public final double deviance(double yr, double ym) {
double y1 = yr == 0 ? .1 : yr;
switch (_family) {
case gaussian:
return (yr - ym) * (yr - ym);
case binomial:
return 2 * ((MathUtils.y_log_y(yr, ym)) + MathUtils.y_log_y(1 - yr, 1 - ym));
case poisson:
if (yr == 0) return 2 * ym;
return 2 * ((yr * Math.log(yr / ym)) - (yr - ym));
case gamma:
if (yr == 0) return -2;
return -2 * (Math.log(yr / ym) - (yr - ym) / ym);
case tweedie:
double theta =
_var_power == 1
? Math.log(y1 / ym)
: (Math.pow(y1, 1. - _var_power) - Math.pow(ym, 1 - _var_power))
/ (1 - _var_power);
double kappa =
_var_power == 2
? Math.log(y1 / ym)
: (Math.pow(yr, 2 - _var_power) - Math.pow(ym, 2 - _var_power))
/ (2 - _var_power);
return 2 * (yr * theta - kappa);
default:
throw new RuntimeException("unknown family " + _family);
}
}
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);
}
}
@Test
public void testQuantile() {
Frame f = null;
try {
Frame fr =
frame(
ard(
ard(1.223292e-02),
ard(1.635312e-25),
ard(1.601522e-11),
ard(8.452298e-10),
ard(2.643733e-10),
ard(2.671520e-06),
ard(1.165381e-06),
ard(7.193265e-10),
ard(3.383532e-04),
ard(2.561221e-05)));
double[] probs = new double[] {0.001, 0.005, .01, .02, .05, .10, .50, .8883, .90, .99};
String x =
String.format("(quantile %%%s %s \"interpolate\")", fr._key, Arrays.toString(probs));
Val val = Exec.exec(x);
fr.delete();
f = val.getFrame();
Assert.assertEquals(2, f.numCols());
// Expected values computed as golden values from R's quantile call
double[] exp =
ard(
1.4413698000016206E-13,
7.206849000001562E-13,
1.4413698000001489E-12,
2.882739600000134E-12,
7.20684900000009E-12,
1.4413698000000017E-11,
5.831131148999999E-07,
3.3669567275300000E-04,
0.00152780988,
0.011162408988);
for (int i = 0; i < exp.length; i++)
Assert.assertTrue(
"expected " + exp[i] + " got " + f.vec(1).at(i),
water.util.MathUtils.compare(exp[i], f.vec(1).at(i), 1e-6, 1e-6));
} finally {
if (f != null) f.delete();
}
}
public static Frame[] shuffleSplitFrame(
Frame fr, Key[] keys, final double ratios[], final long seed) {
// Sanity check the ratios
assert keys.length == ratios.length;
double sum = ratios[0];
for (int i = 1; i < ratios.length; i++) {
sum += ratios[i];
ratios[i] = sum;
}
assert water.util.MathUtils.equalsWithinOneSmallUlp(sum, 1.0);
// Do the split, into ratios.length groupings of NewChunks
final int ncols = fr.numCols();
MRTask mr =
new MRTask() {
@Override
public void map(Chunk cs[], NewChunk ncs[]) {
Random rng = new Random(seed * cs[0].cidx());
int nrows = cs[0]._len;
for (int i = 0; i < nrows; i++) {
double r = rng.nextDouble();
int x = 0; // Pick the NewChunk split
for (; x < ratios.length - 1; x++) if (r < ratios[x]) break;
x *= ncols;
// Helper string holder
ValueString vstr = new ValueString();
// Copy row to correct set of NewChunks
for (int j = 0; j < ncols; j++) {
byte colType = cs[j].vec().get_type();
switch (colType) {
case Vec.T_BAD:
break; /* NOP */
case Vec.T_STR:
ncs[x + j].addStr(cs[j], i);
break;
case Vec.T_UUID:
ncs[x + j].addUUID(cs[j], i);
break;
case Vec.T_NUM: /* fallthrough */
case Vec.T_ENUM:
case Vec.T_TIME:
ncs[x + j].addNum(cs[j].atd(i));
break;
default:
if (colType > Vec.T_TIME && colType <= Vec.T_TIMELAST)
ncs[x + j].addNum(cs[j].atd(i));
else throw new IllegalArgumentException("Unsupported vector type: " + colType);
break;
}
}
}
}
}.doAll(ncols * ratios.length, fr);
// Build output frames
Frame frames[] = new Frame[ratios.length];
Vec[] vecs = fr.vecs();
String[] names = fr.names();
Futures fs = new Futures();
for (int i = 0; i < ratios.length; i++) {
Vec[] nvecs = new Vec[ncols];
for (int c = 0; c < ncols; c++) {
mr.appendables()[i * ncols + c].setDomain(vecs[c].domain());
nvecs[c] = mr.appendables()[i * ncols + c].close(fs);
}
frames[i] = new Frame(keys[i], fr.names(), nvecs);
DKV.put(frames[i], fs);
}
fs.blockForPending();
return frames;
}
/** Re-do the TwoDim table generation with updated model. */
public TwoDimTable generateSummary(Key train, int iter) {
String[] names =
new String[] {
"Family",
"Link",
"Regularization",
"Number of Predictors Total",
"Number of Active Predictors",
"Number of Iterations",
"Training Frame"
};
String[] types = new String[] {"string", "string", "string", "int", "int", "int", "string"};
String[] formats = new String[] {"%s", "%s", "%s", "%d", "%d", "%d", "%s"};
if (_parms._lambda_search) {
names =
new String[] {
"Family",
"Link",
"Regularization",
"Lambda Search",
"Number of Predictors Total",
"Number of Active Predictors",
"Number of Iterations",
"Training Frame"
};
types = new String[] {"string", "string", "string", "string", "int", "int", "int", "string"};
formats = new String[] {"%s", "%s", "%s", "%s", "%d", "%d", "%d", "%s"};
}
_output._model_summary =
new TwoDimTable("GLM Model", "summary", new String[] {""}, names, types, formats, "");
_output._model_summary.set(0, 0, _parms._family.toString());
_output._model_summary.set(0, 1, _parms._link.toString());
String regularization = "None";
if (_parms._lambda != null
&& !(_parms._lambda.length == 1 && _parms._lambda[0] == 0)) { // have regularization
if (_parms._alpha[0] == 0) regularization = "Ridge ( lambda = ";
else if (_parms._alpha[0] == 1) regularization = "Lasso (lambda = ";
else
regularization =
"Elastic Net (alpha = " + MathUtils.roundToNDigits(_parms._alpha[0], 4) + ", lambda = ";
regularization =
regularization
+ MathUtils.roundToNDigits(_parms._lambda[_output._best_lambda_idx], 4)
+ " )";
}
_output._model_summary.set(0, 2, regularization);
int lambdaSearch = 0;
if (_parms._lambda_search) {
lambdaSearch = 1;
_output._model_summary.set(
0,
3,
"nlambda = "
+ _parms._nlambdas
+ ", lambda_max = "
+ MathUtils.roundToNDigits(_lambda_max, 4)
+ ", best_lambda = "
+ MathUtils.roundToNDigits(_output.bestSubmodel().lambda_value, 4));
}
int intercept = _parms._intercept ? 1 : 0;
if (_output.nclasses() > 2) {
_output._model_summary.set(0, 3 + lambdaSearch, _output.bestSubmodel().beta.length);
} else {
_output._model_summary.set(0, 3 + lambdaSearch, beta().length);
}
_output._model_summary.set(0, 4 + lambdaSearch, Integer.toString(_output.rank() - intercept));
_output._model_summary.set(0, 5 + lambdaSearch, Integer.valueOf(iter));
_output._model_summary.set(0, 6 + lambdaSearch, train.toString());
return _output._model_summary;
}
/** Compute statistics about this model on all nodes */
public void computeStats() {
if (!get_params()._diagnostics) return;
float[][] rate = get_params()._adaptive_rate ? new float[units.length - 1][] : null;
if (get_params()._autoencoder && get_params()._sparsity_beta > 0) {
for (int k = 0; k < get_params()._hidden.length; k++) {
mean_a[k] = 0;
for (int j = 0; j < avg_activations[k].size(); j++) mean_a[k] += avg_activations[k].get(j);
mean_a[k] /= avg_activations[k].size();
}
}
for (int y = 1; y < units.length; y++) {
mean_rate[y] = rms_rate[y] = 0;
mean_bias[y] = rms_bias[y] = 0;
mean_weight[y] = rms_weight[y] = 0;
for (int u = 0; u < biases[y - 1].size(); u++) {
mean_bias[y] += biases[y - 1].get(u);
}
if (rate != null) rate[y - 1] = new float[get_weights(y - 1).raw().length];
for (int u = 0; u < get_weights(y - 1).raw().length; u++) {
mean_weight[y] += get_weights(y - 1).raw()[u];
if (rate != null) {
// final float RMS_dx =
// (float)Math.sqrt(ada[y-1][2*u]+(float)get_params().epsilon);
// final float invRMS_g =
// (float)(1/Math.sqrt(ada[y-1][2*u+1]+(float)get_params().epsilon));
final float RMS_dx =
MathUtils.approxSqrt(
get_ada_dx_g(y - 1).raw()[2 * u] + (float) get_params()._epsilon);
final float invRMS_g =
MathUtils.approxInvSqrt(
get_ada_dx_g(y - 1).raw()[2 * u + 1] + (float) get_params()._epsilon);
rate[y - 1][u] =
RMS_dx
* invRMS_g; // not exactly right, RMS_dx should be from the previous time step ->
// but close enough for diagnostics.
mean_rate[y] += rate[y - 1][u];
}
}
mean_bias[y] /= biases[y - 1].size();
mean_weight[y] /= get_weights(y - 1).size();
if (rate != null) mean_rate[y] /= rate[y - 1].length;
for (int u = 0; u < biases[y - 1].size(); u++) {
final double db = biases[y - 1].get(u) - mean_bias[y];
rms_bias[y] += db * db;
}
for (int u = 0; u < get_weights(y - 1).size(); u++) {
final double dw = get_weights(y - 1).raw()[u] - mean_weight[y];
rms_weight[y] += dw * dw;
if (rate != null) {
final double drate = rate[y - 1][u] - mean_rate[y];
rms_rate[y] += drate * drate;
}
}
rms_bias[y] = MathUtils.approxSqrt(rms_bias[y] / biases[y - 1].size());
rms_weight[y] = MathUtils.approxSqrt(rms_weight[y] / get_weights(y - 1).size());
if (rate != null) rms_rate[y] = MathUtils.approxSqrt(rms_rate[y] / rate[y - 1].length);
// rms_bias[y] = (float)Math.sqrt(rms_bias[y]/biases[y-1].length);
// rms_weight[y] = (float)Math.sqrt(rms_weight[y]/weights[y-1].length);
// if (rate != null) rms_rate[y] = (float)Math.sqrt(rms_rate[y]/rate[y-1].length);
// Abort the run if weights or biases are unreasonably large (Note that all input values are
// normalized upfront)
// This can happen with Rectifier units when L1/L2/max_w2 are all set to 0, especially when
// using more than 1 hidden layer.
final double thresh = 1e10;
unstable |=
mean_bias[y] > thresh
|| isNaN(mean_bias[y])
|| rms_bias[y] > thresh
|| isNaN(rms_bias[y])
|| mean_weight[y] > thresh
|| isNaN(mean_weight[y])
|| rms_weight[y] > thresh
|| isNaN(rms_weight[y]);
}
}
