/// <summary>
/// 调用BeginSampleEvent
/// </summary>
/// <param name="currentIteration">
/// 当前训练迭代
/// </param>
/// <param name="currentSample">
/// 当前样品已成功训练
/// </param>
protected virtual void OnEndSample(int currentIteration, TrainingSample currentSample)
{
if (EndSampleEvent != null)
{
EndSampleEvent(this, new TrainingSampleEventArgs(currentIteration, currentSample));
}
}
/// <summary>
/// 训练给定训练集的神经网络（批处理训练）
/// </summary>
/// <param name="trainingSet">
/// 使用的训练集
/// </param>
/// <param name="trainingEpochs">
/// 训练时期数。 （所有样本在每个训练时期以一些随机顺序训练）
/// </param>
/// <exception cref="ArgumentNullException">
/// 如果<c> trainingSet </ c>是<c> null </ c>
/// </exception>
/// <exception cref="ArgumentException">
/// 如果<c> trainingEpochs </ c>为零或负值
/// </exception>
public virtual void Learn(TrainingSet trainingSet, int trainingEpochs)
{
// 验证
Helper.ValidateNotNull(trainingSet, "trainingSet");
Helper.ValidatePositive(trainingEpochs, "trainingEpochs");
if ((trainingSet.InputVectorLength != inputLayer.NeuronCount) ||
(trainingMethod == TrainingMethod.Supervised && trainingSet.OutputVectorLength != outputLayer.NeuronCount) ||
(trainingMethod == TrainingMethod.Unsupervised && trainingSet.OutputVectorLength != 0))
{
throw new ArgumentException("Invalid training set");
}
// 重置isStopping
isStopping = false;
// 重新初始化网络
Initialize();
for (int currentIteration = 0; currentIteration < trainingEpochs; currentIteration++)
{
int[] randomOrder = Helper.GetRandomOrder(trainingSet.TrainingSampleCount);
// 开始新的训练时期
OnBeginEpoch(currentIteration, trainingSet);
// 检查抖动时期
if (jitterEpoch > 0 && currentIteration % jitterEpoch == 0)
{
for (int i = 0; i < connectors.Count; i++)
{
connectors[i].Jitter(jitterNoiseLimit);
}
}
for (int index = 0; index < trainingSet.TrainingSampleCount; index++)
{
TrainingSample randomSample = trainingSet[randomOrder[index]];
// 学习随机训练样本
OnBeginSample(currentIteration, randomSample);
LearnSample(trainingSet[randomOrder[index]], currentIteration, trainingEpochs);
OnEndSample(currentIteration, randomSample);
// 检查我们是否需要停止
if (isStopping)
{
isStopping = false; return;
}
}
// 训练时期成功完成
OnEndEpoch(currentIteration, trainingSet);
// 检查我们是否需要停止
if (isStopping)
{
isStopping = false; return;
}
}
}
/// <summary>
/// Trains the neural network for the given training set (Batch Training)
/// </summary>
/// <param name="trainingSet">
/// The training set to use
/// </param>
/// <param name="trainingEpochs">
/// Number of training epochs. (All samples are trained in some random order, in every
/// training epoch)
/// </param>
/// <exception cref="ArgumentNullException">
/// if <c>trainingSet</c> is <c>null</c>
/// </exception>
/// <exception cref="ArgumentException">
/// if <c>trainingEpochs</c> is zero or negative
/// </exception>
public virtual void Learn(TrainingSet trainingSet, int trainingEpochs)
{
// Validate
Helper.ValidateNotNull(trainingSet, "trainingSet");
Helper.ValidatePositive(trainingEpochs, "trainingEpochs");
if ((trainingSet.InputVectorLength != inputLayer.NeuronCount) ||
(trainingMethod == TrainingMethod.Supervised && trainingSet.OutputVectorLength != outputLayer.NeuronCount) ||
(trainingMethod == TrainingMethod.Unsupervised && trainingSet.OutputVectorLength != 0))
{
throw new ArgumentException("Invalid training set");
}
// Reset isStopping
isStopping = false;
// Re-Initialize the network
Initialize();
for (int currentIteration = 0; currentIteration < trainingEpochs; currentIteration++)
{
int[] randomOrder = Helper.GetRandomOrder(trainingSet.TrainingSampleCount);
// Beginning a new training epoch
OnBeginEpoch(currentIteration, trainingSet);
// Check for Jitter Epoch
if (jitterEpoch > 0 && currentIteration % jitterEpoch == 0)
{
for (int i = 0; i < connectors.Count; i++)
{
connectors[i].Jitter(jitterNoiseLimit);
}
}
for (int index = 0; index < trainingSet.TrainingSampleCount; index++)
{
TrainingSample randomSample = trainingSet[randomOrder[index]];
// Learn a random training sample
OnBeginSample(currentIteration, randomSample);
LearnSample(trainingSet[randomOrder[index]], currentIteration, trainingEpochs);
OnEndSample(currentIteration, randomSample);
// Check if we need to stop
if (isStopping)
{
isStopping = false; return;
}
}
// Training Epoch successfully complete
OnEndEpoch(currentIteration, trainingSet);
// Check if we need to stop
if (isStopping)
{
isStopping = false; return;
}
}
}
/// <summary>
/// 训练给定训练样本的网络（在线训练模式）。 注意这个方法仅训练样本一次，而不管当前的时期是什么。 参数只是用来找出训练进度，并根据它调整参数。
/// </summary>
/// <param name="trainingSample">
/// 培训样品使用
/// </param>
/// <param name="currentIteration">
/// 当前训练迭代
/// </param>
/// <param name="trainingEpochs">
/// 训练时期数
/// </param>
/// <exception cref="ArgumentNullException">
/// 如果<c> trainingSample </ c>为<c> null </ c>
/// </exception>
/// <exception cref="ArgumentException">
/// 如果<c> trainingEpochs </ c>不为正
/// </exception>
/// <exception cref="ArgumentOutOfRangeException">
/// 如果<c> currentIteration </ c>为负，或者如果它不小于<c> trainingEpochs </ c>
/// </exception>
public virtual void Learn(TrainingSample trainingSample, int currentIteration, int trainingEpochs)
{
Helper.ValidateNotNull(trainingSample, "trainingSample");
Helper.ValidatePositive(trainingEpochs, "trainingEpochs");
Helper.ValidateWithinRange(currentIteration, 0, trainingEpochs - 1, "currentIteration");
OnBeginSample(currentIteration, trainingSample);
LearnSample(trainingSample, currentIteration, trainingEpochs);
OnEndSample(currentIteration, trainingSample);
}
/// <summary>
/// 向训练集添加新的监督训练样本。 如果已经存在，它将被替换。
/// </summary>
/// <param name="sample">
/// 要添加的示例
/// </param>
/// <exception cref="ArgumentNullException">
/// 如果<c> sample </ c>为<c> null </ c>
/// </exception>
/// <exception cref="ArgumentException">
/// 如果输入向量或输出向量的大小与其预期大小不同
/// </exception>
public void Add(TrainingSample sample)
{
// 验证
Helper.ValidateNotNull(sample, "sample");
if (sample.InputVector.Length != inputVectorLength)
{
throw new ArgumentException
("Input vector must be of size " + inputVectorLength, "sample");
}
if (sample.OutputVector.Length != outputVectorLength)
{
throw new ArgumentException
("Output vector must be of size " + outputVectorLength, "sample");
}
// 请注意，正在添加引用。 （样本是不可变的）
trainingSamples.Add(sample);
}
/// <summary>
/// Adds a new supervised training sample to the training set. If already exists, it will
/// be replaced.
/// </summary>
/// <param name="sample">
/// The sample to add
/// </param>
/// <exception cref="ArgumentNullException">
/// If <c>sample</c> is <c>null</c>
/// </exception>
/// <exception cref="ArgumentException">
/// If sizes of input vector or output vector are different from their expected sizes
/// </exception>
public void Add(TrainingSample sample)
{
// Validation
Helper.ValidateNotNull(sample, "sample");
if (sample.InputVector.Length != inputVectorLength)
{
throw new ArgumentException
("Input vector must be of size " + inputVectorLength, "sample");
}
if (sample.OutputVector.Length != outputVectorLength)
{
throw new ArgumentException
("Output vector must be of size " + outputVectorLength, "sample");
}
// Note that the reference is being added. (Sample is immutable)
trainingSamples.Add(sample);
}
/// <summary>
/// Determine whether the given object is equal to this instance
/// </summary>
/// <param name="obj">
/// The object to compare with this instance
/// </param>
/// <returns>
/// <c>true</c> if the given object is equal to this instance, <c>false</c> otherwise
/// </returns>
public override bool Equals(object obj)
{
if (obj is TrainingSample)
{
TrainingSample sample = (TrainingSample)obj;
int size;
if ((size = sample.inputVector.Length) == inputVector.Length)
{
for (int i = 0; i < size; i++)
{
if (inputVector[i] != sample.inputVector[i])
{
return(false);
}
}
return(true);
}
}
return(false);
}
/// <summary> /// A protected helper function used to train single learning sample /// </summary> /// <param name="trainingSample"> /// Training sample to use /// </param> /// <param name="currentIteration"> /// Current training epoch (Assumed to be positive and less than <c>trainingEpochs</c>) /// </param> /// <param name="trainingEpochs"> /// Number of training epochs (Assumed to be positive) /// </param> protected abstract void LearnSample(TrainingSample trainingSample, int currentIteration, int trainingEpochs);
/// <summary>
/// Determines whether the training sample is present in the set
/// </summary>
/// <param name="sample">
/// The sample to locate
/// </param>
/// <returns>
/// <c>true</c> if present, <c>false</c> otherwise
/// </returns>
public bool Contains(TrainingSample sample)
{
return trainingSamples.Contains(sample);
}
/// <summary>
/// Adds a new supervised training sample to the training set. If already exists, it will
/// be replaced.
/// </summary>
/// <param name="sample">
/// The sample to add
/// </param>
/// <exception cref="ArgumentNullException">
/// If <c>sample</c> is <c>null</c>
/// </exception>
/// <exception cref="ArgumentException">
/// If sizes of input vector or output vector are different from their expected sizes
/// </exception>
public void Add(TrainingSample sample)
{
// Validation
Helper.ValidateNotNull(sample, "sample");
if (sample.InputVector.Length != inputVectorLength)
{
throw new ArgumentException
("Input vector must be of size " + inputVectorLength, "sample");
}
if (sample.OutputVector.Length != outputVectorLength)
{
throw new ArgumentException
("Output vector must be of size " + outputVectorLength, "sample");
}
// Note that the reference is being added. (Sample is immutable)
trainingSamples.Add(sample);
}
/// <summary>
/// Creates a new instance of this class
/// </summary>
/// <param name="trainingIteration">
/// Current training iteration
/// </param>
/// <param name="trainingSample">
/// The training sample associated with the event
/// </param>
public TrainingSampleEventArgs(int trainingIteration, TrainingSample trainingSample)
{
this.trainingIteration = trainingIteration;
this.trainingSample = trainingSample;
}
/// <summary>
/// 确定训练样本是否存在于集合中
/// </summary>
/// <param name="sample">
/// 要定位的示例
/// </param>
/// <returns>
/// <c> true </ c>如果存在，<c> false </ c>
/// </returns>
public bool Contains(TrainingSample sample)
{
return(trainingSamples.Contains(sample));
}
/// <summary>
/// 删除给定的训练样本
/// </summary>
/// <param name="sample">
/// 要除去的样品
/// </param>
/// <returns>
/// <c> true </ c>如果成功，<c> false </ c>
/// </returns>
public bool Remove(TrainingSample sample)
{
return(trainingSamples.Remove(sample));
}
/// <summary>
/// Roda as redes para cada um dos treinamentos, separando os treinamentos por dia
/// Calcula também a taxa média de acerto de cada rede neural por dia
/// </summary>
/// <param name="dadosTreinamento"></param>
/// <param name="redesPrevisaoFinanceira"></param>
/// <returns></returns>
private static Dictionary<int, TrainingSet> SelecionarTreinamentosPorDia(ConfiguracaoCaptacaoRedes configuracaoCaptacao)
{
int tamanhoEntrada = configuracaoCaptacao.RedesPrevisao.Max(rp => rp.JanelaEntrada);
List<Treinamento> treinamentos = DataBaseUtils.DataBaseUtils.SelecionarTreinamentos(configuracaoCaptacao.Dados, tamanhoEntrada, totalDiasPrevisao, 2);
//Dictionary<int, Dictionary<List<double>, List<double>>> dadosTreinamentoPorShift = new Dictionary<int, Dictionary<List<double>, List<double>>>();
////Resgata o maior tamanho das entradas das redes neurais
////Cria o dicionario de input output por shift
//for (int numShift = 0; numShift < quantidadeShifts; numShift++)
//{
// //Seleciona oa dados antes do shift
// dadosTreinamentoPorShift.Add(numShift, SelecionarInput_Output(configuracaoCaptacao.Dados.Take((configuracaoCaptacao.Dados.Count / (quantidadeShifts - 1)) * numShift).ToList(), tamanhoEntrada));
// //Seleciona os dados depois do shift
// dadosTreinamentoPorShift[numShift].ToList().AddRange(SelecionarInput_Output(configuracaoCaptacao.Dados.Skip((configuracaoCaptacao.Dados.Count / (quantidadeShifts - 1)) * (numShift + 1)).ToList(), tamanhoEntrada));
//}
//Inicializa todos os dias da previsao
for (int diaTreinamento = 0; diaTreinamento < totalDiasPrevisao; diaTreinamento++)
{
foreach (RedePrevisaoFinanceira redePrevisao in configuracaoCaptacao.RedesPrevisao)
{
redePrevisao.TaxaMediaAcertoPorDia.Add(diaTreinamento, 0);
}
}
Dictionary<int, TrainingSet> treinamentosPorDia = new Dictionary<int, TrainingSet>();
for (int dia = 0; dia < totalDiasPrevisao; dia++)
{
treinamentosPorDia.Add(dia, new TrainingSet(tamanhoEntrada, configuracaoCaptacao.RedesPrevisao.Count));
}
//Roda as redes para cada um dos dados de treinamento selecionados
foreach (Treinamento treinamento in treinamentos)
{
//Guarda um historico dos outputs das redes para tratar as redes que tem output menor do que o tamanho da previsao
List<List<double>> outputsRedes = new List<List<double>>();
List<double[]> taxasAcertoRedes = new List<double[]>();
foreach (RedePrevisaoFinanceira redePrevisao in configuracaoCaptacao.RedesPrevisao)
{
Network redeNeural = redePrevisao.RedeNeuralPrevisaoFinanceiraPorDivisaoCrossValidation[treinamento.DivisaoCrossValidation];
//Roda a rede neural, gerando a previsao \/ como a rede pode receber um input menor do que os treinamentos, pegamos apenas os ultimos registros do mesmo
double[] outputPrevisao = redeNeural.Run(treinamento.Input.Skip(treinamento.Input.Count - redePrevisao.JanelaEntrada).Take(redePrevisao.JanelaEntrada).ToArray());
//Alimenta o historico de previsoes de cada rede
outputsRedes.Add(outputPrevisao.ToList());
double[] taxaAcertoPorDia = new double[totalDiasPrevisao];
for (int diasPrevistosRN = 0; diasPrevistosRN < outputPrevisao.Length; diasPrevistosRN++)
{
//Calcula a taxa de acerto da previsao
double ta = Math.Min(outputPrevisao[diasPrevistosRN], treinamento.Output[diasPrevistosRN]) / Math.Max(outputPrevisao[diasPrevistosRN], treinamento.Output[diasPrevistosRN]);
taxaAcertoPorDia[diasPrevistosRN] = ta;
redePrevisao.TaxaMediaAcertoPorDia[diasPrevistosRN] += ta;
}
taxasAcertoRedes.Add(taxaAcertoPorDia);
}
//Lista com o melhor desultado de cada dia (cada item da lista corresponde a um dia..)
List<double> melhoresResultadosDia = new List<double>(treinamento.Input);
//Trata as taxas de acerto que não foram calculadas pois a rede tem output menor do que a quantidade de dias da previsao
for (int dia = 1; dia < totalDiasPrevisao; dia++)
{
//Recupera o indice da rede que tem a melhor taxa de acerto para o dia anterior
int indMelhorRedeParaODiaAnterior = taxasAcertoRedes.IndexOf(taxasAcertoRedes.OrderByDescending(taxasAcertoRede => taxasAcertoRede[dia - 1]).First());
//Adiciona o melhor resultado a lista de melhores resultados
melhoresResultadosDia.Add(outputsRedes[indMelhorRedeParaODiaAnterior][dia - 1]);
for (int indRede = 0; indRede < configuracaoCaptacao.RedesPrevisao.Count; indRede++)
{
//Verifica se a taxa de acerto do dia para a rede ja foi calculada
if (taxasAcertoRedes[indRede][dia] == 0)
{
//Ultiliza os ultimos melhores dados para fazer uma nova previsao
double[] inputRede = melhoresResultadosDia.Skip(melhoresResultadosDia.Count - configuracaoCaptacao.RedesPrevisao[indRede].JanelaEntrada).Take(configuracaoCaptacao.RedesPrevisao[indRede].JanelaEntrada).ToArray();
double[] outputRede = configuracaoCaptacao.RedesPrevisao[indRede].RedeNeuralPrevisaoFinanceiraPorDivisaoCrossValidation[treinamento.DivisaoCrossValidation].Run(inputRede);
//Adiciona o dia previsto na lista de outputs
outputsRedes[indRede].Add(outputRede.Last());
//Calcula a taxa de acerto da previsao
double ta = Math.Min(outputRede.Last(), treinamento.Output[dia]) / Math.Max(outputRede.Last(), treinamento.Output[dia]);
//Adiciona a taxa de acerto da rede para o dia
taxasAcertoRedes[indRede][dia] = ta;
//Atualiza a taxa de acerto da rede
configuracaoCaptacao.RedesPrevisao[indRede].TaxaMediaAcertoPorDia[dia] += ta;
}
}
}
for (int dia = 0; dia < totalDiasPrevisao; dia++)
{
double[] outputCaptacao = new double[configuracaoCaptacao.RedesPrevisao.Count];
for (int indRede = 0; indRede < configuracaoCaptacao.RedesPrevisao.Count; indRede++)
{
outputCaptacao[indRede] = taxasAcertoRedes[indRede][dia];
}
TrainingSample ts = new TrainingSample(treinamento.Input.ToArray(), outputCaptacao);
treinamentosPorDia[dia].Add(ts);
}
}
//Divide a taxa de acerto pela quantidade de treinamentos para saber a taxa média
foreach (RedePrevisaoFinanceira redePrevisao in configuracaoCaptacao.RedesPrevisao)
{
for (int i = 0; i < totalDiasPrevisao; i++)
{
redePrevisao.TaxaMediaAcertoPorDia[i] /= treinamentos.Count;
}
}
return treinamentosPorDia;
}
/// <summary> /// 一个受保护的帮助函数，用于训练单个学习样本 /// </summary> /// <param name="trainingSample"> /// 培训样品使用 /// </param> /// <param name="currentIteration"> /// 当前训练时期（假设为正且小于<c> trainingEpochs </ c>） /// </param> /// <param name="trainingEpochs"> /// 训练时期数（假定为正） /// </param> protected abstract void LearnSample(TrainingSample trainingSample, int currentIteration, int trainingEpochs);
/// <summary>
/// Removes the given training sample
/// </summary>
/// <param name="sample">
/// The sample to remove
/// </param>
/// <returns>
/// <c>true</c> if successful, <c>false</c> otherwise
/// </returns>
public bool Remove(TrainingSample sample)
{
return trainingSamples.Remove(sample);
}
/// <summary>
/// Trains the network for the given training sample (Online training mode). Note that this
/// method trains the sample only once, irrespective of what current epoch is. The arguments
/// are just used to find out training progress and adjust parameters depending on it.
/// </summary>
/// <param name="trainingSample">
/// Training sample to use
/// </param>
/// <param name="currentIteration">
/// Current training iteration
/// </param>
/// <param name="trainingEpochs">
/// Number of training epochs
/// </param>
/// <exception cref="ArgumentNullException">
/// If <c>trainingSample</c> is <c>null</c>
/// </exception>
/// <exception cref="ArgumentException">
/// If <c>trainingEpochs</c> is not positive
/// </exception>
/// <exception cref="ArgumentOutOfRangeException">
/// If <c>currentIteration</c> is negative or, if it is not less than <c>trainingEpochs</c>
/// </exception>
public virtual void Learn(TrainingSample trainingSample, int currentIteration, int trainingEpochs)
{
Helper.ValidateNotNull(trainingSample, "trainingSample");
Helper.ValidatePositive(trainingEpochs, "trainingEpochs");
Helper.ValidateWithinRange(currentIteration, 0, trainingEpochs - 1, "currentIteration");
OnBeginSample(currentIteration, trainingSample);
LearnSample(trainingSample, currentIteration, trainingEpochs);
OnEndSample(currentIteration, trainingSample);
}
/// <summary>
/// 创建此类的新实例
/// </summary>
/// <param name="trainingIteration">
/// 当前训练迭代
/// </param>
/// <param name="trainingSample">
/// 与事件相关联的训练样本
/// </param>
public TrainingSampleEventArgs(int trainingIteration, TrainingSample trainingSample)
{
this.trainingIteration = trainingIteration;
this.trainingSample = trainingSample;
}
/// <summary>
/// Invokes BeginSampleEvent
/// </summary>
/// <param name="currentIteration">
/// Current training iteration
/// </param>
/// <param name="currentSample">
/// Current sample which got trained successfully
/// </param>
protected virtual void OnEndSample(int currentIteration, TrainingSample currentSample)
{
if (EndSampleEvent != null)
{
EndSampleEvent(this, new TrainingSampleEventArgs(currentIteration, currentSample));
}
}
//Train agent's neural network for specific input and desired output
private void trainNeural(double[] input, double output)
{
double[] tmp = { output };
//Create the training sample for the neural network
TrainingSample sample = new TrainingSample(input, tmp);
//Train nn
network.Learn(sample, 0, currentEpoch);
}
