hex.deeplearning.Neurons Maven / Gradle / Ivy
package hex.deeplearning;
import hex.DataInfo;
import water.*;
import water.fvec.Chunk;
import water.fvec.FileVec;
import water.fvec.Frame;
import water.fvec.Vec;
import static water.fvec.Vec.makeCon;
import water.util.ArrayUtils;
import water.util.MathUtils;
import water.util.RandomUtils;
import java.awt.geom.RoundRectangle2D;
import java.nio.ByteBuffer;
import java.util.*;
/**
* This class implements the concept of a Neuron layer in a Neural Network
* During training, every MRTask F/J thread is expected to create these neurons for every map call (Cheap to make).
* These Neurons are NOT sent over the wire.
* The weights connecting the neurons are in a separate class (DeepLearningModel.DeepLearningModelInfo), and will be shared per node.
*/
public abstract class Neurons {
protected int units;
/**
* Constructor of a Neuron Layer
* @param units How many neurons are in this layer?
*/
Neurons(int units) {
this.units = units;
}
/**
* Print the status of this neuron layer
* @return populated String
*/
@Override
public String toString() {
String s = this.getClass().getSimpleName();
s += "\nNumber of Neurons: " + units;
s += "\nParameters:\n" + params.toString();
if (_dropout != null) s += "\nDropout:\n" + _dropout.toString();
return s;
}
/**
* Parameters (deep-cloned() from the user input, can be modified here, e.g. learning rate decay)
*/
protected transient DeepLearningModel.DeepLearningParameters params;
protected transient int _index; //which hidden layer it is
/**
* Layer state (one per neuron): activity, error
*/
public transient Vector _a; //can be sparse for input layer
public transient DenseVector _e;
/**
* References for feed-forward connectivity
*/
public Neurons _previous;
public Neurons _input;
DeepLearningModel.DeepLearningModelInfo _minfo; //reference to shared model info
public Matrix _w;
public DenseVector _b;
/**
* References for momentum training
*/
Matrix _wm;
DenseVector _bm;
/**
* References for ADADELTA
*/
Matrix _ada_dx_g;
DenseVector _bias_ada_dx_g;
/**
* For Dropout training
*/
protected Dropout _dropout;
/**
* Helper to shortcut bprop
*/
private boolean _shortcut = false;
public DenseVector _avg_a;
public static final int missing_int_value = Integer.MAX_VALUE; //encode missing label
public static final Float missing_real_value = Float.NaN; //encode missing regression target
/**
* Helper to check sanity of Neuron layers
* @param training whether training or testing is done
*/
void sanityCheck(boolean training) {
if (this instanceof Input) {
assert(_previous == null);
assert (!training || _dropout != null);
} else {
assert(_previous != null);
if (_minfo.has_momenta()) {
assert(_wm != null);
assert(_bm != null);
assert(_ada_dx_g == null);
}
if (_minfo.adaDelta()) {
if (params._rho == 0) throw new IllegalArgumentException("rho must be > 0 if epsilon is >0.");
if (params._epsilon == 0) throw new IllegalArgumentException("epsilon must be > 0 if rho is >0.");
assert(_minfo.adaDelta());
assert(_bias_ada_dx_g != null);
assert(_wm == null);
assert(_bm == null);
}
if (this instanceof MaxoutDropout || this instanceof TanhDropout || this instanceof RectifierDropout) {
assert (!training || _dropout != null);
}
}
}
/**
* Initialization of the parameters and connectivity of a Neuron layer
* @param neurons Array of all neuron layers, to establish feed-forward connectivity
* @param index Which layer am I?
* @param p User-given parameters (Job parental object hierarchy is not used)
* @param minfo Model information (weights/biases and their momenta)
* @param training Whether training is done or just testing (no need for dropout)
*/
public final void init(Neurons[] neurons, int index, DeepLearningModel.DeepLearningParameters p, final DeepLearningModel.DeepLearningModelInfo minfo, boolean training) {
_index = index-1;
params = (DeepLearningModel.DeepLearningParameters)p.clone();
params._hidden_dropout_ratios = minfo.get_params()._hidden_dropout_ratios;
params._rate *= Math.pow(params._rate_decay, index-1);
_a = new DenseVector(units);
if (!(this instanceof Output) && !(this instanceof Input)) {
_e = new DenseVector(units);
}
if (training && (this instanceof MaxoutDropout || this instanceof TanhDropout
|| this instanceof RectifierDropout || this instanceof Input) ) {
_dropout = this instanceof Input ? new Dropout(units, params._input_dropout_ratio) : new Dropout(units, params._hidden_dropout_ratios[_index]);
}
if (!(this instanceof Input)) {
_previous = neurons[_index]; //incoming neurons
_minfo = minfo;
_w = minfo.get_weights(_index); //incoming weights
_b = minfo.get_biases(_index); //bias for this layer (starting at hidden layer)
if(params._autoencoder && params._sparsity_beta > 0 && _index < params._hidden.length) {
_avg_a = minfo.get_avg_activations(_index);
}
if (minfo.has_momenta()) {
_wm = minfo.get_weights_momenta(_index); //incoming weights
_bm = minfo.get_biases_momenta(_index); //bias for this layer (starting at hidden layer)
}
if (minfo.adaDelta()) {
_ada_dx_g = minfo.get_ada_dx_g(_index);
_bias_ada_dx_g = minfo.get_biases_ada_dx_g(_index);
}
_shortcut = (params._fast_mode || (
// not doing fast mode, but also don't have anything else to update (neither momentum nor ADADELTA history), and no L1/L2
!params._adaptive_rate && !_minfo.has_momenta() && params._l1 == 0.0 && params._l2 == 0.0));
}
sanityCheck(training);
}
/**
* Forward propagation
* @param seed For seeding the RNG inside (for dropout)
* @param training Whether training is done or just testing (no need for dropout)
*/
protected abstract void fprop(long seed, boolean training);
/**
* Back propagation
*/
protected abstract void bprop();
void bprop_sparse(float r, float m) {
SparseVector prev_a = (SparseVector) _previous._a;
int start = prev_a.begin()._idx;
int end = prev_a.end()._idx;
for (int it = start; it < end; ++it) {
final int col = prev_a._indices[it];
final float previous_a = prev_a._values[it];
bprop_col(col, previous_a, r, m);
}
final int rows = _a.size();
final float max_w2 = params._max_w2;
for (int row = 0; row < rows; row++) {
if (max_w2 != Float.POSITIVE_INFINITY)
rescale_weights(_w, row, max_w2);
}
}
/**
* Backpropagation: w -= rate * dE/dw, where dE/dw = dE/dy * dy/dnet * dnet/dw
* This method adds the dnet/dw = activation term per unit
* @param row row index (update weights feeding to this neuron)
* @param partial_grad partial derivative dE/dnet = dE/dy * dy/net
* @param rate learning rate
* @param momentum momentum factor (needed only if ADADELTA isn't used)
*/
final void bprop(final int row, final float partial_grad, final float rate, final float momentum) {
// only correct weights if the gradient is large enough
if (_shortcut && partial_grad == 0f) return;
if (_w instanceof DenseRowMatrix && _previous._a instanceof DenseVector)
bprop_dense_row_dense(
(DenseRowMatrix) _w, (DenseRowMatrix) _wm, (DenseRowMatrix) _ada_dx_g,
(DenseVector) _previous._a, _previous._e, _b, _bm, row, partial_grad, rate, momentum);
else if (_w instanceof DenseRowMatrix && _previous._a instanceof SparseVector)
bprop_dense_row_sparse(
(DenseRowMatrix)_w, (DenseRowMatrix)_wm, (DenseRowMatrix)_ada_dx_g,
(SparseVector)_previous._a, _previous._e, _b, _bm, row, partial_grad, rate, momentum);
else
throw new UnsupportedOperationException("bprop for types not yet implemented.");
}
final void bprop_col(final int col, final float previous_a, final float rate, final float momentum) {
if (_w instanceof DenseColMatrix && _previous._a instanceof SparseVector)
bprop_dense_col_sparse(
(DenseColMatrix)_w, (DenseColMatrix)_wm, (DenseColMatrix)_ada_dx_g,
(SparseVector)_previous._a, _previous._e, _b, _bm, col, previous_a, rate, momentum);
else
throw new UnsupportedOperationException("bprop_col for types not yet implemented.");
}
/**
* Specialization of backpropagation for DenseRowMatrices and DenseVectors
* @param _w weight matrix
* @param _wm weight momentum matrix
* @param adaxg ADADELTA matrix (2 floats per weight)
* @param prev_a activation of previous layer
* @param prev_e error of previous layer
* @param _b bias vector
* @param _bm bias momentum vector
* @param row index of the neuron for which we back-propagate
* @param partial_grad partial derivative dE/dnet = dE/dy * dy/net
* @param rate learning rate
* @param momentum momentum factor (needed only if ADADELTA isn't used)
*/
private void bprop_dense_row_dense(
final DenseRowMatrix _w, final DenseRowMatrix _wm, final DenseRowMatrix adaxg,
final DenseVector prev_a, final DenseVector prev_e, final DenseVector _b, final DenseVector _bm,
final int row, final float partial_grad, float rate, final float momentum)
{
final float rho = (float)params._rho;
final float eps = (float)params._epsilon;
final float l1 = (float)params._l1;
final float l2 = (float)params._l2;
final float max_w2 = params._max_w2;
final boolean have_momenta = _minfo.has_momenta();
final boolean have_ada = _minfo.adaDelta();
final boolean nesterov = params._nesterov_accelerated_gradient;
final boolean update_prev = prev_e != null;
final boolean fast_mode = params._fast_mode;
final int cols = prev_a.size();
final int idx = row * cols;
float avg_grad2 = 0;
for( int col = 0; col < cols; col++ ) {
final float weight = _w.get(row,col);
if( update_prev ) prev_e.add(col, partial_grad * weight); // propagate the error dE/dnet to the previous layer, via connecting weights
final float previous_a = prev_a.get(col);
if (fast_mode && previous_a == 0) continue;
//this is the actual gradient dE/dw
final float grad = partial_grad * previous_a - Math.signum(weight) * l1 - weight * l2;
final int w = idx + col;
if (have_ada) {
final float grad2 = grad*grad;
avg_grad2 += grad2;
float brate = computeAdaDeltaRateForWeight(grad, w, adaxg, rho, eps);
_w.raw()[w] += brate * grad;
} else {
if (!nesterov) {
final float delta = rate * grad;
_w.raw()[w] += delta;
if( have_momenta ) {
_w.raw()[w] += momentum * _wm.raw()[w];
_wm.raw()[w] = delta;
}
} else {
float tmp = grad;
if( have_momenta ) {
_wm.raw()[w] *= momentum;
_wm.raw()[w] += tmp;
tmp = _wm.raw()[w];
}
_w.raw()[w] += rate * tmp;
}
}
}
if (max_w2 != Float.POSITIVE_INFINITY)
rescale_weights(_w, row, max_w2);
if (have_ada) avg_grad2 /= cols;
update_bias(_b, _bm, row, partial_grad, avg_grad2, rate, momentum);
}
/**
* Specialization of backpropagation for DenseColMatrices and SparseVector for previous layer's activation and DenseVector for everything else
* @param w Weight matrix
* @param wm Momentum matrix
* @param adaxg ADADELTA matrix (2 floats per weight)
* @param prev_a sparse activation of previous layer
* @param prev_e error of previous layer
* @param b bias
* @param bm bias momentum
* @param rate learning rate
* @param momentum momentum factor (needed only if ADADELTA isn't used)
*/
private void bprop_dense_col_sparse(
final DenseColMatrix w, final DenseColMatrix wm, final DenseColMatrix adaxg,
final SparseVector prev_a, final DenseVector prev_e, final DenseVector b, final DenseVector bm,
final int col, final float previous_a, float rate, final float momentum)
{
final float rho = (float)params._rho;
final float eps = (float)params._epsilon;
final float l1 = (float)params._l1;
final float l2 = (float)params._l2;
final boolean have_momenta = _minfo.has_momenta();
final boolean have_ada = _minfo.adaDelta();
final boolean nesterov = params._nesterov_accelerated_gradient;
final boolean update_prev = prev_e != null;
final int cols = prev_a.size();
final int rows = _a.size();
for (int row = 0; row < rows; row++) {
final float partial_grad = _e.get(row) * (1f - _a.get(row) * _a.get(row));
final float weight = w.get(row,col);
if( update_prev ) prev_e.add(col, partial_grad * weight); // propagate the error dE/dnet to the previous layer, via connecting weights
assert (previous_a != 0); //only iterate over non-zeros!
if (_shortcut && partial_grad == 0f) continue;
//this is the actual gradient dE/dw
final float grad = partial_grad * previous_a - Math.signum(weight) * l1 - weight * l2;
if (have_ada) {
assert(!have_momenta);
float brate = computeAdaDeltaRateForWeight(grad, row, col, adaxg, rho, eps);
w.add(row,col, brate * grad);
} else {
if (!nesterov) {
final float delta = rate * grad;
w.add(row, col, delta);
// Log.info("for row = " + row + ", col = " + col + ", partial_grad = " + partial_grad + ", grad = " + grad);
if( have_momenta ) {
w.add(row, col, momentum * wm.get(row, col));
wm.set(row, col, delta);
}
} else {
float tmp = grad;
if( have_momenta ) {
float val = wm.get(row, col);
val *= momentum;
val += tmp;
tmp = val;
wm.set(row, col, val);
}
w.add(row, col, rate * tmp);
}
}
//this is called cols times, so we divide the (repeated) contribution by 1/cols
update_bias(b, bm, row, partial_grad/cols, grad*grad/cols, rate, momentum);
}
}
/**
* Specialization of backpropagation for DenseRowMatrices and SparseVector for previous layer's activation and DenseVector for everything else
* @param _w weight matrix
* @param _wm weight momentum matrix
* @param adaxg ADADELTA matrix (2 floats per weight)
* @param prev_a sparse activation of previous layer
* @param prev_e error of previous layer
* @param _b bias vector
* @param _bm bias momentum vector
* @param row index of the neuron for which we back-propagate
* @param partial_grad partial derivative dE/dnet = dE/dy * dy/net
* @param rate learning rate
* @param momentum momentum factor (needed only if ADADELTA isn't used)
*/
private void bprop_dense_row_sparse(
final DenseRowMatrix _w, final DenseRowMatrix _wm, final DenseRowMatrix adaxg,
final SparseVector prev_a, final DenseVector prev_e, final DenseVector _b, final DenseVector _bm,
final int row, final float partial_grad, float rate, final float momentum)
{
final float rho = (float)params._rho;
final float eps = (float)params._epsilon;
final float l1 = (float)params._l1;
final float l2 = (float)params._l2;
final float max_w2 = params._max_w2;
final boolean have_momenta = _minfo.has_momenta();
final boolean have_ada = _minfo.adaDelta();
final boolean nesterov = params._nesterov_accelerated_gradient;
final boolean update_prev = prev_e != null;
final int cols = prev_a.size();
final int idx = row * cols;
float avg_grad2 = 0;
int start = prev_a.begin()._idx;
int end = prev_a.end()._idx;
for (int it = start; it < end; ++it) {
final int col = prev_a._indices[it];
final float weight = _w.get(row,col);
if( update_prev ) prev_e.add(col, partial_grad * weight); // propagate the error dE/dnet to the previous layer, via connecting weights
final float previous_a = prev_a._values[it];
assert (previous_a != 0); //only iterate over non-zeros!
//this is the actual gradient dE/dw
final float grad = partial_grad * previous_a - Math.signum(weight) * l1 - weight * l2;
final int w = idx + col;
if (have_ada) {
assert(!have_momenta);
final float grad2 = grad*grad;
avg_grad2 += grad2;
float brate = computeAdaDeltaRateForWeight(grad, w, adaxg, rho, eps);
_w.raw()[w] += brate * grad;
} else {
if (!nesterov) {
final float delta = rate * grad;
_w.raw()[w] += delta;
if( have_momenta ) {
_w.raw()[w] += momentum * _wm.raw()[w];
_wm.raw()[w] = delta;
}
} else {
float tmp = grad;
if( have_momenta ) {
_wm.raw()[w] *= momentum;
_wm.raw()[w] += tmp;
tmp = _wm.raw()[w];
}
_w.raw()[w] += rate * tmp;
}
}
}
if (max_w2 != Float.POSITIVE_INFINITY)
rescale_weights(_w, row, max_w2);
if (have_ada) avg_grad2 /= prev_a.nnz();
update_bias(_b, _bm, row, partial_grad, avg_grad2, rate, momentum);
}
/**
* Helper to scale down incoming weights if their squared sum exceeds a given value (by a factor of 10 -> to avoid doing costly rescaling too often)
* C.f. Improving neural networks by preventing co-adaptation of feature detectors
* @param row index of the neuron for which to scale the weights
*/
private static void rescale_weights(final Matrix w, final int row, final float max_w2) {
final int cols = w.cols();
if (w instanceof DenseRowMatrix) {
rescale_weights((DenseRowMatrix)w, row, max_w2);
} else if (w instanceof DenseColMatrix) {
float r2 = 0;
for (int col=0; col max_w2) {
final float scale = MathUtils.approxSqrt(max_w2 / r2);
for( int col=0; col < cols; col++ ) w.set(row, col, w.get(row,col) * scale);
}
}
else throw new UnsupportedOperationException("rescale weights for " + w.getClass().getSimpleName() + " not yet implemented.");
}
// Specialization for DenseRowMatrix
private static void rescale_weights(final DenseRowMatrix w, final int row, final float max_w2) {
final int cols = w.cols();
final int idx = row * cols;
float r2 = MathUtils.sumSquares(w.raw(), idx, idx + cols);
// float r2 = MathUtils.approxSumSquares(w.raw(), idx, idx + cols);
if( r2 > max_w2) {
final float scale = MathUtils.approxSqrt(max_w2 / r2);
for( int c = 0; c < cols; c++ ) w.raw()[idx + c] *= scale;
}
}
/**
* Helper to compute the reconstruction error for auto-encoders (part of the gradient computation)
* @param row neuron index
* @return difference between the output (auto-encoder output layer activation) and the target (input layer activation)
*/
protected float autoEncoderError(int row) {
assert (_minfo.get_params()._autoencoder && _index == _minfo.get_params()._hidden.length);
final float t = _input._a.get(row);
final float y = _a.get(row);
float g;
if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.MeanSquare) {
g = t - y;
} else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.Absolute) {
g = y > t ? -1f : 1f;
}
// Huber:
// L = (y-t)^2 for |t-y| < 1, -dL/dy = t - y
// L = 2*|t-y|-1 for |t-y| >= 1, -dL/dy = +/- 2
else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.Huber) {
if (Math.abs(y-t) < 1) {
g = t - y;
} else {
g = y >= t + 1f ? -2f : 2f;
}
} else throw H2O.unimpl("Loss " + params._loss + " not implemented for Auto-Encoder.");
return g;
}
/**
* Compute learning rate with AdaDelta
* http://www.matthewzeiler.com/pubs/googleTR2012/googleTR2012.pdf
* @param grad gradient
* @param row which neuron is to be updated
* @param col weight from which incoming neuron
* @param ada_dx_g Matrix holding helper values (2 floats per weight)
* @param rho hyper-parameter #1
* @param eps hyper-parameter #2
* @return learning rate
*/
private static float computeAdaDeltaRateForWeight(final float grad, final int row, final int col,
final DenseColMatrix ada_dx_g,
final float rho, final float eps) {
ada_dx_g.set(2*row+1, col, rho * ada_dx_g.get(2*row+1, col) + (1f - rho) * grad * grad);
final float rate = MathUtils.approxSqrt((ada_dx_g.get(2*row, col) + eps)/(ada_dx_g.get(2*row+1, col) + eps));
ada_dx_g.set(2*row, col, rho * ada_dx_g.get(2*row, col) + (1f - rho) * rate * rate * grad * grad);
return rate;
}
/**
* Compute learning rate with AdaDelta, specialized for DenseRowMatrix
* @param grad gradient
* @param w neuron index
* @param ada_dx_g Matrix holding helper values (2 floats per weight)
* @param rho hyper-parameter #1
* @param eps hyper-parameter #2
* @return learning rate
*/
private static float computeAdaDeltaRateForWeight(final float grad, final int w,
final DenseRowMatrix ada_dx_g,
final float rho, final float eps) {
ada_dx_g.raw()[2*w+1] = rho * ada_dx_g.raw()[2*w+1] + (1f - rho) * grad * grad;
final float rate = MathUtils.approxSqrt((ada_dx_g.raw()[2*w] + eps)/(ada_dx_g.raw()[2*w+1] + eps));
ada_dx_g.raw()[2*w] = rho * ada_dx_g.raw()[2*w] + (1f - rho) * rate * rate * grad * grad;
return rate;
}
/**
* Compute learning rate with AdaDelta, specialized for DenseVector (Bias)
* @param grad2 squared gradient
* @param row neuron index
* @param bias_ada_dx_g Matrix holding helper values (2 floats per weight)
* @param rho hyper-parameter #1
* @param eps hyper-parameter #2
* @return learning rate
*/
private static float computeAdaDeltaRateForBias(final float grad2, final int row,
final DenseVector bias_ada_dx_g,
final float rho, final float eps) {
bias_ada_dx_g.raw()[2*row+1] = rho * bias_ada_dx_g.raw()[2*row+1] + (1f - rho) * grad2;
final float rate = MathUtils.approxSqrt((bias_ada_dx_g.raw()[2*row ] + eps)/(bias_ada_dx_g.raw()[2*row+1] + eps));
bias_ada_dx_g.raw()[2*row] = rho * bias_ada_dx_g.raw()[2*row ] + (1f - rho) * rate * rate * grad2;
return rate;
}
/**
* Helper to enforce learning rule to satisfy sparsity constraint:
* Computes the (rolling) average activation for each (hidden) neuron.
*/
void compute_sparsity() {
if (_avg_a != null) {
for (int row = 0; row < _avg_a.size(); row++) {
_avg_a.set(row, (float) 0.999 * (_avg_a.get(row)) + (float) 0.001 * (_a.get(row)));
}
}
}
/**
* Helper to update the bias values
* @param _b bias vector
* @param _bm bias momentum vector
* @param row index of the neuron for which we back-propagate
* @param partial_grad partial derivative dE/dnet = dE/dy * dy/net
* @param avg_grad2 average squared gradient for this neuron's incoming weights (only for ADADELTA)
* @param rate learning rate
* @param momentum momentum factor (needed only if ADADELTA isn't used)
*/
void update_bias(final DenseVector _b, final DenseVector _bm, final int row,
float partial_grad, final float avg_grad2, float rate, final float momentum) {
final boolean have_momenta = _minfo.has_momenta();
final boolean have_ada = _minfo.adaDelta();
final float l1 = (float)params._l1;
final float l2 = (float)params._l2;
final float bias = _b.get(row);
partial_grad -= Math.signum(bias) * l1 + bias * l2;
if (have_ada) {
final float rho = (float)params._rho;
final float eps = (float)params._epsilon;
rate = computeAdaDeltaRateForBias(avg_grad2, row, _bias_ada_dx_g, rho, eps);
}
if (!params._nesterov_accelerated_gradient) {
final float delta = rate * partial_grad;
_b.add(row, delta);
if (have_momenta) {
_b.add(row, momentum * _bm.get(row));
_bm.set(row, delta);
}
} else {
float d = partial_grad;
if (have_momenta) {
_bm.set(row, _bm.get(row) * momentum);
_bm.add(row, d);
d = _bm.get(row);
}
_b.add(row, rate * d);
}
//update for sparsity constraint
if (params._autoencoder && params._sparsity_beta > 0 && !(this instanceof Output) && !(this instanceof Input) && (_index != params._hidden.length)) {
_b.add(row, -(float) (rate * params._sparsity_beta * (_avg_a._data[row] - params._average_activation)));
}
if (Float.isInfinite(_b.get(row))) _minfo.set_unstable();
}
/**
* The learning rate
* @param n The number of training samples seen so far (for rate_annealing greater than 0)
* @return Learning rate
*/
public float rate(long n) {
return (float)(params._rate / (1 + params._rate_annealing * n));
}
protected float momentum() {
return momentum(-1);
}
/**
* The momentum - real number in [0, 1)
* Can be a linear ramp from momentum_start to momentum_stable, over momentum_ramp training samples
* @param n The number of training samples seen so far
* @return momentum
*/
public float momentum(long n) {
double m = params._momentum_start;
if( params._momentum_ramp > 0 ) {
final long num = n != -1 ? _minfo.get_processed_total() : n;
if( num >= params._momentum_ramp)
m = params._momentum_stable;
else
m += (params._momentum_stable - params._momentum_start) * num / params._momentum_ramp;
}
return (float)m;
}
/**
* Input layer of the Neural Network
* This layer is different from other layers as it has no incoming weights,
* but instead gets its activation values from the training points.
*/
public static class Input extends Neurons {
private DataInfo _dinfo; //training data
SparseVector _svec;
DenseVector _dvec;
Input(int units, final DataInfo d) {
super(units);
_dinfo = d;
_a = new DenseVector(units);
_dvec = (DenseVector)_a;
}
@Override protected void bprop() { throw new UnsupportedOperationException(); }
@Override protected void fprop(long seed, boolean training) { throw new UnsupportedOperationException(); }
/**
* One of two methods to set layer input values. This one is for raw double data, e.g. for scoring
* @param seed For seeding the RNG inside (for input dropout)
* @param data Data (training columns and responses) to extract the training columns
* from to be mapped into the input neuron layer
*/
public void setInput(long seed, final double[] data) {
// Log.info("Data: " + ArrayUtils.toString(data));
assert(_dinfo != null);
double [] nums = MemoryManager.malloc8d(_dinfo._nums); // a bit wasteful - reallocated each time
int [] cats = MemoryManager.malloc4(_dinfo._cats); // a bit wasteful - reallocated each time
int i = 0, ncats = 0;
for(; i < _dinfo._cats; ++i){
assert(_dinfo._catMissing[i] != 0); //we now *always* have a categorical level for NAs, just in case.
if (Double.isNaN(data[i])) {
cats[ncats] = (_dinfo._catOffsets[i+1]-1); //use the extra level for NAs made during training
} else {
int c = (int)data[i];
if (_dinfo._useAllFactorLevels)
cats[ncats] = c + _dinfo._catOffsets[i];
else if (c!=0)
cats[ncats] = c + _dinfo._catOffsets[i] - 1;
// If factor level in test set was not seen by training, then turn it into an NA
if (cats[ncats] >= _dinfo._catOffsets[i+1]) {
cats[ncats] = (_dinfo._catOffsets[i+1]-1);
}
}
ncats++;
}
final int n = data.length; // data contains only input features - no response is included
for(;i < n;++i){
double d = data[i];
if(_dinfo._normMul != null) d = (d - _dinfo._normSub[i-_dinfo._cats])*_dinfo._normMul[i-_dinfo._cats];
nums[i-_dinfo._cats] = d; //can be NaN for missing numerical data
}
setInput(seed, nums, ncats, cats);
}
/**
* The second method used to set input layer values. This one is used directly by FrameTask.processRow() and by the method above.
* @param seed For seeding the RNG inside (for input dropout)
* @param nums Array containing numerical values, can be NaN
* @param numcat Number of horizontalized categorical non-zero values (i.e., those not being the first factor of a class)
* @param cats Array of indices, the first numcat values are the input layer unit (==column) indices for the non-zero categorical values
* (This allows this array to be re-usable by the caller, without re-allocating each time)
*/
public void setInput(long seed, final double[] nums, final int numcat, final int[] cats) {
_a = _dvec;
Arrays.fill(_a.raw(), 0f);
// random projection from fullN down to max_categorical_features
if (params._max_categorical_features < _dinfo.fullN() - _dinfo._nums) {
assert(nums.length == _dinfo._nums);
final int M = nums.length + params._max_categorical_features;
final boolean random_projection = false;
final boolean hash_trick = true;
if (random_projection) {
final int N = _dinfo.fullN();
assert (_a.size() == M);
// sparse random projection
for (int i = 0; i < M; ++i) {
for (int c = 0; c < numcat; ++c) {
int j = cats[c];
Random rng = RandomUtils.getRNG(params._seed + i * N + j); //TODO: re-use same pool of random numbers with some hashing
float val = 0;
final float rnd = rng.nextFloat();
if (rnd < 1. / 6.) val = (float) Math.sqrt(3);
if (rnd > 5. / 6.) val = -(float) Math.sqrt(3);
_a.add(i, 1f * val);
}
Random rng = RandomUtils.getRNG(params._seed + i*N + _dinfo.numStart());
for (int n = 0; n < nums.length; ++n) {
float val = 0;
final float rnd = rng.nextFloat();
if (rnd < 1. / 6.) val = (float) Math.sqrt(3);
if (rnd > 5. / 6.) val = -(float) Math.sqrt(3);
_a.set(i, (Double.isNaN(nums[n]) ? 0f /*Always do MeanImputation during scoring*/ : (float) nums[n]) * val);
}
}
} else if (hash_trick) {
// Use hash trick for categorical features
assert (_a.size() == M);
// hash N-nums.length down to M-nums.length = cM (#categorical slots - always use all numerical features)
final int cM = params._max_categorical_features;
assert (_a.size() == M);
MurmurHash murmur = MurmurHash.getInstance();
for (int i = 0; i < numcat; ++i) {
ByteBuffer buf = ByteBuffer.allocate(4);
int hashval = murmur.hash(buf.putInt(cats[i]).array(), 4, (int)params._seed); // turn horizontalized categorical integer into another integer, based on seed
_a.add(Math.abs(hashval % cM), 1f); // restrict to limited range
}
for (int i = 0; i < nums.length; ++i)
_a.set(cM + i, Double.isNaN(nums[i]) ? 0f /*Always do MeanImputation during scoring*/ : (float) nums[i]);
}
} else {
assert(_a.size() == _dinfo.fullN());
for (int i = 0; i < numcat; ++i) _a.set(cats[i], 1f); // one-hot encode categoricals
for (int i = 0; i < nums.length; ++i)
_a.set(_dinfo.numStart() + i, Double.isNaN(nums[i]) ? 0f /*Always do MeanImputation during scoring*/ : (float) nums[i]);
}
// Input Dropout
if (_dropout == null) return;
seed += params._seed + 0x1337B4BE;
_dropout.randomlySparsifyActivation(_a, seed);
if (params._sparse) {
_svec = new SparseVector(_dvec);
_a = _svec;
}
}
}
/**
* Tanh neurons - most common, most stable
*/
public static class Tanh extends Neurons {
public Tanh(int units) { super(units); }
@Override protected void fprop(long seed, boolean training) {
gemv((DenseVector)_a, _w, _previous._a, _b, _dropout != null ? _dropout.bits() : null);
final int rows = _a.size();
for( int row = 0; row < rows; row++ )
_a.set(row, 1f - 2f / (1f + (float)Math.exp(2*_a.get(row)))); //evals faster than tanh(x), but is slightly less numerically stable - OK
compute_sparsity();
}
// Computing partial derivative g = dE/dnet = dE/dy * dy/dnet, where dE/dy is the backpropagated error
// dy/dnet = (1 - a^2) for y(net) = tanh(net)
@Override protected void bprop() {
float m = momentum();
float r = _minfo.adaDelta() ? 0 : rate(_minfo.get_processed_total()) * (1f - m);
if (_w instanceof DenseRowMatrix) {
final int rows = _a.size();
for (int row = 0; row < rows; row++) {
if (_minfo.get_params()._autoencoder && _index == _minfo.get_params()._hidden.length)
_e.set(row, autoEncoderError(row));
float g = _e.get(row) * (1f - _a.get(row) * _a.get(row));
bprop(row, g, r, m);
}
}
else {
bprop_sparse(r, m);
}
}
}
/**
* Tanh neurons with dropout
*/
public static class TanhDropout extends Tanh {
public TanhDropout(int units) { super(units); }
@Override protected void fprop(long seed, boolean training) {
if (training) {
seed += params._seed + 0xDA7A6000;
_dropout.fillBytes(seed);
super.fprop(seed, true);
}
else {
super.fprop(seed, false);
ArrayUtils.mult(_a.raw(), (float)(1-params._hidden_dropout_ratios[_index]));
}
}
}
/**
* Maxout neurons
*/
public static class Maxout extends Neurons {
public Maxout(int units) { super(units); }
@Override protected void fprop(long seed, boolean training) {
float max = 0;
final int rows = _a.size();
if (_previous._a instanceof DenseVector) {
for( int row = 0; row < rows; row++ ) {
_a.set(row, 0);
if( !training || _dropout == null || _dropout.unit_active(row) ) {
_a.set(row, Float.NEGATIVE_INFINITY);
for( int i = 0; i < _previous._a.size(); i++ )
_a.set(row, Math.max(_a.get(row), _w.get(row, i) * _previous._a.get(i)));
if (Float.isInfinite(-_a.get(row))) _a.set(row, 0); //catch the case where there is dropout (and/or input sparsity) -> no max found!
_a.add(row, _b.get(row));
max = Math.max(_a.get(row), max);
}
}
if( max > 1 ) ArrayUtils.div(_a.raw(), max);
}
else {
SparseVector x = (SparseVector)_previous._a;
for( int row = 0; row < _a.size(); row++ ) {
_a.set(row, 0);
if( !training || _dropout == null || _dropout.unit_active(row) ) {
// _a.set(row, Float.NEGATIVE_INFINITY);
// for( int i = 0; i < _previous._a.size(); i++ )
// _a.set(row, Math.max(_a.get(row), _w.get(row, i) * _previous._a.get(i)));
float mymax = Float.NEGATIVE_INFINITY;
int start = x.begin()._idx;
int end = x.end()._idx;
for (int it = start; it < end; ++it) {
mymax = Math.max(mymax, _w.get(row, x._indices[it]) * x._values[it]);
}
_a.set(row, mymax);
if (Float.isInfinite(-_a.get(row))) _a.set(row, 0); //catch the case where there is dropout (and/or input sparsity) -> no max found!
_a.add(row, _b.get(row));
max = Math.max(_a.get(row), max);
}
}
if( max > 1f ) ArrayUtils.div(_a.raw(), max);
}
compute_sparsity();
}
@Override protected void bprop() {
float m = momentum();
float r = _minfo.adaDelta() ? 0 : rate(_minfo.get_processed_total()) * (1f - m);
if (_w instanceof DenseRowMatrix) {
final int rows = _a.size();
for( int row = 0; row < rows; row++ ) {
assert (!_minfo.get_params()._autoencoder);
// if (_minfo.get_params().autoencoder && _index == _minfo.get_params().hidden.length)
// _e.set(row, autoEncoderError(row));
float g = _e.get(row);
// if( _a[o] < 0 ) Not sure if we should be using maxout with a hard zero bottom
// g = 0;
bprop(row, g, r, m);
}
}
else {
bprop_sparse(r, m);
}
}
}
/**
* Maxout neurons with dropout
*/
public static class MaxoutDropout extends Maxout {
public MaxoutDropout(int units) { super(units); }
@Override protected void fprop(long seed, boolean training) {
if (training) {
seed += params._seed + 0x51C8D00D;
_dropout.fillBytes(seed);
super.fprop(seed, true);
}
else {
super.fprop(seed, false);
ArrayUtils.mult(_a.raw(), (float)(1-params._hidden_dropout_ratios[_index]));
}
}
}
/**
* Rectifier linear unit (ReLU) neurons
*/
public static class Rectifier extends Neurons {
public Rectifier(int units) { super(units); }
@Override protected void fprop(long seed, boolean training) {
gemv((DenseVector)_a, _w, _previous._a, _b, _dropout != null ? _dropout.bits() : null);
final int rows = _a.size();
for( int row = 0; row < rows; row++ ) {
_a.set(row, Math.max(_a.get(row), 0f));
compute_sparsity();
}
}
@Override protected void bprop() {
float m = momentum();
float r = _minfo.adaDelta() ? 0 : rate(_minfo.get_processed_total()) * (1f - m);
final int rows = _a.size();
if (_w instanceof DenseRowMatrix) {
for (int row = 0; row < rows; row++) {
if (_minfo.get_params()._autoencoder && _index == _minfo.get_params()._hidden.length)
_e.set(row, autoEncoderError(row));
//(d/dx)(max(0,x)) = 1 if x > 0, otherwise 0
float g = _a.get(row) > 0f ? _e.get(row) : 0f;
bprop(row, g, r, m);
}
}
else {
bprop_sparse(r, m);
}
}
}
/**
* Rectifier linear unit (ReLU) neurons with dropout
*/
public static class RectifierDropout extends Rectifier {
public RectifierDropout(int units) { super(units); }
@Override protected void fprop(long seed, boolean training) {
if (training) {
seed += params._seed + 0x3C71F1ED;
_dropout.fillBytes(seed);
super.fprop(seed, true);
}
else {
super.fprop(seed, false);
ArrayUtils.mult(_a.raw(), (float)(1-params._hidden_dropout_ratios[_index]));
}
}
}
/**
* Abstract class for Output neurons
*/
public static abstract class Output extends Neurons {
Output(int units) { super(units); }
protected void fprop(long seed, boolean training) { throw new UnsupportedOperationException(); }
protected void bprop() { throw new UnsupportedOperationException(); }
}
/**
* Output neurons for classification - Softmax
*/
public static class Softmax extends Output {
public Softmax(int units) { super(units); }
protected void fprop() {
gemv((DenseVector) _a, (DenseRowMatrix) _w, (DenseVector) _previous._a, _b, null);
final float max = ArrayUtils.maxValue(_a.raw());
float scale = 0f;
final float rows = _a.size();
for( int row = 0; row < rows; row++ ) {
_a.set(row, (float)Math.exp(_a.get(row) - max));
scale += _a.get(row);
}
for( int row = 0; row < rows; row++ ) {
_a.raw()[row] /= scale;
if (Float.isNaN(_a.get(row))) {
_minfo.set_unstable();
throw new RuntimeException("Numerical instability, predicted NaN.");
}
}
}
/**
* Backpropagation for classification
* Update every weight as follows: w += -rate * dE/dw
* Compute dE/dw via chain rule: dE/dw = dE/dy * dy/dnet * dnet/dw, where net = sum(xi*wi)+b and y = activation function
* @param target actual class label
*/
protected void bprop(int target) {
assert (target != missing_int_value); // no correction of weights/biases for missing label
float m = momentum();
float r = _minfo.adaDelta() ? 0 : rate(_minfo.get_processed_total()) * (1f - m);
float g; //partial derivative dE/dy * dy/dnet
final float rows = _a.size();
for( int row = 0; row < rows; row++ ) {
final float t = (row == target ? 1f : 0f);
final float y = _a.get(row);
//dy/dnet = derivative of softmax = (1-y)*y
if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.CrossEntropy) {
//nothing else needed, -dCE/dy * dy/dnet = target - y
//cf. http://www.stanford.edu/group/pdplab/pdphandbook/handbookch6.html
g = t - y;
} else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.Absolute) {
g = (2*t-1) * (1f - y) * y; //-dL/dy = 2*t-1
} else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.MeanSquare) {
//-dMSE/dy = target-y
g = (t - y) * (1f - y) * y;
} else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.Huber) {
if (t==0) {
if (y<0.5) {
g = -4*y; //L=2*y^2 for y<0.5
} else {
g = -2; //L=2*y-0.5 for y>=0.5
}
} else {
if (y>0.5) {
g = 4*(1-y); //L=2*(1-y)^2 for y<0.5
} else {
g = 2; //L=2*(1-y)-0.5 for y>=0.5
}
}
g *= (1f - y) * y;
} else throw H2O.unimpl("Loss " + params._loss + " not implemented for classification.");
// this call expects dE/dnet
bprop(row, g, r, m);
}
}
}
/**
* Output neurons for regression - Linear units
*/
public static class Linear extends Output {
public Linear(int units) { super(units); }
protected void fprop() {
gemv((DenseVector)_a, _w, _previous._a, _b, _dropout != null ? _dropout.bits() : null);
}
/**
* Backpropagation for regression
* @param target floating-point target value
*/
protected void bprop(float target) {
assert (target != missing_real_value);
final int row = 0;
final float t = target;
final float y = _a.get(row);
float g;
// Computing partial derivative: dE/dnet = dE/dy * dy/dnet = dE/dy * 1
if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.MeanSquare) {
g = t - y; //for MSE -dMSE/dy = target-y
}
// L = |y-t|, -dL/dy = -/+1
else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.Absolute) {
g = y > t ? -1f : 1f;
}
// Huber:
// L = (y-t)^2 for |t-y| < 1, -dL/dy = t - y
// L = 2*|t-y|-1 for |t-y| >= 1, -dL/dy = +/- 2
else if (params._loss == DeepLearningModel.DeepLearningParameters.Loss.Huber) {
if (Math.abs(y-t) < 1) {
g = t - y;
} else {
g = y >= t + 1f ? -2f : 2f;
}
} else throw H2O.unimpl("Loss " + params._loss + " not implemented for regression.");
float m = momentum();
float r = _minfo.adaDelta() ? 0 : rate(_minfo.get_processed_total()) * (1f - m);
bprop(row, g, r, m);
}
}
/**
* Mat-Vec Plus Add (with optional row dropout)
* @param res = a*x+y (pre-allocated, will be overwritten)
* @param a matrix of size rows x cols
* @param x vector of length cols
* @param y vector of length rows
* @param row_bits if not null, check bits of this byte[] to determine whether a row is used or not
*/
static void gemv_naive(final float[] res, final float[] a, final float[] x, final float[] y, byte[] row_bits) {
final int cols = x.length;
final int rows = y.length;
assert(res.length == rows);
for(int row = 0; row row = a.row(r);
Set> set = row.entrySet();
for (Map.Entry e : set) {
final float val = x.get(e.getKey());
if (val != 0f) res.add(r, e.getValue() * val); //TODO: iterate over both iterators and only add where there are matching indices
}
res.add(r, y.get(r));
}
}
static void gemv(final DenseVector res, final SparseColMatrix a, final SparseVector x, final DenseVector y, byte[] row_bits) {
final int rows = y.size();
assert(res.size() == rows);
for(int r = 0; r col = a.col(c);
final float val = x.get(c);
if (val == 0f) continue;
for (Map.Entry e : col.entrySet()) {
final int r = e.getKey();
if( row_bits != null && (row_bits[r / 8] & (1 << (r % 8))) == 0) continue;
// iterate over all non-empty columns for this row
res.add(r, e.getValue() * val);
}
}
for(int r = 0; r " + value();
}
float value() { return _values[_idx]; }
int index() { return _indices[_idx]; }
void setValue(float val) { _values[_idx] = val; }
}
public Iterator begin() { return new Iterator(0); }
public Iterator end() { return new Iterator(_indices.length); }
@Override public Frame toFrame(Key key) { return Neurons.toFrame(this, key); }
}
/**
* Abstract matrix interface
*/
public abstract interface Matrix {
abstract float get(int row, int col);
abstract void set(int row, int col, float val);
abstract void add(int row, int col, float val);
abstract int cols();
abstract int rows();
abstract long size();
abstract float[] raw();
public Frame toFrame(Key key);
}
/**
* Helper to convert the Matrix to a Frame using MRTask
*/
private static class FrameFiller extends MRTask {
final DenseColMatrix dcm;
final DenseRowMatrix drm;
final SparseRowMatrix srm;
final SparseColMatrix scm;
FrameFiller(Matrix m) {
if (m instanceof DenseColMatrix) {
dcm = (DenseColMatrix)m;
drm = null;
srm = null;
scm = null;
}
else if (m instanceof DenseRowMatrix) {
dcm = null;
drm = (DenseRowMatrix)m;
srm = null;
scm = null;
}
else if (m instanceof SparseRowMatrix) {
dcm = null;
drm = null;
srm = (SparseRowMatrix)m;
scm = null;
}
else {
dcm = null;
drm = null;
srm = null;
scm = (SparseColMatrix)m;
}
}
@Override public void map(Chunk[] cs) {
Matrix m=null;
if (dcm != null) m = dcm;
if (drm != null) m = drm;
if (scm != null) m = scm;
if (srm != null) m = srm;
int off = (int)cs[0].start();
assert(m.cols() == cs.length);
for (int c = 0; c < cs.length; ++c) {
for (int r = 0; r < cs[0]._len; ++r) {
cs[c].set(r, m.get(off + r, c));
}
}
}
}
/**
* Helper to convert a Matrix into a Frame
* @param m Matrix
* @param key Key for output Frame
* @return Reference to Frame (which is also in DKV)
*/
static Frame toFrame(Matrix m, Key key) {
final int log_rows_per_chunk = Math.max(1, FileVec.DFLT_LOG2_CHUNK_SIZE - (int) Math.floor(Math.log(m.cols()) / Math.log(2.)));
Vec v[] = new Vec[m.cols()];
for (int i = 0; i < m.cols(); ++i) {
v[i] = makeCon(0, m.rows(), log_rows_per_chunk);
}
Frame f = new FrameFiller(m).doAll(new Frame(key, v, true))._fr;
DKV.put(key, f);
return f;
}
/**
* Dense row matrix implementation
*/
public final static class DenseRowMatrix extends Iced implements Matrix {
private float[] _data;
private int _cols;
private int _rows;
DenseRowMatrix(int rows, int cols) { this(new float[cols*rows], rows, cols); }
DenseRowMatrix(float[] v, int rows, int cols) { _data = v; _rows = rows; _cols = cols; }
@Override public float get(int row, int col) {
assert(row<_rows && col<_cols) : "_data.length: " + _data.length + ", checking: " + row + " < " + _rows + " && " + col + " < " + _cols;
return _data[row*_cols + col];
}
@Override public void set(int row, int col, float val) { assert(row<_rows && col<_cols); _data[row*_cols + col] = val; }
@Override public void add(int row, int col, float val) { assert(row<_rows && col<_cols); _data[row*_cols + col] += val; }
@Override public int cols() { return _cols; }
@Override public int rows() { return _rows; }
@Override public long size() { return (long)_rows*(long)_cols; }
public float[] raw() { return _data; }
@Override public Frame toFrame(Key key) { return Neurons.toFrame(this, key); }
}
/**
* Dense column matrix implementation
*/
public final static class DenseColMatrix extends Iced implements Matrix {
private float[] _data;
private int _cols;
private int _rows;
DenseColMatrix(int rows, int cols) { this(new float[cols*rows], rows, cols); }
DenseColMatrix(float[] v, int rows, int cols) { _data = v; _rows = rows; _cols = cols; }
DenseColMatrix(DenseRowMatrix m, int rows, int cols) { this(rows, cols); for (int row=0;row[] _rows;
private int _cols;
SparseRowMatrix(int rows, int cols) { this(null, rows, cols); }
SparseRowMatrix(Matrix v, int rows, int cols) {
_rows = new TreeMap[rows];
for (int row=0;row();
_cols = cols;
if (v!=null)
for (int row=0;row row(int row) { return _rows[row]; }
public float[] raw() { throw new UnsupportedOperationException("raw access to the data in a sparse matrix is not implemented."); }
@Override public Frame toFrame(Key key) { return Neurons.toFrame(this, key); }
}
/**
* Sparse column matrix implementation
*/
static final class SparseColMatrix extends Iced implements Matrix {
private TreeMap[] _cols;
private int _rows;
SparseColMatrix(int rows, int cols) { this(null, rows, cols); }
SparseColMatrix(Matrix v, int rows, int cols) {
_rows = rows;
_cols = new TreeMap[cols];
for (int col=0;col();
if (v!=null)
for (int row=0;row col(int col) { return _cols[col]; }
public float[] raw() { throw new UnsupportedOperationException("raw access to the data in a sparse matrix is not implemented."); }
@Override public Frame toFrame(Key key) { return Neurons.toFrame(this, key); }
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy