private void getnegphase() {
/*
* It does the negative phase of unsupervised RBM training algorithm
*
* For details, please refer to Dr. Hinton's paper:
* Reducing the dimensionality of data with neural networks. Science, Vol. 313. no. 5786, pp. 504 - 507, 28 July 2006.
*/
// start calculate the negative phase
// calculate the curved value of v1,h1
// find the vector of v1
Matrix negdata = poshidstates.times(vishid.transpose());
// (1 * numhid) * (numhid * numdims) = (1 * numdims)
negdata.plusEquals(visbiases);
// poshidstates*vishid' + visbiases
double[][] tmp1 = negdata.getArray();
int i1 = 0;
while (i1 < numdims) {
tmp1[0][i1] = 1 / (1 + Math.exp(-tmp1[0][i1]));
i1++;
}
// find the vector of h1
neghidprobs = negdata.times(vishid);
// (1 * numdims) * (numdims * numhid) = (1 * numhid)
neghidprobs.plusEquals(hidbiases);
double[][] tmp2 = neghidprobs.getArray();
int i2 = 0;
while (i2 < numhid) {
tmp2[0][i2] = 1 / (1 + Math.exp(-tmp2[0][i2]));
i2++;
}
negprods = negdata.transpose().times(neghidprobs);
// (numdims * 1) *(1 * numhid) = (numdims * numhid)
}
public void map(
LongWritable key,
Text value,
OutputCollector<IntWritable, DoubleWritable> output,
Reporter reporter)
throws IOException {
/*
* It implements the mapper. It outputs the numbers of weight and updated weights.
*
* Note that the format of intermediate output is <IntWritable, DoubleWritable>,
* because the key is the number of weight (an integer), and the value is the weight's value (double)
*/
inputData = value.toString();
// go through the process
initialize();
getposphase();
getnegphase();
update();
// output the intermediate data
// The <key, value> pairs are <weightID, weightUpdate>
double[][] vishidinc_array = vishidinc.getArray();
for (int i = 0; i < numdims; i++) {
for (int j = 0; j < numhid; j++) {
weightPos.set(i * numhid + j);
weightValue.set(vishidinc_array[i][j]);
output.collect(weightPos, weightValue);
}
}
}
private void prop2nextLayer() {
/*
* It computes the forward propagation algorithm.
*/
poshidprobs = data.times(vishid);
// (1 * numdims) * (numdims * numhid)
poshidprobs.plusEquals(hidbiases);
// data*vishid + hidbiases
double[][] product_tmp2 = poshidprobs.getArray();
for (int i2 = 0; i2 < numhid; i2++) {
/*
* compute the updated input, and write them to newinput
*/
product_tmp2[0][i2] = 1 / (1 + Math.exp(-product_tmp2[0][i2]));
newinput[i2] = (int) (product_tmp2[0][i2] * 255.0);
}
}
private void getposphase() {
/*
* It does the positive phase of unsupervised RBM training algorithm
*
* For details, please refer to Dr. Hinton's paper:
* Reducing the dimensionality of data with neural networks. Science, Vol. 313. no. 5786, pp. 504 - 507, 28 July 2006.
*/
// Start calculate the positive phase
// calculate the cured value of h0
poshidprobs = data.times(vishid);
// (1 * numdims) * (numdims * numhid)
poshidprobs.plusEquals(hidbiases);
// data*vishid + hidbiases
double[][] product_tmp2 = poshidprobs.getArray();
int i2 = 0;
while (i2 < numhid) {
product_tmp2[0][i2] = 1 / (1 + Math.exp(-product_tmp2[0][i2]));
i2++;
}
posprods = data.transpose().times(poshidprobs);
// (numdims * 1) * (1 * numhid)
// end of the positive phase calculation, find the binary presentation of h0
int i3 = 0;
double[][] tmp1 = poshidprobs.getArray();
double[][] tmp2 = new double[1][numhid];
Random randomgenerator = new Random();
while (i3 < numhid) {
/*
* a sampling according to possiblity given by poshidprobs
*/
if (tmp1[0][i3] > randomgenerator.nextDouble()) tmp2[0][i3] = 1;
else tmp2[0][i3] = 0;
i3++;
}
// poshidstates is a binary sampling according to possiblity given by poshidprobs
poshidstates = new Matrix(tmp2);
}
// update the weights and biases
// This serves as a reducer
private void update() {
/*
* It computes the update of weights using previous results and parameters
*
* For details, please refer to Dr. Hinton's paper:
* Reducing the dimensionality of data with neural networks. Science, Vol. 313. no. 5786, pp. 504 - 507, 28 July 2006.
*/
double momentum;
// if (epoch > 5)
// momentum = finalmomentum;
// else
// momentum = initialmomentum;
// vishidinc = momentum*vishidinc + epsilonw*( (posprods-negprods)/numcases -
// weightcost*vishid);
// vishidinc.timesEquals(momentum);
Matrix temp1 = posprods.minus(negprods);
Matrix temp2 = vishid.times(weightcost);
temp1.minusEquals(temp2);
temp1.timesEquals(epsilonw);
// the final updates of weights are written in vishidinc
vishidinc.plusEquals(temp1);
}
