/// <summary>
///
/// Scala listę kroków stosując średnią ważoną gdzie wagą jest długość kroku
/// </summary>
/// <param name="st1"></param>
/// <param name="newAlphas"></param>
/// <param name="a1"></param>
/// <param name="modAlphas">new merged alpha</param>
/// <param name="weighSum"></param>
/// <returns>average value for a1</returns>
private float WeightedDistMerge(AlphaInfo st1, List <StepPairVariable> newAlphas, AlphaInfo[] modAlphas)
{
int size = modAlphas.Length;
float weighSum = 0;
float a1 = 0f;
float[] weights = new float[size];
modAlphas[0] = new AlphaInfo(st1.Index, 0, st1.Y, st1.Error, st1.Product);
for (int k = 1; k < size; k++)
{
var secAlpha = newAlphas[k - 1].Second;
modAlphas[k] = new AlphaInfo(secAlpha.Index, 0, secAlpha.Y, secAlpha.Error, secAlpha.Product);
var it = newAlphas[k - 1];
//we compute change in distance for two alphas
float sq1 = it.First.Alpha - alpha[it.First.Index];
sq1 *= sq1;
float sq2 = it.Second.Alpha - alpha[it.Second.Index];
sq2 *= sq2;
weights[k] = sq1 + sq2;
weighSum += weights[k];
}
weighSum = Math.Abs(weighSum);
a1 = MergeWithWeight(newAlphas, modAlphas, weighSum, weights);
return(a1);
}
private float AvgMerge(AlphaInfo st1, List <StepPairVariable> newAlphas, AlphaInfo[] modAlphas)
{
int size = modAlphas.Length;
float weighSum = 0;
float a1 = 0f;
modAlphas[0] = new AlphaInfo(st1.Index, 0, st1.Y, st1.Error, st1.Product);
float[] weights = new float[size];
float weight = 1.0f / newAlphas.Count;
for (int k = 1; k < size; k++)
{
var secAlpha = newAlphas[k - 1].Second;
modAlphas[k] = new AlphaInfo(secAlpha.Index, 0, secAlpha.Y, secAlpha.Error, secAlpha.Product);
weights[k] = weight;
}
weighSum = 1;
a1 = MergeWithWeight(newAlphas, modAlphas, weighSum, weights);
return(a1);
}
/// <summary>
///
/// </summary>
/// <param name="i1">index for first alpha</param>
/// <param name="E1">decision error for first alpah</param>
/// <param name="y1">label coresponding to first alpha</param>
/// <param name="alph1">alpha</param>
/// <param name="i2">index second alpha</param>
/// <returns></returns>
private StepPairVariable ComputeAlphaStep(AlphaInfo step1, int i2)
{
float y1 = 0, y2 = 0, s = 0;
float alph1, alph2 = 0; /* old_values of alpha_1, alpha_2 */
float a1 = 0, a2 = 0; /* new values of alpha_1, alpha_2 */
float E1 = 0, E2 = 0, //error for second alpha
L = 0, H = 0, //lower upper bounds
k11 = 0, k22 = 0, k12 = 0, //kernel products
eta = 0, // step taken,
Lobj = 0, Hobj = 0;
int i1 = step1.Index;
if (i1 == i2)
{
return(null); // no step taken
}
alph1 = step1.Alpha;
y1 = step1.Y;
E1 = step1.Error;
alph2 = alpha[i2];
y2 = problem.Y[i2];
E2 = ErrorForAlpha(i2);
s = y1 * y2;
//compute bounds
if (y1 == y2)
{
float gamma = alph1 + alph2;
if (gamma > C)
{
L = gamma - C;
H = C;
}
else
{
L = 0;
H = gamma;
}
}
else
{
float gamma = alph2 - alph1;
if (gamma > 0)
{
L = gamma;
H = C;
}
else
{
L = 0;
H = C - gamma;
}
}
if (L == H)
{
return(null); // no step take
}
k11 = step1.Product;
k12 = Product(i1, i2);
k22 = Product(i2, i2);
eta = 2 * k12 - k11 - k22;
if (eta < 0)
{
//original version with plus
//a2 = alph2 + y2 * (E2 - E1) / eta;
a2 = alph2 - y2 * (E1 - E2) / eta;
if (a2 < L)
{
a2 = L;
}
else if (a2 > H)
{
a2 = H;
}
}
else
{
{
float c1 = eta / 2;
float c2 = y2 * (E1 - E2) - eta * alph2;
Lobj = c1 * L * L + c2 * L;
Hobj = c1 * H * H + c2 * H;
}
if (Lobj > Hobj + epsilon)
{
a2 = L;
}
else if (Lobj < Hobj - epsilon)
{
a2 = H;
}
else
{
a2 = alph2;
}
}
if (Math.Abs(a2 - alph2) < epsilon * (a2 + alph2 + epsilon))
{
return(null); // no step taken
}
a1 = alph1 - s * (a2 - alph2);
if (a1 < 0)
{
a2 += s * a1;
a1 = 0;
}
else if (a1 > C)
{
float t = a1 - C;
a2 += s * t;
a1 = C;
}
var st1 = new AlphaInfo(i1, a1, y1, E1, k11);
var st2 = new AlphaInfo(i2, a2, y2, E2, k22);
return(new StepPairVariable(st1, st2, k12, s, eta));
}
private AlphaInfo[] MergeSteps(AlphaInfo st1, List <StepPairVariable> newAlphas)
{
//a1 - new alpha, alpha1-old alpha
float a1 = 0f, alpha1 = 0;
int index1 = st1.Index;
float y1 = st1.Y;
alpha1 = alpha[st1.Index];
float k11 = st1.Product,
E1 = st1.Error;
int size = newAlphas.Count + 1;
AlphaInfo[] modAlphas = new AlphaInfo[size];
a1 = WeightedEtaMerge(st1, newAlphas, modAlphas);
//a1 = WeightedSmallEtaMerge(st1, newAlphas, modAlphas);
//a1 = AvgMerge(st1, newAlphas, modAlphas);
//a1 = WeightedDistMerge(st1, newAlphas, modAlphas);
if (a1 < 0)
{
for (int i = 1; i < size; i++)
{
modAlphas[i].Alpha += newAlphas[i - 1].Si * a1;
}
//a2 += s * a1;
a1 = 0;
}
else if (a1 > C)
{
float t = a1 - C;
for (int i = 1; i < size; i++)
{
modAlphas[i].Alpha += newAlphas[i - 1].Si * t;
}
//a2 += s * t;
a1 = C;
}
// modAlphas.AddFirst(new Pair<int, float>(index1, a1));
modAlphas[0].Alpha = a1;
//float b1 = 0, b2 = 0;
float bnew = 0;
if (a1 > 0 && a1 < C)
{
// bnew = b + E1 + y1 * (a1 - alph1) * k11 + y2 * (a2 - alph2) * k12;
bnew = b + E1 + y1 * (a1 - alpha1) * k11;
// int j = 0;
for (int i = 1; i < size; i++)
{
var item = modAlphas[i];
float al2 = item.Alpha;
int ind = item.Index;
float oldAl = alpha[ind];
bnew += item.Y * (al2 - oldAl) * newAlphas[i - 1].Product;
}
}
else
{
for (int i = 1; i < size; i++)
{
var item = modAlphas[i];
float prodI1 = newAlphas[i - 1].Product;
if (item.Alpha > 0 && item.Alpha < C)
{
int mIndex = item.Index;
bnew = item.Error + b + y1 * (a1 - alpha1) * prodI1;
for (int j = 1; j < modAlphas.Length; j++)
{
var st = modAlphas[j];
//if (mIndex == st.Index)
// bnew += st.Error;
bnew += st.Y * (st.Alpha - alpha[st.Index]) * Product(mIndex, st.Index);
}
//we compute new B;
break;
}
}
//if (a2 > 0 && a2 < C)
//{
// bnew = b + E2 + y1 * (a1 - alph1) * k12 + y2 * (a2 - alph2) * k22;
//}
//else
//{
// b1 = b + E1 + y1 * (a1 - alph1) * k11 + y2 * (a2 - alph2) * k12;
// b2 = b + E2 + y1 * (a1 - alph1) * k12 + y2 * (a2 - alph2) * k22;
// bnew = (b1 + b2) / 2;
//}
}
//todo: plus or minus?
delta_b = bnew - b;
b = bnew;
return(modAlphas);
}
private List <StepPairVariable> FindAlphaSteps(int cores, AlphaInfo st1)
{
List <StepPairVariable> newAlphas = new List <StepPairVariable>(cores);
object lockObj = new object();
int k = st1.Index;
//foreach (var i2 in GlobalHeuristic(k))
//{
// StepPairVariable ap = ComputeAlphaStep(st1, i2);
// if (ap == null) continue;
// newAlphas.Add(ap);
// if (newAlphas.Count >= cores)
// break;
//}
#region Parallel version
Parallel.ForEach(GlobalHeuristic(k), (i2, loopState) =>
{
if (loopState.ShouldExitCurrentIteration)
{
return;
}
StepPairVariable ap = ComputeAlphaStep(st1, i2);
if (ap == null)
{
return;
}
lock (lockObj)
{
if (newAlphas.Count >= cores)
{
//stop searching
loopState.Stop();
return;
}
else
{
newAlphas.Add(ap);
if (newAlphas.Count >= cores)
{
loopState.Stop();
}
}
}
});
#endregion
return(newAlphas);
}
/*
* public override Model<TProblemElement> ComputeModel()
* {
* errorCache = new float[problem.ElementsCount];
* alpha = new float[problem.ElementsCount];
* // SMO algorithm
* int numChange = 0;
* int examineAll = 1;
* int kktViolatiors = 0;
* int cores = System.Environment.ProcessorCount;
*
* while (numChange > 0 || examineAll > 0)
* {
* numChange = 0;
*
* for (int k = 0; k < problem.ElementsCount; k++)
* {
*
* int i2 = -1;
* //find first index
* if (ExamineExample(k))
* {
* List<AlphaPair> newAlphasPair = FindSubIndexes(cores, k);
*
* if (newAlphasPair.Count == 0)
* continue;
* else
* numChange += newAlphasPair.Count;
*
*
*
* var newAlphas= MergeSteps(newAlphasPair);
*
*
*
* int modSize = newAlphas.Count;
* //list of modified alphas
* //list for changes in alpha_i - = y1 * (a1 - alph1);
* //key - alpha index, value- alpha step
* List<KeyValuePair<int, float>> alphaStep = new List<KeyValuePair<int, float>>();
*
*
* int index = k;
* float yi, oldAlpha, newAlpha;
* //error for first index
* //errors for second indexes, we start from One to #cores+1
* //in order to fill up the alphaStep array, we have to remeber
* //that modAlphas start form 0 to #cores so its is nessesary to substrac One
* foreach (var item in newAlphas)
* {
* index = item.First;
* newAlpha = item.Second;
* yi = problem.Labels[index];
* oldAlpha = alpha[index];
*
* alphaStep.Add(new KeyValuePair<int, float>(index, yi * (newAlpha - oldAlpha)));
* }
*
*
* //float t1 = y1 * (a1 - alph1);
* //float t2 = y2 * (a2 - alph2);
*
* for (int i = 0; i < problem.ElementsCount; i++)
* {
* if (0 < alpha[i] && alpha[i] < C)
* {
* for (int j = 0; j < alphaStep.Count; j++)
* {
* int alphaIndex = alphaStep[j].Key;
* errorCache[i] += alphaStep[j].Value * Product(i, alphaIndex);
* }
* errorCache[i] -= delta_b;
* //old version -errorCache[i] +=t1 * Product(i1, i)+ t2 * Product(i2, i) - delta_b;
* }
* }
*
* //update erroCache for alpha's
* for (int j = 0; j < alphaStep.Count; j++)
* {
* int alphaIndex = alphaStep[j].Key;
* //zero error on modified alphas
* errorCache[alphaIndex] = 0f;
* }
* //errorCache[i1] = 0f;
* //errorCache[i2] = 0f;
*
* //update alphas
* foreach (var item in newAlphas)
* {
* //take arithmetic average of alphas
* alpha[item.First] = item.Second;
* }
*
* //update alpha, first index
* //alpha[k] = avgAlpha1;
* }
* }
* if (examineAll == 1)
* {
* examineAll = 0;
* }
* else if (numChange == 0)
* {
* examineAll = 1;
* }
#region Old code
* //else
* //{
* // for (int k = 0; k < problem.ElementsCount; k++)
* // {
* // if (alpha[k] != 0 && alpha[k] != C)
* // {
* // if (ExamineExample(k)) numChange++;
* // }
* // }
* //}
*
* //if (examineAll == 1)
* //{
* // examineAll = 0;
* //}
* //else if (numChange == 0)
* //{
* // examineAll = 1;
* //}
#endregion
* }
*
* // cleaning
* errorCache = null;
*
#region Building Model
* Model<TProblemElement> model = new Model<TProblemElement>();
* model.NumberOfClasses = 2;
* model.Alpha = alpha;
* model.Rho = b;
*
*
* List<TProblemElement> supportElements = new List<TProblemElement>();
* List<int> suporrtIndexes = new List<int>(alpha.Length);
* for (int i = 0; i < alpha.Length; i++)
* {
* if (alpha[i] > 0)
* {
* supportElements.Add(problem.Elements[i]);
* suporrtIndexes.Add(i);
* }
*
* }
* model.SupportElements = supportElements.ToArray();
* model.SupportElementsIndexes = suporrtIndexes.ToArray();
#endregion
*
* return model;
* }
*/
public override Model <TProblemElement> ComputeModel()
{
errorCache = new float[problem.ElementsCount];
alpha = new float[problem.ElementsCount];
// SMO algorithm
int numChange = 0;
int examineAll = 1;
//int kktViolatiors = 0;
int cores = System.Environment.ProcessorCount;
float E1 = float.MinValue;
float y1, alpha1;
long mainIter = 0;
long subIter = 0;
while (numChange > 0 || examineAll > 0)
{
numChange = 0;
mainIter++;
//if (examineAll > 0)
//{
// for (int k = 0; k < problem.ElementsCount; k++)
// {
// if (ExamineExample(k))
// {
// numChange++;
// subIter++;
// }
// }
//}
//else
//{
// for (int k = 0; k < problem.ElementsCount; k++)
// {
// if (alpha[k] != 0 && alpha[k] != C)
// {
// if (ExamineExample(k))
// {
// numChange++;
// subIter++;
// }
// }
// }
//}
if (examineAll > 0)
{
for (int k = 0; k < problem.ElementsCount; k++)
{
y1 = problem.Y[k];
alpha1 = alpha[k];
E1 = ErrorForAlpha(k);
//find first index
if (!KKTViolator(E1, y1, alpha1))
{
continue;
}
subIter++;
AlphaInfo st1 = new AlphaInfo(k, alpha1, y1, E1, Product(k, k));
var newSteps = FindAlphaSteps(cores, st1);
if (newSteps.Count == 0)
{
continue;
}
else
{
numChange += newSteps.Count;
}
var newAlphas = MergeSteps(st1, newSteps);
UpdateAlpha(newAlphas);
}
}
else
{
for (int k = 0; k < problem.ElementsCount; k++)
{
if (alpha[k] == 0 || alpha[k] == C)
{
continue;
}
y1 = problem.Y[k];
alpha1 = alpha[k];
E1 = ErrorForAlpha(k);
//find first index
if (!KKTViolator(E1, y1, alpha1))
{
continue;
}
subIter++;
AlphaInfo st1 = new AlphaInfo(k, alpha1, y1, E1, Product(k, k));
var newSteps = FindAlphaSteps(cores, st1);
if (newSteps.Count == 0)
{
continue;
}
else
{
numChange += newSteps.Count;
}
var newAlphas = MergeSteps(st1, newSteps);
UpdateAlpha(newAlphas);
}
}
if (examineAll == 1)
{
examineAll = 0;
}
else if (numChange == 0)
{
examineAll = 1;
}
}
// cleaning
errorCache = null;
#region Building Model
Model <TProblemElement> model = new Model <TProblemElement>();
model.NumberOfClasses = 2;
model.Alpha = alpha;
model.Bias = b;
//List<TProblemElement> supportElements = new List<TProblemElement>();
//List<int> suporrtIndexes = new List<int>(alpha.Length);
//for (int i = 0; i < alpha.Length; i++)
//{
// if (alpha[i] > 0)
// {
// supportElements.Add(problem.Elements[i]);
// suporrtIndexes.Add(i);
// }
//}
//model.SupportElements = supportElements.ToArray();
//model.SupportElementsIndexes = suporrtIndexes.ToArray();
List <TProblemElement> supportElements = new List <TProblemElement>(alpha.Length);
List <int> suporrtIndexes = new List <int>(alpha.Length);
List <float> supportLabels = new List <float>(alpha.Length);
for (int j = 0; j < alpha.Length; j++)
{
if (Math.Abs(alpha[j]) > 0)
{
supportElements.Add(problem.Elements[j]);
suporrtIndexes.Add(j);
supportLabels.Add(problem.Y[j]);
}
}
model.SupportElements = supportElements.ToArray();
model.SupportElementsIndexes = suporrtIndexes.ToArray();
model.Y = supportLabels.ToArray();
#endregion
Debug.WriteLine("Main iteration=" + mainIter + " subIteration=" + subIter);
Console.WriteLine("Main iteration=" + mainIter + " subIteration=" + subIter);
Console.WriteLine("SV=" + model.SupportElementsIndexes.Length);
return(model);
}
