/// <summary>
/// Performs training iteration.
/// </summary>
public bool Iteration()
{
//Primary stop condition
if (AttemptEpoch == MaxAttemptEpoch)
{
return(false);
}
//New lambda to be tested
double newLambda = _lambdaSeeker.Next;
//Secondary stop condition
if (AttemptEpoch > 0 && Math.Abs(_currLambda - newLambda) < StopLambdaDifference)
{
return(false);
}
//Next epoch allowed
_currLambda = newLambda;
++AttemptEpoch;
InfoMessage = $"lambda={_currLambda.ToString(CultureInfo.InvariantCulture)}";
//Inverse _XTdotX matrix
Matrix I;
if (_currLambda > 0)
{
Matrix B = new Matrix(_XTdotX);
double tmp = B.Data[0][0];
B.AddScalarToDiagonal(_currLambda);
B.Data[0][0] = tmp;
I = B.Inverse(true);
}
else
{
I = _XTdotX.Inverse(true);
}
//New weights buffer
double[] newWeights = new double[_net.NumOfWeights];
//Weights for each output neuron
for (int outputIdx = 0; outputIdx < _net.NumOfOutputValues; outputIdx++)
{
//Weights solution
Vector weights = I * _XTdotY[outputIdx];
//Store weights
//Bias
newWeights[_net.NumOfOutputValues * _net.NumOfInputValues + outputIdx] = weights.Data[0];
//Predictors
for (int i = 0; i < _net.NumOfInputValues; i++)
{
newWeights[outputIdx * _net.NumOfInputValues + i] = weights.Data[i + 1];
}
}
//Set new weights and compute error
_net.SetWeights(newWeights);
MSE = _net.ComputeBatchErrorStat(_inputVectorCollection, _outputVectorCollection).MeanSquare;
//Update lambda seeker
_lambdaSeeker.ProcessError(MSE);
return(true);
}
/// <inheritdoc/>
public bool Iteration()
{
//Primary stop condition
if (AttemptEpoch == MaxAttemptEpoch)
{
return(false);
}
//Next epoch allowed
++AttemptEpoch;
InfoMessage = $"Lambda={_cfg.Lambda.ToString(CultureInfo.InvariantCulture)} Alpha={_cfg.Alpha.ToString(CultureInfo.InvariantCulture)}";
//Whole network new weights buffer
double[] newWeights = _net.GetWeightsCopy();
//Optimization of the weights for each output separeatelly
for (int outputIdx = 0; outputIdx < _net.NumOfOutputValues; outputIdx++)
{
//New weights for specific output neuron
double[] weights = new double[_net.NumOfInputValues + 1];
//Copy weights and compute the sum
//Bias first
weights[0] = newWeights[_net.NumOfOutputValues * _net.NumOfInputValues + outputIdx];
//Inputs next
Parallel.For(0, _net.NumOfInputValues, i =>
{
weights[i + 1] = newWeights[outputIdx * _net.NumOfInputValues + i];
});
//Elastic iteration
//Compute and store current output values and new bias
double oldBias = weights[0];
double newBias = 0;
double[] parallelSubResults1 = new double[_parallelRanges.Count];
double[] parallelSubResults2 = new double[_parallelRanges.Count];
double[] computedOutputs = new double[_outputVectorCollection.Count];
parallelSubResults1.Populate(0);
Parallel.For(0, _parallelRanges.Count, rangeIdx =>
{
for (int i = _parallelRanges[rangeIdx].Item1; i < _parallelRanges[rangeIdx].Item2; i++)
{
computedOutputs[i] = ComputeLinOutput(i, weights);
parallelSubResults1[rangeIdx] += (_outputVectorCollection[i][outputIdx] - computedOutputs[i] + oldBias);
}
});
//New bias finalization
for (int i = 0; i < _parallelRanges.Count; i++)
{
newBias += parallelSubResults1[i];
}
newBias /= _outputVectorCollection.Count;
weights[0] = newBias;
//Update computed outputs if bias has changed
double biasDifference = newBias - oldBias;
if (biasDifference != 0)
{
Parallel.For(0, _parallelRanges.Count, rangeIdx =>
{
for (int i = _parallelRanges[rangeIdx].Item1; i < _parallelRanges[rangeIdx].Item2; i++)
{
computedOutputs[i] += biasDifference;
}
});
}
//Optimization
for (int inputValueIdx = 0; inputValueIdx < _net.NumOfInputValues; inputValueIdx++)
{
//Fit and denominator computation
double oldWeight = weights[1 + inputValueIdx];
double fit = 0, denominator = 0;
parallelSubResults1.Populate(0);
parallelSubResults2.Populate(0);
Parallel.For(0, _parallelRanges.Count, rangeIdx =>
{
for (int i = _parallelRanges[rangeIdx].Item1; i < _parallelRanges[rangeIdx].Item2; i++)
{
double x = _inputVectorCollection[i][inputValueIdx];
if (x != 0)
{
parallelSubResults1[rangeIdx] += x * (_outputVectorCollection[i][outputIdx] - computedOutputs[i] + x * oldWeight);
parallelSubResults2[rangeIdx] += x * x;
}
}
});
//Fit and denominator finalization
for (int i = 0; i < _parallelRanges.Count; i++)
{
fit += parallelSubResults1[i];
denominator += parallelSubResults2[i];
}
fit /= _outputVectorCollection.Count;
denominator /= _outputVectorCollection.Count;
denominator += _cfg.Lambda * (1d - _cfg.Alpha);
double newWeight = 0;
if (denominator != 0)
{
newWeight = SoftThreshold(fit) / denominator;
}
//Set new weight
weights[1 + inputValueIdx] = newWeight;
//Update computed values
double weightsDiff = newWeight - oldWeight;
if (weightsDiff != 0)
{
Parallel.For(0, _parallelRanges.Count, rangeIdx =>
{
for (int i = _parallelRanges[rangeIdx].Item1; i < _parallelRanges[rangeIdx].Item2; i++)
{
double x = _inputVectorCollection[i][inputValueIdx];
if (x != 0)
{
computedOutputs[i] += weightsDiff * x;
}
}
});
}
}
//Put optimized weights back to newWaights buffer
//Bias
newWeights[_net.NumOfOutputValues * _net.NumOfInputValues + outputIdx] = weights[0];
//Inputs
for (int i = 0; i < _net.NumOfInputValues; i++)
{
newWeights[outputIdx * _net.NumOfInputValues + i] = weights[i + 1];
}
}
//Set new weights and compute final error
_net.SetWeights(newWeights);
MSE = _net.ComputeBatchErrorStat(_inputVectorCollection, _outputVectorCollection).MeanSquare;
return(true);
}
/// <summary>
/// Performs training iteration.
/// </summary>
public bool Iteration()
{
//Fetch next noise intensity
double newNoise = _noiseSeeker.Next;
//Check continue conditions
if (AttemptEpoch == MaxAttemptEpoch || Math.Abs(_currNoise - newNoise) < StopNoiseDifference)
{
//Try new attempt
if (!NextAttempt())
{
//Next attempt is not available
return(false);
}
else
{
newNoise = _noiseSeeker.Next;
}
}
//Next epoch
++AttemptEpoch;
_currNoise = newNoise;
InfoMessage = $"noise={_currNoise.ToString(CultureInfo.InvariantCulture)}";
//Adjusted predictors
Matrix predictors = PreparePredictors(_currNoise);
//Decomposition
QRD decomposition = null;
bool useableQRD = true;
try
{
//Try to create QRD. Any exception signals numerical unstability
decomposition = new QRD(predictors);
}
catch
{
//Creation of QRD object throws exception. QRD object is not ready for use.
useableQRD = false;
if (AttemptEpoch == 1)
{
//No previous successful epoch so stop training
throw;
}
}
if (useableQRD)
{
//QRD is ready for use (low probability of numerical unstability)
//New weights
double[] newWeights = new double[_net.NumOfWeights];
//Regression for each output neuron
for (int outputIdx = 0; outputIdx < _net.NumOfOutputValues; outputIdx++)
{
//Regression
Matrix solution = decomposition.Solve(_outputSingleColMatrixCollection[outputIdx]);
//Store weights
//Input weights
for (int i = 0; i < solution.NumOfRows - 1; i++)
{
newWeights[outputIdx * _net.NumOfInputValues + i] = solution.Data[i][0];
}
//Bias weight
newWeights[_net.NumOfOutputValues * _net.NumOfInputValues + outputIdx] = solution.Data[solution.NumOfRows - 1][0];
}
//Set new weights and compute error
_net.SetWeights(newWeights);
MSE = _net.ComputeBatchErrorStat(_inputVectorCollection, _outputVectorCollection).MeanSquare;
}
_noiseSeeker.ProcessError(MSE);
return(true);
}
