public static Job createTimesSquaredJob(
Configuration initialConf,
Vector v,
int outputVectorDim,
Path matrixInputPath,
Path outputVectorPathBase,
Class<? extends TimesSquaredMapper> mapClass,
Class<? extends VectorSummingReducer> redClass)
throws IOException {
FileSystem fs = FileSystem.get(matrixInputPath.toUri(), initialConf);
matrixInputPath = fs.makeQualified(matrixInputPath);
outputVectorPathBase = fs.makeQualified(outputVectorPathBase);
long now = System.nanoTime();
Path inputVectorPath = new Path(outputVectorPathBase, INPUT_VECTOR + '/' + now);
SequenceFile.Writer inputVectorPathWriter = null;
try {
inputVectorPathWriter =
new SequenceFile.Writer(
fs, initialConf, inputVectorPath, NullWritable.class, VectorWritable.class);
inputVectorPathWriter.append(NullWritable.get(), new VectorWritable(v));
} finally {
Closeables.close(inputVectorPathWriter, false);
}
URI ivpURI = inputVectorPath.toUri();
DistributedCache.setCacheFiles(new URI[] {ivpURI}, initialConf);
Job job =
HadoopUtil.prepareJob(
matrixInputPath,
new Path(outputVectorPathBase, OUTPUT_VECTOR_FILENAME),
SequenceFileInputFormat.class,
mapClass,
NullWritable.class,
VectorWritable.class,
redClass,
NullWritable.class,
VectorWritable.class,
SequenceFileOutputFormat.class,
initialConf);
job.setCombinerClass(redClass);
job.setJobName("TimesSquaredJob: " + matrixInputPath);
Configuration conf = job.getConfiguration();
conf.set(INPUT_VECTOR, ivpURI.toString());
conf.setBoolean(IS_SPARSE_OUTPUT, !v.isDense());
conf.setInt(OUTPUT_VECTOR_DIMENSION, outputVectorDim);
return job;
}
/**
* A version to compute yRow as a sparse vector in case of extremely sparse matrices
*
* @param aRow
* @param yRowOut
*/
public void computeYRow(Vector aRow, Vector yRowOut) {
yRowOut.assign(0.0);
if (aRow.isDense()) {
int n = aRow.size();
for (int j = 0; j < n; j++) {
accumDots(j, aRow.getQuick(j), yRowOut);
}
} else {
for (Iterator<Element> iter = aRow.iterateNonZero(); iter.hasNext(); ) {
Element el = iter.next();
accumDots(el.index(), el.get(), yRowOut);
}
}
}
/**
* compute YRow=ARow*Omega.
*
* @param aRow row of matrix A (size n)
* @param yRow row of matrix Y (result) must be pre-allocated to size of (k+p)
*/
@Deprecated
public void computeYRow(Vector aRow, double[] yRow) {
// assert yRow.length == kp;
Arrays.fill(yRow, 0.0);
if (aRow.isDense()) {
int n = aRow.size();
for (int j = 0; j < n; j++) {
accumDots(j, aRow.getQuick(j), yRow);
}
} else {
for (Iterator<Element> iter = aRow.iterateNonZero(); iter.hasNext(); ) {
Element el = iter.next();
accumDots(el.index(), el.get(), yRow);
}
}
}
@Override
protected void map(Writable key, VectorWritable value, Context context)
throws IOException, InterruptedException {
omega.computeYRow(value.get(), yRow);
// compute outer product update for YtY
if (yRow.isDense()) {
for (int i = 0; i < kp; i++) {
double yi;
if ((yi = yRow.getQuick(i)) == 0.0) {
continue; // avoid densing up here unnecessarily
}
for (int j = i; j < kp; j++) {
double yj;
if ((yj = yRow.getQuick(j)) != 0.0) {
mYtY.setQuick(i, j, mYtY.getQuick(i, j) + yi * yj);
}
}
}
} else {
/*
* the disadvantage of using sparse vector (aside from the fact that we
* are creating some short-lived references) here is that we obviously
* do two times more iterations then necessary if y row is pretty dense.
*/
for (Iterator<Vector.Element> iterI = yRow.iterateNonZero(); iterI.hasNext(); ) {
Vector.Element eli = iterI.next();
int i = eli.index();
for (Iterator<Vector.Element> iterJ = yRow.iterateNonZero(); iterJ.hasNext(); ) {
Vector.Element elj = iterJ.next();
int j = elj.index();
if (j < i) {
continue;
}
mYtY.setQuick(i, j, mYtY.getQuick(i, j) + eli.get() * elj.get());
}
}
}
}
