public static void train_one(Problem prob, Parameters param, out double[] w, double Cp, double Cn)
{
double[][] inputs = prob.Inputs;
int[] labels = prob.Outputs.Apply(x => x >= 0 ? 1 : -1);
double eps = param.Tolerance;
int pos = 0;
for (int i = 0; i < labels.Length; i++)
if (labels[i] >= 0) pos++;
int neg = prob.Outputs.Length - pos;
double primal_solver_tol = eps * Math.Max(Math.Min(pos, neg), 1.0) / prob.Inputs.Length;
SupportVectorMachine svm = new SupportVectorMachine(prob.Dimensions);
ISupportVectorMachineLearning teacher = null;
switch (param.Solver)
{
case LibSvmSolverType.L2RegularizedLogisticRegression:
// l2r_lr_fun
teacher = new ProbabilisticNewtonMethod(svm, inputs, labels)
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol
}; break;
case LibSvmSolverType.L2RegularizedL2LossSvc:
// fun_obj=new l2r_l2_svc_fun(prob, C);
teacher = new LinearNewtonMethod(svm, inputs, labels)
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol
}; break;
case LibSvmSolverType.L2RegularizedL2LossSvcDual:
// solve_l2r_l1l2_svc(prob, w, eps, Cp, Cn, L2R_L2LOSS_SVC_DUAL);
teacher = new LinearCoordinateDescent(svm, inputs, labels)
{
Loss = Loss.L2,
PositiveWeight = Cp,
NegativeWeight = Cn,
}; break;
case LibSvmSolverType.L2RegularizedL1LossSvcDual:
// solve_l2r_l1l2_svc(prob, w, eps, Cp, Cn, L2R_L1LOSS_SVC_DUAL);
teacher = new LinearCoordinateDescent(svm, inputs, labels)
{
Loss = Loss.L1,
PositiveWeight = Cp,
NegativeWeight = Cn,
}; break;
case LibSvmSolverType.L1RegularizedLogisticRegression:
// solve_l1r_lr(&prob_col, w, primal_solver_tol, Cp, Cn);
teacher = new ProbabilisticCoordinateDescent(svm, inputs, labels)
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol
}; break;
case LibSvmSolverType.L2RegularizedLogisticRegressionDual:
// solve_l2r_lr_dual(prob, w, eps, Cp, Cn);
teacher = new ProbabilisticDualCoordinateDescent(svm, inputs, labels)
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol,
}; break;
}
Trace.WriteLine("Training " + param.Solver);
// run the learning algorithm
var sw = Stopwatch.StartNew();
double error = teacher.Run();
sw.Stop();
// save the solution
w = svm.ToWeights();
Trace.WriteLine(String.Format("Finished {0}: {1} in {2}",
param.Solver, error, sw.Elapsed));
}
public static LibSvmModel train(Problem prob, Parameters parameters)
{
double[] w;
double Cp = parameters.Complexity;
double Cn = parameters.Complexity;
if (parameters.ClassWeights != null)
{
for (int i = 0; i < parameters.ClassLabels.Count; i++)
{
if (parameters.ClassLabels[i] == -1)
Cn *= parameters.ClassWeights[i];
else if (parameters.ClassLabels[i] == +1)
Cn *= parameters.ClassWeights[i];
}
}
train_one(prob, parameters, out w, Cp, Cn);
return new LibSvmModel()
{
Dimension = prob.Dimensions,
Classes = 2,
Labels = new[] { +1, -1 },
Solver = parameters.Solver,
Weights = w,
Bias = 0
};
}
public static string check_parameter(Problem prob, Parameters param)
{
if (param.Epsilon <= 0)
return "eps <= 0";
if (param.Complexity <= 0)
return "C <= 0";
if (param.Epsilon < 0)
return "p < 0";
if (!Enum.IsDefined(typeof(LibSvmSolverType), param.Solver))
return "unknown solver type";
if (param.CrossValidation)
return "cross-validation is not supported at this time.";
return null;
}
public static Parameters parse_command_line(string[] args,
out string input_file_name, out string model_file_name)
{
// default values
var parameters = new Parameters()
{
Solver = LibSvmSolverType.L2RegularizedL1LossSvcDual,
Complexity = 1,
Tolerance = Double.PositiveInfinity,
Epsilon = 0.1,
Bias = -1
};
input_file_name = null;
model_file_name = null;
int i;
try
{
// parse options
for (i = 0; i < args.Length; i++)
{
if (args[i][0] != '-')
break;
if (++i >= args.Length)
exit_with_help();
switch (args[i - 1][1])
{
case 's':
parameters.Solver = (LibSvmSolverType)Int32.Parse(args[i]);
break;
case 'c':
parameters.Complexity = Double.Parse(args[i]);
break;
case 'p':
parameters.Epsilon = Double.Parse(args[i]);
break;
case 'e':
parameters.Tolerance = Double.Parse(args[i]);
break;
case 'B':
parameters.Bias = Double.Parse(args[i]);
break;
case 'w':
parameters.ClassLabels.Add(Int32.Parse(args[i - 1][2].ToString()));
parameters.ClassWeights.Add(Double.Parse(args[i]));
break;
case 'v':
parameters.CrossValidation = true;
parameters.ValidationFolds = Int32.Parse(args[i]);
if (parameters.ValidationFolds < 2)
{
Console.WriteLine("n-fold cross validation: n must >= 2");
exit_with_help();
}
break;
case 'q':
Trace.Listeners.Clear();
i--;
break;
default:
Console.WriteLine("unknown option: -%c" + args[i - 1][1]);
exit_with_help();
break;
}
}
// determine filenames
if (i >= args.Length)
exit_with_help();
input_file_name = args[i];
if (i < args.Length - 1)
model_file_name = args[i + 1];
}
catch
{
exit_with_help();
}
if (parameters.Tolerance == Double.PositiveInfinity)
{
switch (parameters.Solver)
{
case LibSvmSolverType.L2RegularizedLogisticRegression:
case LibSvmSolverType.L2RegularizedL2LossSvc:
parameters.Tolerance = 0.01;
break;
case LibSvmSolverType.L2RegularizedL2LossSvr:
parameters.Tolerance = 0.001;
break;
case LibSvmSolverType.L2RegularizedL2LossSvcDual:
case LibSvmSolverType.L2RegularizedL1LossSvcDual:
case LibSvmSolverType.MulticlassSvmCrammerSinger:
case LibSvmSolverType.L2RegularizedLogisticRegressionDual:
parameters.Tolerance = 0.1;
break;
case LibSvmSolverType.L1RegularizedL2LossSvc:
case LibSvmSolverType.L1RegularizedLogisticRegression:
parameters.Tolerance = 0.01;
break;
case LibSvmSolverType.L2RegularizedL1LossSvrDual:
case LibSvmSolverType.L2RegularizedL2LossSvrDual:
parameters.Tolerance = 0.1;
break;
}
}
return parameters;
}
private static ISupervisedLearning<SupportVectorMachine, double[], int> create(
Parameters param, double Cp, double Cn, double[][] inputs, int[] outputs)
{
double eps = param.Tolerance;
int n = outputs.Length;
int pos = 0;
for (int i = 0; i < outputs.Length; i++)
{
if (outputs[i] >= 0)
pos++;
}
int neg = n - pos;
double primal_solver_tol = eps * Math.Max(Math.Min(pos, neg), 1.0) / n;
switch (param.Solver)
{
case LibSvmSolverType.L2RegularizedLogisticRegression:
// l2r_lr_fun
return new ProbabilisticNewtonMethod()
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol
};
case LibSvmSolverType.L2RegularizedL2LossSvc:
// fun_obj=new l2r_l2_svc_fun(prob, C);
return new LinearNewtonMethod()
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol
};
case LibSvmSolverType.L2RegularizedL2LossSvcDual:
// solve_l2r_l1l2_svc(prob, w, eps, Cp, Cn, L2R_L2LOSS_SVC_DUAL);
return new LinearDualCoordinateDescent()
{
Loss = Loss.L2,
PositiveWeight = Cp,
NegativeWeight = Cn,
};
case LibSvmSolverType.L2RegularizedL1LossSvcDual:
// solve_l2r_l1l2_svc(prob, w, eps, Cp, Cn, L2R_L1LOSS_SVC_DUAL);
return new LinearDualCoordinateDescent()
{
Loss = Loss.L1,
PositiveWeight = Cp,
NegativeWeight = Cn,
};
case LibSvmSolverType.L1RegularizedLogisticRegression:
// solve_l1r_lr(&prob_col, w, primal_solver_tol, Cp, Cn);
return new ProbabilisticCoordinateDescent()
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol
};
case LibSvmSolverType.L2RegularizedLogisticRegressionDual:
// solve_l2r_lr_dual(prob, w, eps, Cp, Cn);
return new ProbabilisticDualCoordinateDescent()
{
PositiveWeight = Cp,
NegativeWeight = Cn,
Tolerance = primal_solver_tol,
};
}
throw new InvalidOperationException("Unknown solver type: {0}".Format(param.Solver));
}
public static void train_one(Problem prob, Parameters param, out double[] w, double Cp, double Cn)
{
double[][] inputs = prob.Inputs;
int[] labels = prob.Outputs.Apply(x => x >= 0 ? 1 : -1);
// Create the learning algorithm from the parameters
var teacher = create(param, Cp, Cn, inputs, labels);
Trace.WriteLine("Training " + param.Solver);
// Run the learning algorithm
var sw = Stopwatch.StartNew();
SupportVectorMachine svm = teacher.Learn(inputs, labels);
sw.Stop();
double error = new HingeLoss(labels).Loss(svm.Score(inputs));
// Save the solution
w = svm.ToWeights();
Trace.WriteLine(String.Format("Finished {0}: {1} in {2}",
param.Solver, error, sw.Elapsed));
}
