DNN的准确性较低

如何解决DNN的准确性较低

我最近一直基于http://neuralnetworksanddeeplearning.com/实现NN。我已经完成了关于反向传播和SGD的整个算法，几乎与本书作者所采用的方法相同。问题是，尽管他具有相同的hiperparameters，但在一个历元之后他获得90％左右的准确性时，在5个历元之后我却获得了30％的准确性。您知道可能是什么原因吗？ 这是我的资料库。 https://github.com/PiPower/Deep-Neural-Network 这是Network.cpp中实现的反向传播算法和SGD算法的一部分：

    void Network::Train(MatrixD_Array& TrainingData,MatrixD_Array& TrainingLabels,int BatchSize,int epochs,double LearningRate)
{
    

assert(TrainingData.size() == TrainingLabels.size() && CostFunc != nullptr && CostFuncDer != nullptr && LearningRate > 0);
    std::vector<long unsigned int > indexes;
    for (int i = 0; i < TrainingData.size(); i++)  indexes.push_back(i);

    std::random_device rd;
    std::mt19937 g(rd());

    std::vector<Matrix<double>> NablaWeights;
    std::vector<Matrix<double>> NablaBiases;


    NablaWeights.resize(Layers.size());
    NablaBiases.resize(Layers.size());


    for (int i = 0; i < Layers.size(); i++)
    {
        NablaWeights[i] = Matrix<double>(Layers[i].GetInDim(),Layers[i].GetoutDim());
        NablaBiases[i] = Matrix<double>(1,Layers[i].GetoutDim());
    }

    //---- Epoch iterating
    for (int i = 0; i < epochs; i++)
    {
        cout << "Epoch number: " << i << endl;
        shuffle(indexes.begin(),indexes.end(),g);

        // Batch iterating 
        for (int batch = 0; batch < TrainingData.size(); batch = batch + BatchSize)
        {
            for (int i = 0; i < Layers.size(); i++)
            {
                NablaWeights[i].Clear();
                NablaBiases[i].Clear();
            }


            int i = 0;
            while( i < BatchSize && (i+batch)< TrainingData.size())
            {
                
                std::vector<Matrix<double>> ActivationOutput;
                std::vector<Matrix<double>> Z_Output;
                ActivationOutput.resize(Layers.size() + 1);
                Z_Output.resize(Layers.size());

                ActivationOutput[0] = TrainingData[indexes[i + batch]];
                int index = 0;
                // Pushing values through
                for (auto layer : Layers)
                {
                    Z_Output[index] = layer.Mul(ActivationOutput[index]);
                    ActivationOutput[index + 1] = layer.ApplyActivation(Z_Output[index]);
                    index++;
                }
                // ---- Calculating Nabla that will be later devided by batch size element wise 
                auto DeltaNabla = BackPropagation(ActivationOutput,Z_Output,TrainingLabels[indexes[i + batch]]);

                for (int i = 0; i < Layers.size(); i++)
                {
                    NablaWeights[i] = NablaWeights[i] + DeltaNabla.first[i];
                    NablaBiases[i] = NablaBiases[i] + DeltaNabla.second[i];
                }
            i++;
            }

            for (int g = 0; g < Layers.size(); g++)
            {
                
                Layers[g].Weights = Layers[g].Weights - NablaWeights[g] * LearningRate;
                Layers[g].Biases = Layers[g].Biases - NablaBiases[g] * LearningRate;
            }
           // std::transform(NablaWeights.begin(),NablaWeights.end(),NablaWeights.begin(),[BatchSize,LearningRate](Matrix<double>& Weight) {return Weight * (LearningRate / BatchSize);});
            //std::transform(NablaBiases.begin(),NablaBiases.end(),NablaBiases.begin(),LearningRate](Matrix<double>& Bias) {return Bias * (LearningRate / BatchSize); });

        }
    }
}

std::pair<MatrixD_Array,MatrixD_Array> Network::BackPropagation( MatrixD_Array& ActivationOutput,MatrixD_Array& Z_Output,Matrix<double>& label)
{
    MatrixD_Array NablaWeight;
    MatrixD_Array NablaBias;

    NablaWeight.resize(Layers.size());
    NablaBias.resize(Layers.size());

    auto zs = Layers[Layers.size() - 1].ActivationPrime(Z_Output[Z_Output.size() - 1]);
    Matrix<double> Delta_L = Hadamard(CostFuncDer(ActivationOutput[ActivationOutput.size() - 1],label),zs);

    NablaWeight[Layers.size() - 1] = Delta_L * ActivationOutput[ActivationOutput.size() - 2].Transpose();
    NablaBias[Layers.size() - 1] = Delta_L;

    for (int j = 2; j <= Layers.size() ; j++)
    {
        auto sp = Layers[Layers.size() - j].ActivationPrime(Z_Output[Layers.size() -j]);
        Delta_L = Hadamard(Layers[Layers.size() -  j+1 ].Weights.Transpose() * Delta_L,sp);

        NablaWeight[Layers.size() - j] = Delta_L * ActivationOutput[ActivationOutput.size() -j-1].Transpose();
        NablaBias[Layers.size() - j] = Delta_L;
    }

    return make_pair(NablaWeight,NablaBias);
}

解决方法

事实证明，mnist加载程序无法正常工作。