length - 1) sb.p(',');
}
return sb.p(']').toString();
}
@Override
protected void toJavaUnifyPreds(SB bodyCtxSB) {
if (isClassifier()) {
bodyCtxSB
.i()
.p(
"// Compute Probabilities for classifier (scale via http://www.hongliangjie.com/2011/01/07/logsum/)")
.nl();
bodyCtxSB.i().p("float dsum = 0, maxval = Float.NEGATIVE_INFINITY;").nl();
if (nclasses() == 2) {
bodyCtxSB.i().p("preds[2] = -preds[1];").nl();
}
bodyCtxSB
.i()
.p("for(int i=1; i<preds.length; i++) maxval = Math.max(maxval, preds[i]);")
.nl();
bodyCtxSB
.i()
.p(
"for(int i=1; i<preds.length; i++) dsum += (preds[i]=(float) Math.exp(preds[i] - maxval));")
.nl();
bodyCtxSB.i().p("for(int i=1; i<preds.length; i++) preds[i] = preds[i] / dsum;").nl();
}
}
@Override
public String str() {
SB sb = new SB().p('(');
for (AST ast : _asts) sb.p(ast.toString()).p(' ');
return sb.p(')').toString();
}
@Override
protected void toJavaPredictBody(final SB bodySb, final SB classCtxSb, final SB fileCtxSb) {
SB model = new SB();
bodySb.i().p("java.util.Arrays.fill(preds,0);").nl();
final int cats = _output._ncats;
final int nums = _output._nnums;
bodySb.i().p("final int nstart = CATOFFS[CATOFFS.length-1];").nl();
bodySb.i().p("for(int i = 0; i < ").p(_parms._k).p("; i++) {").nl();
// Categorical columns
bodySb.i(1).p("for(int j = 0; j < ").p(cats).p("; j++) {").nl();
bodySb.i(2).p("double d = data[PERMUTE[j]];").nl();
bodySb.i(2).p("if(Double.isNaN(d)) continue;").nl();
bodySb.i(2).p("int last = CATOFFS[j+1]-CATOFFS[j]-1;").nl();
bodySb.i(2).p("int c = (int)d").p(_parms._use_all_factor_levels ? ";" : "-1;").nl();
bodySb.i(2).p("if(c < 0 || c > last) continue;").nl();
bodySb.i(2).p("preds[i] += EIGVECS[CATOFFS[j]+c][i];").nl();
bodySb.i(1).p("}").nl();
// Numeric columns
bodySb.i(1).p("for(int j = 0; j < ").p(nums).p("; j++) {").nl();
bodySb
.i(2)
.p(
"preds[i] += (data[PERMUTE[j"
+ (cats > 0 ? "+" + cats : "")
+ "]]-NORMSUB[j])*NORMMUL[j]*EIGVECS[j"
+ (cats > 0 ? "+ nstart" : "")
+ "][i];")
.nl();
bodySb.i(1).p("}").nl();
bodySb.i().p("}").nl();
fileCtxSb.p(model);
}
// Note: POJO scoring code doesn't support per-row offsets (the scoring API would need to be
// changed to pass in offsets)
@Override
protected void toJavaUnifyPreds(SB body, SB file) {
// Preds are filled in from the trees, but need to be adjusted according to
// the loss function.
if (_parms._distribution == Distributions.Family.bernoulli) {
body.ip("preds[2] = preds[1] + ").p(_output._init_f).p(";").nl();
body.ip("preds[2] = " + _parms._distribution.linkInvString("preds[2]") + ";").nl();
body.ip("preds[1] = 1.0-preds[2];").nl();
if (_parms._balance_classes)
body.ip(
"hex.genmodel.GenModel.correctProbabilities(preds, PRIOR_CLASS_DISTRIB, MODEL_CLASS_DISTRIB);")
.nl();
body.ip(
"preds[0] = hex.genmodel.GenModel.getPrediction(preds, data, "
+ defaultThreshold()
+ ");")
.nl();
return;
}
if (_output.nclasses() == 1) { // Regression
body.ip("preds[0] += ").p(_output._init_f).p(";").nl();
body.ip("preds[0] = " + _parms._distribution.linkInvString("preds[0]") + ";").nl();
return;
}
if (_output.nclasses() == 2) { // Kept the initial prediction for binomial
body.ip("preds[1] += ").p(_output._init_f).p(";").nl();
body.ip("preds[2] = - preds[1];").nl();
}
body.ip("hex.genmodel.GenModel.GBM_rescale(preds);").nl();
if (_parms._balance_classes)
body.ip(
"hex.genmodel.GenModel.correctProbabilities(preds, PRIOR_CLASS_DISTRIB, MODEL_CLASS_DISTRIB);")
.nl();
body.ip(
"preds[0] = hex.genmodel.GenModel.getPrediction(preds, data, "
+ defaultThreshold()
+ ");")
.nl();
}