/** Function to compute the spectral distance between two song models.
* This is a fast implementation of the symmetrized Kullback Leibler
* Divergence.
*/
public static float Distance(Scms s1, Scms s2, ScmsConfiguration c)
{
float val = 0;
int i;
int k;
int idx = 0;
int dim = c.Dimension;
int covlen = c.CovarianceLength;
float tmp1;
unsafe {
fixed (float* s1cov = s1.cov, s2icov = s2.icov,
s1icov = s1.icov, s2cov = s2.cov,
s1mean = s1.mean, s2mean = s2.mean,
mdiff = c.MeanDiff, aicov = c.AddInverseCovariance)
{
for (i = 0; i < covlen; i++) {
aicov[i] = s1icov[i] + s2icov[i];
}
for (i = 0; i < dim; i++) {
idx = i*dim - (i*i+i)/2;
val += s1cov[idx+i] * s2icov[idx+i] +
s2cov[idx+i] * s1icov[idx+i];
for (k = i+1; k < dim; k++) {
val += 2*s1cov[idx+k] * s2icov[idx+k] +
2*s2cov[idx+k] * s1icov[idx+k];
}
}
for (i = 0; i < dim; i++) {
mdiff[i] = s1mean[i] - s2mean[i];
}
for (i = 0; i < dim; i++) {
idx = i - dim;
tmp1 = 0;
for (k = 0; k <= i; k++) {
idx += dim - k;
tmp1 += aicov[idx] * mdiff[k];
}
for (k = i + 1; k < dim; k++) {
idx++;
tmp1 += aicov[idx] * mdiff[k];
}
val += tmp1 * mdiff[i];
}
}
}
// FIXME: fix the negative return values
//val = Math.Max(0.0f, (val/2 - s1.cov.dim));
val = val / 4 - c.Dimension / 2;
return val;
}
private static void Test()
{
mirageaudio_initgst();
string song1_filename = "/home/lorentz/Music/Library/Pachelbel/Johann Pachelbel - Canon And Gigue In D Major For 3 Violins And Basso Continuo.mp3";
string song2_filename = "/home/lorentz/Music/Library/Karajan Adagios/CD 1/Pachelbel - Canon in d Major (Kanon And Gigue in d Major = d Dur) av Johann Pachelbel.mp3";
Scms song1 = null;
Scms song2 = null;
DbgTimer t1 = new DbgTimer();
t1.Start();
int runs = 10;
for (int i = 0; i < runs; i++) {
Analyzer.Analyze(song1_filename);
}
long l1 = 0;
t1.Stop(ref l1);
Dbg.WriteLine("Analysis: " + runs + " times - " + l1 + "ms; " +
(double)l1/(double)runs + "ms per analysis");
song1 = Analyzer.Analyze(song1_filename);
song2 = Analyzer.Analyze(song2_filename);
ScmsConfiguration config = new ScmsConfiguration (Analyzer.MFCC_COEFFICIENTS);
Console.WriteLine("Distance = " + Scms.Distance (song1, song2, config));
DbgTimer t2 = new DbgTimer();
t2.Start();
runs = 100000;
for (int i = 0; i < runs; i++) {
Scms.Distance (song1, song2, config);
}
long l2 = 0;
t2.Stop(ref l2);
Dbg.WriteLine("Distance Computation: " + runs + " times - " + l2 + "ms; " +
(double)l2/(double)runs + "ms per comparison");
}
/** Function to compute the spectral distance between two song models.
* This is a fast implementation of the symmetrized Kullback Leibler
* Divergence.
*/
public static float Distance(Scms s1, Scms s2, ScmsConfiguration c)
{
float val = 0;
int i;
int k;
int idx = 0;
int dim = c.Dimension;
int covlen = c.CovarianceLength;
float tmp1;
unsafe
{
fixed(float *s1cov = s1.cov, s2icov = s2.icov,
s1icov = s1.icov, s2cov = s2.cov,
s1mean = s1.mean, s2mean = s2.mean,
mdiff = c.MeanDiff, aicov = c.AddInverseCovariance)
{
for (i = 0; i < covlen; i++)
{
aicov[i] = s1icov[i] + s2icov[i];
}
for (i = 0; i < dim; i++)
{
idx = i * dim - (i * i + i) / 2;
val += s1cov[idx + i] * s2icov[idx + i] +
s2cov[idx + i] * s1icov[idx + i];
for (k = i + 1; k < dim; k++)
{
val += 2 * s1cov[idx + k] * s2icov[idx + k] +
2 * s2cov[idx + k] * s1icov[idx + k];
}
}
for (i = 0; i < dim; i++)
{
mdiff[i] = s1mean[i] - s2mean[i];
}
for (i = 0; i < dim; i++)
{
idx = i - dim;
tmp1 = 0;
for (k = 0; k <= i; k++)
{
idx += dim - k;
tmp1 += aicov[idx] * mdiff[k];
}
for (k = i + 1; k < dim; k++)
{
idx++;
tmp1 += aicov[idx] * mdiff[k];
}
val += tmp1 * mdiff[i];
}
}
}
// FIXME: fix the negative return values
//val = Math.Max(0.0f, (val/2 - s1.cov.dim));
val = val / 4 - c.Dimension / 2;
return(val);
}
