void nextStep()
{
// compute jacobian
if (_retry == 0)
{
_jac = _jac_func(_inputs, _weights);
}
// compute hessian approximation with tykhonov damping coefficient
var jacT = new NRealMatrix(_jac.Rows, _jac.Columns);
jacT.SetArray(_jac.ToArray());
jacT.Transpose();
var dampedHessian = new NRealMatrix(_jac.Columns, _jac.Columns);
dampedHessian = jacT * _jac;
for (int idxRow = 0; idxRow < dampedHessian.Rows; idxRow++)
{
dampedHessian.SetAt(idxRow, idxRow, new NDouble(dampedHessian[idxRow, idxRow] * (1.0 + _lambda) + 1e-10));
}
var adj = new NRealMatrix(dampedHessian.Rows, 1);
var y = new NRealMatrix(dampedHessian.Rows, 1);
y = jacT * _error;
// solve dampedHessian * adj = y
LapackLib.Instance.SolveSle(dampedHessian, y, adj);
var nextWeights = new NRealMatrix(1, _weights.Columns);
for (int idxWeight = 0; idxWeight < nextWeights.Columns; idxWeight++)
{
nextWeights.SetAt(0, idxWeight, new NDouble(_weights[0, idxWeight] - adj[idxWeight, 0]));
}
// compute errors
var error = calcError(nextWeights);
var totalError = calcTotalError(error);
if (totalError > _totalError)
{
// revert step and increase damping factor
if (_retry < 100)
{
_lambda *= 11.0;
_retry++;
}
else
{
updateWeights();
throw new StallException();
}
}
else
{
// accept step and decrease damping factor
_lambda /= 9.0;
_weights.SetArray(nextWeights.ToArray());
_error = error;
_totalError = totalError;
_retry = 0;
}
}
public LevenbergMarquardt(objective_func obj_func, List <double> inputs, List <Value> modelParams, model_func model, model_func model_jac, double lambda = 0.001, double obj_error = 0.00001, int max_iter = 10000, int rnd_seed = 0)
{
if (inputs.Count == 0)
{
throw new ApplicationException("Number of input data must be > 0");
}
if (modelParams.Count == 0)
{
throw new ApplicationException("Number of model parameters must be > 0");
}
_obj_func = obj_func;
_model = model;
_jac_func = model_jac;
_lambda = lambda;
_max_iter = max_iter;
_inputs = new NRealMatrix(inputs.Count, 1);
_inputs.SetArray((from input in inputs select new NDouble[] { new NDouble(input) }).ToArray());
_outputs = _obj_func(_inputs);
_modelParams = modelParams;
// initalize the weights with normal random distibution
var seed = new MLapack.MCJIMatrix(4, 1);
var rndSeed = rnd_seed == 0 ? 321 : rnd_seed;
seed.setAt(0, 0, rndSeed);
seed.setAt(1, 0, rndSeed);
seed.setAt(2, 0, rndSeed);
seed.setAt(3, 0, rndSeed);
// check if a guess has been provided
bool modelParamInitialized = false;
foreach (var weight in modelParams)
{
if (weight.X != 0)
{
modelParamInitialized = true;
}
}
if (modelParamInitialized)
{
_weights = new NRealMatrix(1, modelParams.Count);
_weights.SetArray(new NDouble[][] { (from param in modelParams select new NDouble(param.X)).ToArray() });
}
else
{
_weights = LapackLib.Instance.RandomMatrix(RandomDistributionType.Uniform_0_1, seed, 1, modelParams.Count);
for (int idxWeight = 0; idxWeight < _weights.Columns; idxWeight++)
{
_weights[0, idxWeight] = (_weights[0, idxWeight] * 2.0 - 1.0) / Math.Sqrt(inputs.Count);
}
}
_obj_error = obj_error;
_error = calcError(_weights);
_totalError = calcTotalError(_error);
_startError = _totalError;
}
public LevenbergMarquardt(objective_func obj_func, List<double> inputs, List<Value> modelParams, model_func model, model_func model_jac, double lambda = 0.001, double obj_error = 0.00001, int max_iter = 10000, int rnd_seed = 0)
{
if (inputs.Count == 0)
throw new ApplicationException("Number of input data must be > 0");
if (modelParams.Count == 0)
throw new ApplicationException("Number of model parameters must be > 0");
_obj_func = obj_func;
_model = model;
_jac_func = model_jac;
_lambda = lambda;
_max_iter = max_iter;
_inputs = new NRealMatrix(inputs.Count, 1);
_inputs.SetArray((from input in inputs select new NDouble[] { new NDouble(input) } ).ToArray());
_outputs = _obj_func(_inputs);
_modelParams = modelParams;
// initalize the weights with normal random distibution
var seed = new MLapack.MCJIMatrix(4,1);
var rndSeed = rnd_seed == 0 ? 321 : rnd_seed;
seed.setAt(0, 0, rndSeed);
seed.setAt(1, 0, rndSeed);
seed.setAt(2, 0, rndSeed);
seed.setAt(3, 0, rndSeed);
// check if a guess has been provided
bool modelParamInitialized = false;
foreach (var weight in modelParams)
{
if (weight.X != 0)
modelParamInitialized = true;
}
if (modelParamInitialized)
{
_weights = new NRealMatrix(1, modelParams.Count);
_weights.SetArray(new NDouble[][] {(from param in modelParams select new NDouble(param.X)).ToArray() });
}
else
{
_weights = LapackLib.Instance.RandomMatrix(RandomDistributionType.Uniform_0_1, seed, 1, modelParams.Count);
for (int idxWeight = 0; idxWeight < _weights.Columns; idxWeight++)
_weights[0, idxWeight] = (_weights[0, idxWeight] * 2.0 - 1.0) / Math.Sqrt(inputs.Count);
}
_obj_error = obj_error;
_error = calcError(_weights);
_totalError = calcTotalError(_error);
_startError = _totalError;
}
void nextStep()
{
// compute jacobian
if (_retry == 0)
_jac = _jac_func(_inputs, _weights);
// compute hessian approximation with tykhonov damping coefficient
var jacT = new NRealMatrix(_jac.Rows, _jac.Columns);
jacT.SetArray(_jac.ToArray());
jacT.Transpose();
var dampedHessian = new NRealMatrix(_jac.Columns, _jac.Columns);
dampedHessian = jacT * _jac;
for (int idxRow = 0; idxRow < dampedHessian.Rows; idxRow++)
dampedHessian.SetAt(idxRow, idxRow, new NDouble(dampedHessian[idxRow, idxRow] * (1.0 + _lambda) + 1e-10));
var adj = new NRealMatrix(dampedHessian.Rows, 1);
var y = new NRealMatrix(dampedHessian.Rows, 1);
y = jacT * _error;
// solve dampedHessian * adj = y
LapackLib.Instance.SolveSle(dampedHessian, y, adj);
var nextWeights = new NRealMatrix(1, _weights.Columns);
for (int idxWeight = 0; idxWeight < nextWeights.Columns; idxWeight++)
nextWeights.SetAt(0, idxWeight, new NDouble(_weights[0, idxWeight] - adj[idxWeight, 0]));
// compute errors
var error = calcError(nextWeights);
var totalError = calcTotalError(error);
if (totalError > _totalError)
{
// revert step and increase damping factor
if (_retry < 100)
{
_lambda *= 11.0;
_retry++;
}
else
{
updateWeights();
throw new StallException();
}
}
else
{
// accept step and decrease damping factor
_lambda /= 9.0;
_weights.SetArray(nextWeights.ToArray());
_error = error;
_totalError = totalError;
_retry = 0;
}
}
