public final double loss(double u, double a, Loss loss) {
assert loss.isForNumeric() : "Loss function " + loss + " not applicable to numerics";
switch (loss) {
case Quadratic:
return (u - a) * (u - a);
case Absolute:
return Math.abs(u - a);
case Huber:
return Math.abs(u - a) <= 1 ? 0.5 * (u - a) * (u - a) : Math.abs(u - a) - 0.5;
case Poisson:
assert a >= 0 : "Poisson loss L(u,a) requires variable a >= 0";
return Math.exp(u)
+ (a == 0 ? 0 : -a * u + a * Math.log(a) - a); // Since \lim_{a->0} a*log(a) = 0
case Hinge:
// return Math.max(1-a*u,0);
return Math.max(1 - (a == 0 ? -u : u), 0); // Booleans are coded {0,1} instead of {-1,1}
case Logistic:
// return Math.log(1 + Math.exp(-a * u));
return Math.log(
1 + Math.exp(a == 0 ? u : -u)); // Booleans are coded {0,1} instead of {-1,1}
case Periodic:
return 1 - Math.cos((a - u) * (2 * Math.PI) / _period);
default:
throw new RuntimeException("Unknown loss function " + loss);
}
}
/**
* Compute Variable Importance, based on GEDEON: DATA MINING OF INPUTS: ANALYSING MAGNITUDE AND
* FUNCTIONAL MEASURES
*
* @return variable importances for input features
*/
public float[] computeVariableImportances() {
float[] vi = new float[units[0]];
Arrays.fill(vi, 0f);
float[][] Qik = new float[units[0]][units[2]]; // importance of input i on output k
float[] sum_wj = new float[units[1]]; // sum of incoming weights into first hidden layer
float[] sum_wk =
new float[units[2]]; // sum of incoming weights into output layer (or second hidden layer)
for (float[] Qi : Qik) Arrays.fill(Qi, 0f);
Arrays.fill(sum_wj, 0f);
Arrays.fill(sum_wk, 0f);
// compute sum of absolute incoming weights
for (int j = 0; j < units[1]; j++) {
for (int i = 0; i < units[0]; i++) {
float wij = get_weights(0).get(j, i);
sum_wj[j] += Math.abs(wij);
}
}
for (int k = 0; k < units[2]; k++) {
for (int j = 0; j < units[1]; j++) {
float wjk = get_weights(1).get(k, j);
sum_wk[k] += Math.abs(wjk);
}
}
// compute importance of input i on output k as product of connecting weights going through j
for (int i = 0; i < units[0]; i++) {
for (int k = 0; k < units[2]; k++) {
for (int j = 0; j < units[1]; j++) {
float wij = get_weights(0).get(j, i);
float wjk = get_weights(1).get(k, j);
// Qik[i][k] += Math.abs(wij)/sum_wj[j] * wjk; //Wong,Gedeon,Taggart '95
Qik[i][k] += Math.abs(wij) / sum_wj[j] * Math.abs(wjk) / sum_wk[k]; // Gedeon '97
}
}
}
// normalize Qik over all outputs k
for (int k = 0; k < units[2]; k++) {
float sumQk = 0;
for (int i = 0; i < units[0]; i++) sumQk += Qik[i][k];
for (int i = 0; i < units[0]; i++) Qik[i][k] /= sumQk;
}
// importance for feature i is the sum over k of i->k importances
for (int i = 0; i < units[0]; i++) vi[i] = ArrayUtils.sum(Qik[i]);
// normalize importances such that max(vi) = 1
ArrayUtils.div(vi, ArrayUtils.maxValue(vi));
// zero out missing categorical variables if they were never seen
if (_saw_missing_cats != null) {
for (int i = 0; i < _saw_missing_cats.length; ++i) {
assert (data_info._catMissing[i] == 1); // have a missing bucket for each categorical
if (!_saw_missing_cats[i]) vi[data_info._catOffsets[i + 1] - 1] = 0;
}
}
return vi;
}
public final double regularize(double[] u, Regularizer regularization) {
if (u == null) return 0;
double ureg = 0;
switch (regularization) {
case None:
return 0;
case Quadratic:
for (int i = 0; i < u.length; i++) ureg += u[i] * u[i];
return ureg;
case L2:
for (int i = 0; i < u.length; i++) ureg += u[i] * u[i];
return Math.sqrt(ureg);
case L1:
for (int i = 0; i < u.length; i++) ureg += Math.abs(u[i]);
return ureg;
case NonNegative:
for (int i = 0; i < u.length; i++) {
if (u[i] < 0) return Double.POSITIVE_INFINITY;
}
return 0;
case OneSparse:
int card = 0;
for (int i = 0; i < u.length; i++) {
if (u[i] < 0) return Double.POSITIVE_INFINITY;
else if (u[i] > 0) card++;
}
return card == 1 ? 0 : Double.POSITIVE_INFINITY;
case UnitOneSparse:
int ones = 0, zeros = 0;
for (int i = 0; i < u.length; i++) {
if (u[i] == 1) ones++;
else if (u[i] == 0) zeros++;
else return Double.POSITIVE_INFINITY;
}
return ones == 1 && zeros == u.length - 1 ? 0 : Double.POSITIVE_INFINITY;
case Simplex:
double sum = 0, absum = 0;
for (int i = 0; i < u.length; i++) {
if (u[i] < 0) return Double.POSITIVE_INFINITY;
else {
sum += u[i];
absum += Math.abs(u[i]);
}
}
return MathUtils.equalsWithinRecSumErr(sum, 1.0, u.length, absum)
? 0
: Double.POSITIVE_INFINITY;
default:
throw new RuntimeException("Unknown regularization function " + regularization);
}
}
public final double lgrad(double u, double a, Loss loss) {
assert loss.isForNumeric() : "Loss function " + loss + " not applicable to numerics";
switch (loss) {
case Quadratic:
return 2 * (u - a);
case Absolute:
return Math.signum(u - a);
case Huber:
return Math.abs(u - a) <= 1 ? u - a : Math.signum(u - a);
case Poisson:
assert a >= 0 : "Poisson loss L(u,a) requires variable a >= 0";
return Math.exp(u) - a;
case Hinge:
// return a*u <= 1 ? -a : 0;
return a == 0
? (-u <= 1 ? 1 : 0)
: (u <= 1 ? -1 : 0); // Booleans are coded as {0,1} instead of {-1,1}
case Logistic:
// return -a/(1+Math.exp(a*u));
return a == 0
? 1 / (1 + Math.exp(-u))
: -1 / (1 + Math.exp(u)); // Booleans are coded as {0,1} instead of {-1,1}
case Periodic:
return ((2 * Math.PI) / _period) * Math.sin((a - u) * (2 * Math.PI) / _period);
default:
throw new RuntimeException("Unknown loss function " + loss);
}
}
@Override
public void map(Chunk[] chks) {
_gss = new double[_nclass][];
_rss = new double[_nclass][];
// For all tree/klasses
for (int k = 0; k < _nclass; k++) {
final DTree tree = _trees[k];
final int leaf = _leafs[k];
if (tree == null) continue; // Empty class is ignored
// A leaf-biased array of all active Tree leaves.
final double gs[] = _gss[k] = new double[tree._len - leaf];
final double rs[] = _rss[k] = new double[tree._len - leaf];
final Chunk nids = chk_nids(chks, k); // Node-ids for this tree/class
final Chunk ress = chk_work(chks, k); // Residuals for this tree/class
// If we have all constant responses, then we do not split even the
// root and the residuals should be zero.
if (tree.root() instanceof LeafNode) continue;
for (int row = 0; row < nids._len; row++) { // For all rows
int nid = (int) nids.at80(row); // Get Node to decide from
if (nid < 0) continue; // Missing response
if (tree.node(nid) instanceof UndecidedNode) // If we bottomed out the tree
nid = tree.node(nid)._pid; // Then take parent's decision
DecidedNode dn = tree.decided(nid); // Must have a decision point
if (dn._split._col == -1) // Unable to decide?
dn = tree.decided(nid = dn._pid); // Then take parent's decision
int leafnid = dn.ns(chks, row); // Decide down to a leafnode
assert leaf <= leafnid && leafnid < tree._len;
assert tree.node(leafnid) instanceof LeafNode;
// Note: I can which leaf/region I end up in, but I do not care for
// the prediction presented by the tree. For GBM, we compute the
// sum-of-residuals (and sum/abs/mult residuals) for all rows in the
// leaf, and get our prediction from that.
nids.set0(row, leafnid);
assert !ress.isNA0(row);
double res = ress.at0(row);
double ares = Math.abs(res);
gs[leafnid - leaf] += _nclass > 1 ? ares * (1 - ares) : 1;
rs[leafnid - leaf] += res;
}
}
}
@Override
protected void compute2() {
CoxPHModel model = null;
try {
Scope.enter();
_parms.read_lock_frames(CoxPH.this);
init(true);
applyScoringFrameSideEffects();
// The model to be built
model = new CoxPHModel(dest(), _parms, new CoxPHModel.CoxPHOutput(CoxPH.this));
model.delete_and_lock(_key);
applyTrainingFrameSideEffects();
int nResponses = 1;
boolean useAllFactorLevels = false;
final DataInfo dinfo =
new DataInfo(
Key.make(),
_modelBuilderTrain,
null,
nResponses,
useAllFactorLevels,
DataInfo.TransformType.DEMEAN,
TransformType.NONE,
true,
false,
false,
false,
false,
false);
initStats(model, dinfo);
final int n_offsets =
(model._parms.offset_columns == null) ? 0 : model._parms.offset_columns.length;
final int n_coef = dinfo.fullN() - n_offsets;
final double[] step = MemoryManager.malloc8d(n_coef);
final double[] oldCoef = MemoryManager.malloc8d(n_coef);
final double[] newCoef = MemoryManager.malloc8d(n_coef);
Arrays.fill(step, Double.NaN);
Arrays.fill(oldCoef, Double.NaN);
for (int j = 0; j < n_coef; ++j) newCoef[j] = model._parms.init;
double oldLoglik = -Double.MAX_VALUE;
final int n_time = (int) (model._output.max_time - model._output.min_time + 1);
final boolean has_start_column = (model._parms.start_column != null);
final boolean has_weights_column = (model._parms.weights_column != null);
for (int i = 0; i <= model._parms.iter_max; ++i) {
model._output.iter = i;
final CoxPHTask coxMR =
new CoxPHTask(
self(),
dinfo,
newCoef,
model._output.min_time,
n_time,
n_offsets,
has_start_column,
has_weights_column)
.doAll(dinfo._adaptedFrame);
final double newLoglik = calcLoglik(model, coxMR);
if (newLoglik > oldLoglik) {
if (i == 0) calcCounts(model, coxMR);
calcModelStats(model, newCoef, newLoglik);
calcCumhaz_0(model, coxMR);
if (newLoglik == 0) model._output.lre = -Math.log10(Math.abs(oldLoglik - newLoglik));
else model._output.lre = -Math.log10(Math.abs((oldLoglik - newLoglik) / newLoglik));
if (model._output.lre >= model._parms.lre_min) break;
Arrays.fill(step, 0);
for (int j = 0; j < n_coef; ++j)
for (int k = 0; k < n_coef; ++k)
step[j] -= model._output.var_coef[j][k] * model._output.gradient[k];
for (int j = 0; j < n_coef; ++j)
if (Double.isNaN(step[j]) || Double.isInfinite(step[j])) break;
oldLoglik = newLoglik;
System.arraycopy(newCoef, 0, oldCoef, 0, oldCoef.length);
} else {
for (int j = 0; j < n_coef; ++j) step[j] /= 2;
}
for (int j = 0; j < n_coef; ++j) newCoef[j] = oldCoef[j] - step[j];
}
model.update(_key);
} catch (Throwable t) {
Job thisJob = DKV.getGet(_key);
if (thisJob._state == JobState.CANCELLED) {
Log.info("Job cancelled by user.");
} else {
t.printStackTrace();
failed(t);
throw t;
}
} finally {
updateModelOutput();
_parms.read_unlock_frames(CoxPH.this);
Scope.exit();
done(); // Job done!
}
tryComplete();
}
