private static void run(Callable c, boolean read, int size) {
// Count all i/o time from here, including all retry overheads
long start_io_ms = System.currentTimeMillis();
while (true) {
try {
long start_ns = System.nanoTime(); // Blocking i/o call timing - without counting repeats
c.call();
TimeLine.record_IOclose(start_ns, start_io_ms, read ? 1 : 0, size, Value.HDFS);
break;
// Explicitly ignore the following exceptions but
// fail on the rest IOExceptions
} catch (EOFException e) {
ignoreAndWait(e, false);
} catch (SocketTimeoutException e) {
ignoreAndWait(e, false);
} catch (S3Exception e) {
// Preserve S3Exception before IOException
// Since this is tricky code - we are supporting different HDFS version
// New version declares S3Exception as IOException
// But old versions (0.20.xxx) declares it as RuntimeException
// So we have to catch it before IOException !!!
ignoreAndWait(e, false);
} catch (IOException e) {
ignoreAndWait(e, true);
} catch (Exception e) {
throw Log.errRTExcept(e);
}
}
}
private static void run(Callable c, boolean read, int size) {
// Count all i/o time from here, including all retry overheads
long start_io_ms = System.currentTimeMillis();
while (true) {
try {
long start_ns = System.nanoTime(); // Blocking i/o call timing - without counting repeats
c.call();
TimeLine.record_IOclose(start_ns, start_io_ms, read ? 1 : 0, size, Value.HDFS);
break;
// Explicitly ignore the following exceptions but
// fail on the rest IOExceptions
} catch (EOFException e) {
ignoreAndWait(e, false);
} catch (SocketTimeoutException e) {
ignoreAndWait(e, false);
} catch (IOException e) {
ignoreAndWait(e, true);
} catch (Exception e) {
throw Log.errRTExcept(e);
}
}
}
public static class GLRMParameters extends Model.Parameters {
public String algoName() {
return "GLRM";
}
public String fullName() {
return "Generalized Low Rank Modeling";
}
public String javaName() {
return GLRMModel.class.getName();
}
public DataInfo.TransformType _transform =
DataInfo.TransformType.NONE; // Data transformation (demean to compare with PCA)
public int _k = 1; // Rank of resulting XY matrix
public GLRM.Initialization _init = GLRM.Initialization.PlusPlus; // Initialization of Y matrix
public SVDParameters.Method _svd_method =
SVDParameters.Method.Randomized; // SVD initialization method (for _init = SVD)
public Key<Frame> _user_y; // User-specified Y matrix (for _init = User)
public Key<Frame> _user_x; // User-specified X matrix (for _init = User)
public boolean _expand_user_y =
true; // Should categorical columns in _user_y be expanded via one-hot encoding? (for _init
// = User)
// Loss functions
public Loss _loss = Loss.Quadratic; // Default loss function for numeric cols
public Loss _multi_loss = Loss.Categorical; // Default loss function for categorical cols
public int _period = 1; // Length of the period when _loss = Periodic
public Loss[] _loss_by_col; // Override default loss function for specific columns
public int[] _loss_by_col_idx;
// Regularization functions
public Regularizer _regularization_x = Regularizer.None; // Regularization function for X matrix
public Regularizer _regularization_y = Regularizer.None; // Regularization function for Y matrix
public double _gamma_x = 0; // Regularization weight on X matrix
public double _gamma_y = 0; // Regularization weight on Y matrix
// Optional parameters
public int _max_iterations = 1000; // Max iterations
public int _max_updates = 2 * _max_iterations; // Max number of updates (X or Y)
public double _init_step_size = 1.0; // Initial step size (decrease until we hit min_step_size)
public double _min_step_size = 1e-4; // Min step size
public long _seed = System.nanoTime(); // RNG seed
@Override
protected long nFoldSeed() {
return _seed;
}
// public Key<Frame> _representation_key; // Key to save X matrix
public String _representation_name;
public boolean _recover_svd =
false; // Recover singular values and eigenvectors of XY at the end?
public boolean _impute_original =
false; // Reconstruct original training data by reversing _transform?
public boolean _verbose = true; // Log when objective increases each iteration?
// Quadratic -> Gaussian distribution ~ exp(-(a-u)^2)
// Absolute -> Laplace distribution ~ exp(-|a-u|)
public enum Loss {
Quadratic(true),
Absolute(true),
Huber(true),
Poisson(true),
Periodic(true), // One-dimensional loss (numeric)
Logistic(true, true),
Hinge(true, true), // Boolean loss (categorical)
Categorical(false),
Ordinal(false); // Multi-dimensional loss (categorical)
private boolean forNumeric;
private boolean forBinary;
Loss(boolean forNumeric) {
this(forNumeric, false);
}
Loss(boolean forNumeric, boolean forBinary) {
this.forNumeric = forNumeric;
this.forBinary = forBinary;
}
public boolean isForNumeric() {
return forNumeric;
}
public boolean isForCategorical() {
return !forNumeric;
}
public boolean isForBinary() {
return forBinary;
}
}
// Non-negative matrix factorization (NNMF): r_x = r_y = NonNegative
// Orthogonal NNMF: r_x = OneSparse, r_y = NonNegative
// K-means clustering: r_x = UnitOneSparse, r_y = 0 (\gamma_y = 0)
// Quadratic mixture: r_x = Simplex, r_y = 0 (\gamma_y = 0)
public enum Regularizer {
None,
Quadratic,
L2,
L1,
NonNegative,
OneSparse,
UnitOneSparse,
Simplex
}
// Check if all elements of _loss_by_col are equal to a specific loss function
private final boolean allLossEquals(Loss loss) {
if (null == _loss_by_col) return false;
boolean res = true;
for (int i = 0; i < _loss_by_col.length; i++) {
if (_loss_by_col[i] != loss) {
res = false;
break;
}
}
return res;
}
// Closed form solution only if quadratic loss, no regularization or quadratic regularization
// (same for X and Y), and no missing values
public final boolean hasClosedForm() {
long na_cnt = 0;
Frame train = _train.get();
for (int i = 0; i < train.numCols(); i++) na_cnt += train.vec(i).naCnt();
return hasClosedForm(na_cnt);
}
public final boolean hasClosedForm(long na_cnt) {
boolean loss_quad =
(null == _loss_by_col && _loss == Quadratic)
|| (null != _loss_by_col
&& allLossEquals(Quadratic)
&& (_loss_by_col.length == _train.get().numCols() || _loss == Quadratic));
return na_cnt == 0
&& ((loss_quad
&& (_gamma_x == 0
|| _regularization_x == Regularizer.None
|| _regularization_x == GLRMParameters.Regularizer.Quadratic)
&& (_gamma_y == 0
|| _regularization_y == Regularizer.None
|| _regularization_y == GLRMParameters.Regularizer.Quadratic)));
}
// L(u,a): Loss function
public final double loss(double u, double a) {
return loss(u, a, _loss);
}
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);
}
}
// \grad_u L(u,a): Gradient of loss function with respect to u
public final double lgrad(double u, double a) {
return lgrad(u, a, _loss);
}
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);
}
}
// L(u,a): Multidimensional loss function
public final double mloss(double[] u, int a) {
return mloss(u, a, _multi_loss);
}
public static double mloss(double[] u, int a, Loss multi_loss) {
assert multi_loss.isForCategorical()
: "Loss function " + multi_loss + " not applicable to categoricals";
if (a < 0 || a > u.length - 1)
throw new IllegalArgumentException(
"Index must be between 0 and " + String.valueOf(u.length - 1));
double sum = 0;
switch (multi_loss) {
case Categorical:
for (int i = 0; i < u.length; i++) sum += Math.max(1 + u[i], 0);
sum += Math.max(1 - u[a], 0) - Math.max(1 + u[a], 0);
return sum;
case Ordinal:
for (int i = 0; i < u.length - 1; i++) sum += Math.max(a > i ? 1 - u[i] : 1, 0);
return sum;
default:
throw new RuntimeException("Unknown multidimensional loss function " + multi_loss);
}
}
// \grad_u L(u,a): Gradient of multidimensional loss function with respect to u
public final double[] mlgrad(double[] u, int a) {
return mlgrad(u, a, _multi_loss);
}
public static double[] mlgrad(double[] u, int a, Loss multi_loss) {
assert multi_loss.isForCategorical()
: "Loss function " + multi_loss + " not applicable to categoricals";
if (a < 0 || a > u.length - 1)
throw new IllegalArgumentException(
"Index must be between 0 and " + String.valueOf(u.length - 1));
double[] grad = new double[u.length];
switch (multi_loss) {
case Categorical:
for (int i = 0; i < u.length; i++) grad[i] = (1 + u[i] > 0) ? 1 : 0;
grad[a] = (1 - u[a] > 0) ? -1 : 0;
return grad;
case Ordinal:
for (int i = 0; i < u.length - 1; i++) grad[i] = (a > i && 1 - u[i] > 0) ? -1 : 0;
return grad;
default:
throw new RuntimeException("Unknown multidimensional loss function " + multi_loss);
}
}
// r_i(x_i), r_j(y_j): Regularization function for single row x_i or column y_j
public final double regularize_x(double[] u) {
return regularize(u, _regularization_x);
}
public final double regularize_y(double[] u) {
return regularize(u, _regularization_y);
}
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);
}
}
// \sum_i r_i(x_i): Sum of regularization function for all entries of X
public final double regularize_x(double[][] u) {
return regularize(u, _regularization_x);
}
public final double regularize_y(double[][] u) {
return regularize(u, _regularization_y);
}
public final double regularize(double[][] u, Regularizer regularization) {
if (u == null || regularization == Regularizer.None) return 0;
double ureg = 0;
for (int i = 0; i < u.length; i++) {
ureg += regularize(u[i], regularization);
if (Double.isInfinite(ureg)) return ureg;
}
return ureg;
}
// \prox_{\alpha_k*r}(u): Proximal gradient of (step size) * (regularization function) evaluated
// at vector u
public final double[] rproxgrad_x(double[] u, double alpha, Random rand) {
return rproxgrad(u, alpha, _gamma_x, _regularization_x, rand);
}
public final double[] rproxgrad_y(double[] u, double alpha, Random rand) {
return rproxgrad(u, alpha, _gamma_y, _regularization_y, rand);
}
// public final double[] rproxgrad_x(double[] u, double alpha) { return rproxgrad(u, alpha,
// _gamma_x, _regularization_x, RandomUtils.getRNG(_seed)); }
// public final double[] rproxgrad_y(double[] u, double alpha) { return rproxgrad(u, alpha,
// _gamma_y, _regularization_y, RandomUtils.getRNG(_seed)); }
static double[] rproxgrad(
double[] u, double alpha, double gamma, Regularizer regularization, Random rand) {
if (u == null || alpha == 0 || gamma == 0) return u;
double[] v = new double[u.length];
switch (regularization) {
case None:
return u;
case Quadratic:
for (int i = 0; i < u.length; i++) v[i] = u[i] / (1 + 2 * alpha * gamma);
return v;
case L2:
// Proof uses Moreau decomposition; see section 6.5.1 of Parikh and Boyd
// https://web.stanford.edu/~boyd/papers/pdf/prox_algs.pdf
double weight = 1 - alpha * gamma / ArrayUtils.l2norm(u);
if (weight < 0) return v; // Zero vector
for (int i = 0; i < u.length; i++) v[i] = weight * u[i];
return v;
case L1:
for (int i = 0; i < u.length; i++)
v[i] = Math.max(u[i] - alpha * gamma, 0) + Math.min(u[i] + alpha * gamma, 0);
return v;
case NonNegative:
for (int i = 0; i < u.length; i++) v[i] = Math.max(u[i], 0);
return v;
case OneSparse:
int idx = ArrayUtils.maxIndex(u, rand);
v[idx] = u[idx] > 0 ? u[idx] : 1e-6;
return v;
case UnitOneSparse:
idx = ArrayUtils.maxIndex(u, rand);
v[idx] = 1;
return v;
case Simplex:
// Proximal gradient algorithm by Chen and Ye in http://arxiv.org/pdf/1101.6081v2.pdf
// 1) Sort input vector u in ascending order: u[1] <= ... <= u[n]
int n = u.length;
int[] idxs = new int[n];
for (int i = 0; i < n; i++) idxs[i] = i;
ArrayUtils.sort(idxs, u);
// 2) Calculate cumulative sum of u in descending order
// cumsum(u) = (..., u[n-2]+u[n-1]+u[n], u[n-1]+u[n], u[n])
double[] ucsum = new double[n];
ucsum[n - 1] = u[idxs[n - 1]];
for (int i = n - 2; i >= 0; i--) ucsum[i] = ucsum[i + 1] + u[idxs[i]];
// 3) Let t_i = (\sum_{j=i+1}^n u[j] - 1)/(n - i)
// For i = n-1,...,1, set optimal t* to first t_i >= u[i]
double t = (ucsum[0] - 1) / n; // Default t* = (\sum_{j=1}^n u[j] - 1)/n
for (int i = n - 1; i >= 1; i--) {
double tmp = (ucsum[i] - 1) / (n - i);
if (tmp >= u[idxs[i - 1]]) {
t = tmp;
break;
}
}
// 4) Return max(u - t*, 0) as projection of u onto simplex
double[] x = new double[u.length];
for (int i = 0; i < u.length; i++) x[i] = Math.max(u[i] - t, 0);
return x;
default:
throw new RuntimeException("Unknown regularization function " + regularization);
}
}
// Project X,Y matrices into appropriate subspace so regularizer is finite. Used during
// initialization.
public final double[] project_x(double[] u, Random rand) {
return project(u, _regularization_x, rand);
}
public final double[] project_y(double[] u, Random rand) {
return project(u, _regularization_y, rand);
}
public final double[] project(double[] u, Regularizer regularization, Random rand) {
if (u == null) return u;
switch (regularization) {
// Domain is all real numbers
case None:
case Quadratic:
case L2:
case L1:
return u;
// Proximal operator of indicator function for a set C is (Euclidean) projection onto C
case NonNegative:
case OneSparse:
case UnitOneSparse:
return rproxgrad(u, 1, 1, regularization, rand);
case Simplex:
double reg =
regularize(
u,
regularization); // Check if inside simplex before projecting since algo is
// complicated
if (reg == 0) return u;
return rproxgrad(u, 1, 1, regularization, rand);
default:
throw new RuntimeException("Unknown regularization function " + regularization);
}
}
// \hat A_{i,j} = \argmin_a L_{i,j}(x_iy_j, a): Data imputation for real numeric values
public final double impute(double u) {
return impute(u, _loss);
}
public static double impute(double u, Loss loss) {
assert loss.isForNumeric() : "Loss function " + loss + " not applicable to numerics";
switch (loss) {
case Quadratic:
case Absolute:
case Huber:
case Periodic:
return u;
case Poisson:
return Math.exp(u) - 1;
case Hinge:
case Logistic:
return u > 0 ? 1 : 0; // Booleans are coded as {0,1} instead of {-1,1}
default:
throw new RuntimeException("Unknown loss function " + loss);
}
}
// \hat A_{i,j} = \argmin_a L_{i,j}(x_iy_j, a): Data imputation for categorical values
// {0,1,2,...}
// TODO: Is there a faster way to find the loss minimizer?
public final int mimpute(double[] u) {
return mimpute(u, _multi_loss);
}
public static int mimpute(double[] u, Loss multi_loss) {
assert multi_loss.isForCategorical()
: "Loss function " + multi_loss + " not applicable to categoricals";
switch (multi_loss) {
case Categorical:
case Ordinal:
double[] cand = new double[u.length];
for (int a = 0; a < cand.length; a++) cand[a] = mloss(u, a, multi_loss);
return ArrayUtils.minIndex(cand);
default:
throw new RuntimeException("Unknown multidimensional loss function " + multi_loss);
}
}
}
