public void Train(ILabeledExampleCollection <LblT, SparseVector <double> > dataset)
{
Utils.ThrowException(dataset == null ? new ArgumentNullException("dataset") : null);
Utils.ThrowException(dataset.Count == 0 ? new ArgumentValueException("dataset") : null);
Dispose();
int[] trainSet = new int[dataset.Count];
int[] labels = new int[dataset.Count];
int j = 0;
foreach (LabeledExample <LblT, SparseVector <double> > lblEx in dataset)
{
SparseVector <double> vec = lblEx.Example;
int[] idx = new int[vec.Count];
float[] val = new float[vec.Count];
for (int i = 0; i < vec.Count; i++)
{
idx[i] = vec.InnerIdx[i] + 1; // *** indices are 1-based in SvmLightLib
val[i] = (float)vec.InnerDat[i]; // *** loss of precision (double -> float)
}
int lbl;
if (!mLblToId.TryGetValue(lblEx.Label, out lbl))
{
mLblToId.Add(lblEx.Label, lbl = mLblToId.Count + 1); // *** labels start with 1 in SvmLightLib
mIdxToLbl.Add(lblEx.Label);
}
trainSet[j++] = SvmLightLib.NewFeatureVector(idx.Length, idx, val, lbl);
}
mModelId = SvmLightLib.TrainMulticlassModel(string.Format("-c {0} -e {1}", mC.ToString(CultureInfo.InvariantCulture), mEps.ToString(CultureInfo.InvariantCulture)),
trainSet.Length, trainSet);
// delete training vectors
foreach (int vecIdx in trainSet)
{
SvmLightLib.DeleteFeatureVector(vecIdx);
}
}
public void Load(BinarySerializer reader)
{
Utils.ThrowException(reader == null ? new ArgumentNullException("reader") : null);
Dispose();
// the following statements throw serialization-related exceptions
mVerbosityLevel = (SvmLightVerbosityLevel)reader.ReadInt();
mC = reader.ReadDouble();
mBiasedHyperplane = reader.ReadBool();
mKernelType = (SvmLightKernelType)reader.ReadInt();
mKernelParamGamma = reader.ReadDouble();
mKernelParamD = reader.ReadDouble();
mKernelParamS = reader.ReadDouble();
mKernelParamC = reader.ReadDouble();
mBiasedCostFunction = reader.ReadBool();
mCustomParams = reader.ReadString();
mEps = reader.ReadDouble();
mMaxIter = reader.ReadInt();
mIdxToLbl.Load(reader);
mLblCmp = reader.ReadObject <IEqualityComparer <LblT> >();
if (reader.ReadBool())
{
SvmLightLib.ReadByteCallback rb = delegate() { return(reader.ReadByte()); };
try
{
mModelId = SvmLightLib.LoadModelBinCallback(rb);
}
catch (BadImageFormatException e)
{
string assemblyPath = Path.GetDirectoryName(Assembly.GetExecutingAssembly().Location);
throw new BadImageFormatException(e.Message + "\n assembly path: " + assemblyPath, e);
}
GC.KeepAlive(rb);
}
}
// *** ISerializable interface implementation ***
public void Save(BinarySerializer writer)
{
Utils.ThrowException(writer == null ? new ArgumentNullException("writer") : null);
// the following statements throw serialization-related exceptions
writer.WriteInt((int)mVerbosityLevel);
writer.WriteDouble(mC);
writer.WriteBool(mBiasedHyperplane);
writer.WriteInt((int)mKernelType);
writer.WriteDouble(mKernelParamGamma);
writer.WriteDouble(mKernelParamD);
writer.WriteDouble(mKernelParamS);
writer.WriteDouble(mKernelParamC);
writer.WriteBool(mBiasedCostFunction);
writer.WriteString(mCustomParams);
writer.WriteDouble(mEps);
writer.WriteInt(mMaxIter);
mIdxToLbl.Save(writer);
writer.WriteObject(mLblCmp);
writer.WriteBool(mModelId != -1);
if (mModelId != -1)
{
SvmLightLib.WriteByteCallback wb = delegate(byte b) { writer.WriteByte(b); };
SvmLightLib.SaveModelBinCallback(mModelId, wb);
GC.KeepAlive(wb);
}
}
public void Load(BinarySerializer reader)
{
Utils.ThrowException(reader == null ? new ArgumentNullException("reader") : null);
Dispose();
// the following statements throw serialization-related exceptions
mVerbosityLevel = (SvmLightVerbosityLevel)reader.ReadInt();
mC = reader.ReadDouble();
mBiasedHyperplane = reader.ReadBool();
mKernelType = (SvmLightKernelType)reader.ReadInt();
mKernelParamGamma = reader.ReadDouble();
mKernelParamD = reader.ReadDouble();
mKernelParamS = reader.ReadDouble();
mKernelParamC = reader.ReadDouble();
mBiasedCostFunction = reader.ReadBool();
mCustomParams = reader.ReadString();
mEps = reader.ReadDouble();
mMaxIter = reader.ReadInt();
mIdxToLbl.Load(reader);
mLblCmp = reader.ReadObject <IEqualityComparer <LblT> >();
if (reader.ReadBool())
{
SvmLightLib.ReadByteCallback rb = delegate() { return(reader.ReadByte()); };
mModelId = SvmLightLib.LoadModelBinCallback(rb);
GC.KeepAlive(rb);
}
}
public ArrayList <KeyDat <double, int> > RankFeatures() // Guyon et al. 2002
{
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
ArrayList <KeyDat <double, int> > result = new ArrayList <KeyDat <double, int> >();
if (mKernelType != SvmLightKernelType.Linear)
{
// any kernel
int numFeat = SvmLightLib.GetFeatureCount(mModelId);
double allFeat = 0.5 * ComputeCost(-1);
for (int i = 0; i < numFeat; i++)
{
//Console.WriteLine("{0} / {1}", i + 1, numFeat);
double featScore = Math.Abs(allFeat - 0.5 * ComputeCost(/*rmvFeatIdx=*/ i));
result.Add(new KeyDat <double, int>(featScore, i));
}
}
else
{
// linear kernel (fast)
double[] w = GetLinearWeights();
for (int i = 0; i < w.Length; i++)
{
result.Add(new KeyDat <double, int>(0.5 * w[i] * w[i], i));
}
}
result.Sort(DescSort <KeyDat <double, int> > .Instance);
return(result);
}
public Prediction <LblT> Predict(SparseVector <double> example)
{
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
Utils.ThrowException(example == null ? new ArgumentNullException("example") : null);
Prediction <LblT> result = new Prediction <LblT>();
int[] idx = new int[example.Count];
float[] val = new float[example.Count];
for (int i = 0; i < example.Count; i++)
{
idx[i] = example.InnerIdx[i] + 1; // *** indices are 1-based in SvmLightLib
val[i] = (float)example.InnerDat[i]; // *** loss of precision (double -> float)
}
int vecId = SvmLightLib.NewFeatureVector(idx.Length, idx, val, 0);
SvmLightLib.MulticlassClassify(mModelId, 1, new int[] { vecId });
int n = SvmLightLib.GetFeatureVectorClassifScoreCount(vecId);
for (int i = 0; i < n; i++)
{
double score = SvmLightLib.GetFeatureVectorClassifScore(vecId, i);
LblT lbl = mIdxToLbl[i];
result.Inner.Add(new KeyDat <double, LblT>(score, lbl));
}
result.Inner.Sort(DescSort <KeyDat <double, LblT> > .Instance);
result.Trim();
SvmLightLib.DeleteFeatureVector(vecId); // delete feature vector
return(result);
}
public Prediction <LblT> Predict(SparseVector <double> example)
{
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
Utils.ThrowException(example == null ? new ArgumentNullException("example") : null);
Prediction <LblT> result = new Prediction <LblT>();
int[] idx = new int[example.Count];
float[] val = new float[example.Count];
for (int i = 0; i < example.Count; i++)
{
idx[i] = example.InnerIdx[i] + 1;
val[i] = (float)example.InnerDat[i]; // *** cast to float
}
int vecId = SvmLightLib.NewFeatureVector(idx.Length, idx, val, 0);
SvmLightLib.Classify(mModelId, 1, new int[] { vecId });
double score = SvmLightLib.GetFeatureVectorClassifScore(vecId, 0);
LblT lbl = mIdxToLbl[score > 0 ? 0 : 1];
LblT otherLbl = mIdxToLbl[score > 0 ? 1 : 0];
result.Inner.Add(new KeyDat <double, LblT>(Math.Abs(score), lbl));
result.Inner.Add(new KeyDat <double, LblT>(-Math.Abs(score), otherLbl));
SvmLightLib.DeleteFeatureVector(vecId); // delete feature vector
return(result);
}
public void LoadModel(string fileName)
{
Utils.ThrowException(typeof(LblT) != typeof(int) ? new InvalidOperationException() : null);
Utils.ThrowException(!Utils.VerifyFileNameOpen(fileName) ? new ArgumentValueException("fileName") : null);
Dispose();
mIdxToLbl.Add((LblT)(object)1);
mIdxToLbl.Add((LblT)(object)-1);
mModelId = SvmLightLib.LoadModel(fileName);
}
// *** IDisposable interface implementation ***
public void Dispose()
{
if (mModelId != -1)
{
SvmLightLib.DeleteMulticlassModel(mModelId);
mLblToId.Clear();
mIdxToLbl.Clear();
mModelId = -1;
}
}
// *** IDisposable interface implementation ***
public void Dispose()
{
if (mModelId != -1)
{
SvmLightLib.DeleteModel(mModelId);
mIdxToLbl.Clear();
mModelId = -1;
mWeights = null;
}
}
private SparseVector <double> GetSupportVector(int idx)
{
int featCount = SvmLightLib.GetSupportVectorFeatureCount(mModelId, idx);
SparseVector <double> vec = new SparseVector <double>();
for (int i = 0; i < featCount; i++)
{
vec.InnerIdx.Add(SvmLightLib.GetSupportVectorFeature(mModelId, idx, i));
vec.InnerDat.Add(SvmLightLib.GetSupportVectorWeight(mModelId, idx, i));
}
return(vec);
}
public ArrayList <IdxDat <double> > GetAlphas() // returns pairs (support vector index, alpha * y)
{
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
ArrayList <IdxDat <double> > alphas = new ArrayList <IdxDat <double> >();
for (int i = 0; i < SvmLightLib.GetSupportVectorCount(mModelId); i++)
{
double alpha = SvmLightLib.GetSupportVectorAlpha(mModelId, i);
int idx = SvmLightLib.GetSupportVectorIndex(mModelId, i);
alphas.Add(new IdxDat <double>(idx, alpha));
}
return(alphas);
}
// *** ISerializable interface implementation ***
public void Save(BinarySerializer writer)
{
Utils.ThrowException(writer == null ? new ArgumentNullException("writer") : null);
// the following statements throw serialization-related exceptions
writer.WriteDouble(mC);
writer.WriteDouble(mEps);
mIdxToLbl.Save(writer);
writer.WriteObject(mLblCmp);
writer.WriteBool(mModelId != -1);
if (mModelId != -1)
{
SvmLightLib.WriteByteCallback wb = delegate(byte b) { writer.WriteByte(b); };
SvmLightLib.SaveMulticlassModelBinCallback(mModelId, wb);
GC.KeepAlive(wb);
}
}
public void Train(ILabeledExampleCollection <LblT, SparseVector <double> > dataset)
{
Utils.ThrowException(dataset == null ? new ArgumentNullException("dataset") : null);
Utils.ThrowException(dataset.Count == 0 ? new ArgumentValueException("dataset") : null);
Dispose();
int[] trainSet = new int[dataset.Count];
int[] labels = new int[dataset.Count];
Dictionary <LblT, int> lblToIdx = new Dictionary <LblT, int>(mLblCmp);
MultiSet <int> lblCount = new MultiSet <int>();
int j = 0;
foreach (LabeledExample <LblT, SparseVector <double> > lblEx in dataset)
{
SparseVector <double> vec = lblEx.Example;
int[] idx = new int[vec.Count];
float[] val = new float[vec.Count];
for (int i = 0; i < vec.Count; i++)
{
idx[i] = vec.InnerIdx[i] + 1;
val[i] = (float)vec.InnerDat[i]; // *** cast to float
}
int lbl;
if (!lblToIdx.TryGetValue(lblEx.Label, out lbl))
{
lblToIdx.Add(lblEx.Label, lbl = lblToIdx.Count);
mIdxToLbl.Add(lblEx.Label);
}
Utils.ThrowException(lbl == 2 ? new ArgumentValueException("dataset") : null);
trainSet[j++] = SvmLightLib.NewFeatureVector(idx.Length, idx, val, lbl == 0 ? 1 : -1);
lblCount.Add(lbl == 0 ? 1 : -1);
}
string costFactor = "";
if (mBiasedCostFunction)
{
costFactor = "-j " + ((double)lblCount.GetCount(-1) / (double)lblCount.GetCount(1));
}
mModelId = SvmLightLib.TrainModel(string.Format(CultureInfo.InvariantCulture, "-v {0} -c {1} -t {2} -g {3} -d {4} -s {5} -r {6} -b {7} -e {8} -# {9} {10} {11}",
(int)mVerbosityLevel, mC, (int)mKernelType, mKernelParamGamma, mKernelParamD, mKernelParamS, mKernelParamC, mBiasedHyperplane ? 1 : 0,
mEps, mMaxIter, mCustomParams, costFactor), trainSet.Length, trainSet);
// delete training vectors
foreach (int vecIdx in trainSet)
{
SvmLightLib.DeleteFeatureVector(vecIdx);
}
}
public double[] GetLinearWeights()
{
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
Utils.ThrowException(mKernelType != SvmLightKernelType.Linear ? new InvalidOperationException() : null);
if (mWeights != null)
{
return(mWeights);
}
int featureCount = SvmLightLib.GetFeatureCount(mModelId);
double[] weights = new double[featureCount];
for (int i = 0; i < featureCount; i++)
{
weights[i] = SvmLightLib.GetLinearWeight(mModelId, i);
}
mWeights = weights;
return(weights);
}
public void Load(BinarySerializer reader)
{
Utils.ThrowException(reader == null ? new ArgumentNullException("reader") : null);
Dispose();
// the following statements throw serialization-related exceptions
mC = reader.ReadDouble();
mEps = reader.ReadDouble();
mIdxToLbl.Load(reader);
for (int i = 0; i < mIdxToLbl.Count; i++)
{
mLblToId.Add(mIdxToLbl[i], i + 1);
}
mLblCmp = reader.ReadObject <IEqualityComparer <LblT> >();
if (reader.ReadBool())
{
SvmLightLib.ReadByteCallback rb = delegate() { return(reader.ReadByte()); };
mModelId = SvmLightLib.LoadMulticlassModelBinCallback(rb);
GC.KeepAlive(rb);
}
}
public void SaveModel(string fileName)
{
Utils.ThrowException(!Utils.VerifyFileNameCreate(fileName) ? new ArgumentValueException("fileName") : null);
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
SvmLightLib.SaveModel(mModelId, fileName);
}
public double GetHyperplaneBias()
{
Utils.ThrowException(mModelId == -1 ? new InvalidOperationException() : null);
return(SvmLightLib.GetHyperplaneBias(mModelId));
}
private double[][] GetKernel(int rmvFeatIdx)
{
int numSv = SvmLightLib.GetSupportVectorCount(mModelId);
// initialize matrix
double[][] kernel = new double[numSv][];
// compute linear kernel
SparseMatrix <double> m = new SparseMatrix <double>();
for (int i = 0; i < numSv; i++)
{
SparseVector <double> sv = GetSupportVector(i);
m[i] = sv;
}
if (rmvFeatIdx >= 0)
{
m.RemoveColAt(rmvFeatIdx);
}
SparseMatrix <double> mTr = m.GetTransposedCopy();
for (int i = 0; i < numSv; i++)
{
double[] innerProd = ModelUtils.GetDotProductSimilarity(mTr, numSv, m[i]);
kernel[i] = innerProd;
}
// compute non-linear kernel
switch (mKernelType)
{
case SvmLightKernelType.Polynomial:
for (int row = 0; row < kernel.Length; row++)
{
for (int col = 0; col < kernel.Length; col++)
{
kernel[row][col] = Math.Pow(mKernelParamS * kernel[row][col] + mKernelParamC, mKernelParamD);
}
}
break;
case SvmLightKernelType.RadialBasisFunction:
double[] diag = new double[kernel.Length];
for (int i = 0; i < kernel.Length; i++)
{
diag[i] = kernel[i][i];
} // save diagonal
for (int row = 0; row < kernel.Length; row++)
{
for (int col = 0; col < kernel.Length; col++)
{
kernel[row][col] = Math.Exp(-mKernelParamGamma * (diag[row] + diag[col] - 2.0 * kernel[row][col]));
}
}
break;
case SvmLightKernelType.Sigmoid:
for (int row = 0; row < kernel.Length; row++)
{
for (int col = 0; col < kernel.Length; col++)
{
kernel[row][col] = Math.Tanh(mKernelParamS * kernel[row][col] + mKernelParamC);
}
}
break;
}
return(kernel);
}