@Override
protected SBPrintStream toJavaInit(SBPrintStream sb, CodeGeneratorPipeline fileCtx) {
sb.nl();
sb.ip("public boolean isSupervised() { return true; }").nl();
sb.ip("public int nfeatures() { return " + _output.nfeatures() + "; }").nl();
sb.ip("public int nclasses() { return " + _output.nclasses() + "; }").nl();
return sb;
}
@Override
protected SBPrintStream toJavaInit(SBPrintStream sb, CodeGeneratorPipeline fileCtx) {
sb = super.toJavaInit(sb, fileCtx);
sb.ip("public boolean isSupervised() { return " + isSupervised() + "; }").nl();
sb.ip("public int nfeatures() { return " + _output.nfeatures() + "; }").nl();
sb.ip("public int nclasses() { return " + _output.nclasses() + "; }").nl();
JCodeGen.toStaticVar(sb, "RESCNT", _output._rescnt, "Count of categorical levels in response.");
JCodeGen.toStaticVar(
sb, "APRIORI", _output._apriori_raw, "Apriori class distribution of the response.");
JCodeGen.toStaticVar(sb, "PCOND", _output._pcond_raw, "Conditional probability of predictors.");
double[] dlen = null;
if (_output._ncats > 0) {
dlen = new double[_output._ncats];
for (int i = 0; i < _output._ncats; i++) dlen[i] = _output._domains[i].length;
}
JCodeGen.toStaticVar(
sb, "DOMLEN", dlen, "Number of unique levels for each categorical predictor.");
return sb;
}
@Override
protected SBPrintStream toJavaInit(SBPrintStream sb, CodeGeneratorPipeline fileCtx) {
sb = super.toJavaInit(sb, fileCtx);
sb.ip("public boolean isSupervised() { return " + isSupervised() + "; }").nl();
sb.ip("public int nfeatures() { return " + _output.nfeatures() + "; }").nl();
sb.ip("public int nclasses() { return " + _parms._nv + "; }").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._v, "Eigenvector matrix.");
return sb;
}
@Override
protected void toJavaPredictBody(
SBPrintStream body,
CodeGeneratorPipeline classCtx,
CodeGeneratorPipeline fileCtx,
final boolean verboseCode) {
// Generate static fields
classCtx.add(
new CodeGenerator() {
@Override
public void generate(JCodeSB out) {
JCodeGen.toClassWithArray(out, "static", "BETA", beta_internal()); // "The Coefficients"
JCodeGen.toClassWithArray(
out, "static", "NUM_MEANS", _output._dinfo._numMeans, "Imputed numeric values");
JCodeGen.toClassWithArray(
out,
"static",
"CAT_MODES",
_output._dinfo._catModes,
"Imputed categorical values.");
JCodeGen.toStaticVar(out, "CATOFFS", dinfo()._catOffsets, "Categorical Offsets");
}
});
body.ip("final double [] b = BETA.VALUES;").nl();
if (_parms._missing_values_handling == MissingValuesHandling.MeanImputation) {
body.ip(
"for(int i = 0; i < "
+ _output._dinfo._cats
+ "; ++i) if(Double.isNaN(data[i])) data[i] = CAT_MODES.VALUES[i];")
.nl();
body.ip(
"for(int i = 0; i < "
+ _output._dinfo._nums
+ "; ++i) if(Double.isNaN(data[i + "
+ _output._dinfo._cats
+ "])) data[i+"
+ _output._dinfo._cats
+ "] = NUM_MEANS.VALUES[i];")
.nl();
}
if (_parms._family != Family.multinomial) {
body.ip("double eta = 0.0;").nl();
if (!_parms._use_all_factor_levels) { // skip level 0 of all factors
body.ip("for(int i = 0; i < CATOFFS.length-1; ++i) if(data[i] != 0) {").nl();
body.ip(" int ival = (int)data[i] - 1;").nl();
body.ip(
" if(ival != data[i] - 1) throw new IllegalArgumentException(\"categorical value out of range\");")
.nl();
body.ip(" ival += CATOFFS[i];").nl();
body.ip(" if(ival < CATOFFS[i + 1])").nl();
body.ip(" eta += b[ival];").nl();
} else { // do not skip any levels
body.ip("for(int i = 0; i < CATOFFS.length-1; ++i) {").nl();
body.ip(" int ival = (int)data[i];").nl();
body.ip(
" if(ival != data[i]) throw new IllegalArgumentException(\"categorical value out of range\");")
.nl();
body.ip(" ival += CATOFFS[i];").nl();
body.ip(" if(ival < CATOFFS[i + 1])").nl();
body.ip(" eta += b[ival];").nl();
}
body.ip("}").nl();
final int noff = dinfo().numStart() - dinfo()._cats;
body.ip("for(int i = ").p(dinfo()._cats).p("; i < b.length-1-").p(noff).p("; ++i)").nl();
body.ip(" eta += b[").p(noff).p("+i]*data[i];").nl();
body.ip("eta += b[b.length-1]; // reduce intercept").nl();
if (_parms._family != Family.tweedie)
body.ip("double mu = hex.genmodel.GenModel.GLM_").p(_parms._link.toString()).p("Inv(eta");
else
body.ip(
"double mu = hex.genmodel.GenModel.GLM_tweedieInv(eta," + _parms._tweedie_link_power);
body.p(");").nl();
if (_parms._family == Family.binomial) {
body.ip("preds[0] = (mu > ")
.p(defaultThreshold())
.p(") ? 1 : 0")
.p("; // threshold given by ROC")
.nl();
body.ip("preds[1] = 1.0 - mu; // class 0").nl();
body.ip("preds[2] = mu; // class 1").nl();
} else {
body.ip("preds[0] = mu;").nl();
}
} else {
int P = _output._global_beta_multinomial[0].length;
body.ip("preds[0] = 0;").nl();
body.ip("for(int c = 0; c < " + _output._nclasses + "; ++c){").nl();
body.ip(" preds[c+1] = 0;").nl();
if (dinfo()._cats > 0) {
if (!_parms._use_all_factor_levels) { // skip level 0 of all factors
body.ip(" for(int i = 0; i < CATOFFS.length-1; ++i) if(data[i] != 0) {").nl();
body.ip(" int ival = (int)data[i] - 1;").nl();
body.ip(
" if(ival != data[i] - 1) throw new IllegalArgumentException(\"categorical value out of range\");")
.nl();
body.ip(" ival += CATOFFS[i];").nl();
body.ip(" if(ival < CATOFFS[i + 1])").nl();
body.ip(" preds[c+1] += b[ival+c*" + P + "];").nl();
} else { // do not skip any levels
body.ip(" for(int i = 0; i < CATOFFS.length-1; ++i) {").nl();
body.ip(" int ival = (int)data[i];").nl();
body.ip(
" if(ival != data[i]) throw new IllegalArgumentException(\"categorical value out of range\");")
.nl();
body.ip(" ival += CATOFFS[i];").nl();
body.ip(" if(ival < CATOFFS[i + 1])").nl();
body.ip(" preds[c+1] += b[ival+c*" + P + "];").nl();
}
body.ip(" }").nl();
}
final int noff = dinfo().numStart();
body.ip(" for(int i = 0; i < " + dinfo()._nums + "; ++i)").nl();
body.ip(" preds[c+1] += b[" + noff + "+i + c*" + P + "]*data[i];").nl();
body.ip(" preds[c+1] += b[" + (P - 1) + " + c*" + P + "]; // reduce intercept").nl();
body.ip("}").nl();
body.ip("double max_row = 0;").nl();
body.ip("for(int c = 1; c < preds.length; ++c) if(preds[c] > max_row) max_row = preds[c];")
.nl();
body.ip("double sum_exp = 0;").nl();
body.ip(
"for(int c = 1; c < preds.length; ++c) { sum_exp += (preds[c] = Math.exp(preds[c]-max_row));}")
.nl();
body.ip("sum_exp = 1/sum_exp;").nl();
body.ip("double max_p = 0;").nl();
body.ip(
"for(int c = 1; c < preds.length; ++c) if((preds[c] *= sum_exp) > max_p){ max_p = preds[c]; preds[0] = c-1;};")
.nl();
}
}
@Override
protected void toJavaPredictBody(
SBPrintStream bodySb,
CodeGeneratorPipeline classCtx,
CodeGeneratorPipeline fileCtx,
final boolean verboseCode) {
bodySb.i().p("java.util.Arrays.fill(preds,0);").nl();
bodySb.i().p("double mean, sdev, prob;").nl();
bodySb.i().p("double[] nums = new double[" + _output._levels.length + "];").nl();
bodySb.i().p("for(int i = 0; i < " + _output._levels.length + "; i++) {").nl();
bodySb.i(1).p("nums[i] = Math.log(APRIORI[i]);").nl();
bodySb.i(1).p("for(int j = 0; j < " + _output._ncats + "; j++) {").nl();
bodySb.i(2).p("if(Double.isNaN(data[j])) continue;").nl();
bodySb.i(2).p("int level = (int)data[j];").nl();
bodySb
.i(2)
.p(
"prob = level < "
+ _output._pcond_raw.length
+ " ? PCOND[j][i][level] : "
+ (_parms._laplace == 0
? 0
: _parms._laplace + "/(RESCNT[i] + " + _parms._laplace + "*DOMLEN[j])"))
.p(";")
.nl();
bodySb
.i(2)
.p(
"nums[i] += Math.log(prob <= "
+ _parms._eps_prob
+ " ? "
+ _parms._min_prob
+ " : prob);")
.nl();
bodySb.i(1).p("}").nl();
bodySb.i(1).p("for(int j = " + _output._ncats + "; j < data.length; j++) {").nl();
bodySb.i(2).p("if(Double.isNaN(data[j])) continue;").nl();
bodySb.i(2).p("mean = Double.isNaN(PCOND[j][i][0]) ? 0 : PCOND[j][i][0];").nl();
bodySb
.i(2)
.p(
"sdev = Double.isNaN(PCOND[j][i][1]) ? 1 : (PCOND[j][i][1] <= "
+ _parms._eps_sdev
+ " ? "
+ _parms._min_sdev
+ " : PCOND[j][i][1]);")
.nl();
bodySb
.i(2)
.p(
"prob = Math.exp(-((data[j]-mean)*(data[j]-mean))/(2.*sdev*sdev)) / (sdev*Math.sqrt(2.*Math.PI));")
.nl();
bodySb
.i(2)
.p(
"nums[i] += Math.log(prob <= "
+ _parms._eps_prob
+ " ? "
+ _parms._min_prob
+ " : prob);")
.nl();
bodySb.i(1).p("}").nl();
bodySb.i().p("}").nl();
bodySb.i().p("double sum;").nl();
bodySb.i().p("for(int i = 0; i < nums.length; i++) {").nl();
bodySb.i(1).p("sum = 0;").nl();
bodySb.i(1).p("for(int j = 0; j < nums.length; j++) {").nl();
bodySb.i(2).p("sum += Math.exp(nums[j]-nums[i]);").nl();
bodySb.i(1).p("}").nl();
bodySb.i(1).p("preds[i+1] = 1/sum;").nl();
bodySb.i().p("}").nl();
bodySb
.i()
.p(
"preds[0] = hex.genmodel.GenModel.getPrediction(preds, PRIOR_CLASS_DISTRIB, data, "
+ defaultThreshold()
+ ");")
.nl();
}
@Override
protected void toJavaPredictBody(
SBPrintStream bodySb,
CodeGeneratorPipeline classCtx,
CodeGeneratorPipeline fileCtx,
final boolean verboseCode) {
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._nv).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("int last = CATOFFS[j+1]-CATOFFS[j]-1;").nl();
bodySb
.i(2)
.p("int c = Double.isNaN(d) ? last : (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();
}