Нейронная сеть C# MNIST увеличивает точность примерно до 20 %, а затем снова падает до 10 %. ⇐ C#

[Anonymous]

Я пытался использовать этот код Python MNIST для написания нейронной сети на C# — https://www.kaggle.com/code/wwsalmon/si ... -numpy-no- tf-keras/notebook - но после многих попыток я не могу добиться значительного прогресса, кажется, что точность всегда составляет около 20%, прежде чем она снова падает до 10%? Пожалуйста, помогите.

Код: Выделить всё

namespace neuralNetwork3
{
public partial class Form1 : Form
{
Random rnd = new Random();

string[] trainingData = File.ReadAllLines(@"file path");
string[] testingData = File.ReadAllLines(@"file path");
double[,] X_train;
int m;
double[,] W1 = new double[10, 784];
double[] b1 = new double[10];
double[,] W2 = new double[10, 10];
double[] b2 = new double[10];

double[,] Z1;
double[,] Z2;
double[,] A1;
double[,] A2;
public Form1()
{
InitializeComponent();
}

private void Form1_Load(object sender, EventArgs e)
{
X_train = new double[trainingData.Length,784];
for (int i = 0; i < X_train.GetLength(0); i++)
{
double[] lineVals = trainingData[i][2..].Split(',').Select(x => double.Parse(x) / 255f).ToArray();
for (int j = 0; j < X_train.GetLength(1); j++)
{
X_train[i,j] = lineVals[j];
}
}

m = trainingData.Length;
Z1 = new double[10, m];
Z2 = new double[10, m];
A1 = new double[10, m];
A2 = new double[10, m];
}
void init_params()
{
for (int i = 0; i < W1.GetLength(0); i++)
{
for (int j = 0; j < W1.GetLength(1); j++)
{
W1[i, j] = rnd.NextDouble() - 0.5d;
}
}
for (int i = 0; i < b1.GetLength(0); i++)
{
b1[i] = rnd.NextDouble() - 0.5d;
}
for (int i = 0; i < W2.GetLength(0); i++)
{
for (int j = 0; j < W2.GetLength(1); j++)
{
W2[i, j] = rnd.NextDouble() - 0.5d;
}
}
for (int i = 0; i < b1.GetLength(0); i++)
{
b2[i] = rnd.NextDouble() - 0.5d;
}
}
double ReLU(double Z)
{
return Math.Max(0, Z);
}
double ReLU_deriv(double Z)
{
return Z > 0 ? 1:0;
}
int[,] one_hot(int[] Y)
{
int[,] encoded = new int[10, trainingData.Length];
for (int i = 0; i < encoded.GetLength(0); i++)
{
for (int j = 0; j < encoded.GetLength(1); j++)
{
encoded[i, j] = Y[j] == i ? 1:0;
}
}
return encoded;
}
void backward_prop(int[] Y, double learningRate)
{
int[,] oneHotY = one_hot(Y);
double[,] dZ2 = new double[10, m];
for (int i = 0; i < dZ2.GetLength(0); i++)
{
for (int j = 0; j < dZ2.GetLength(1); j++)
{
dZ2[i, j] = 2*(A2[i,j] - oneHotY[i,j]);
}
}
double[,] dW2 = new double[10, 10];
double[,] dZ2dotA1 = MProduct(dZ2, MTranspose(A1));
for (int i = 0; i < dW2.GetLength(0); i++)
{
for (int j = 0; j < dW2.GetLength(1); j++)
{
dW2[i, j] = (double)1 / m * dZ2dotA1[i,j];
}
}
double[] dB2 = new double[10];
for (int i = 0; i < dB2.Length; i++)
{
dB2[i] = (double)1 / m * Enumerable.Range(0, m).Select(x => dZ2[i,x]).Sum();//dZ2[i, j];
}
double[,] dZ1 = new double[10, m];
double[,] W2dotdZ2 = MProduct(MTranspose(W2), dZ2);
double[,] reluZ1 = new double[Z1.GetLength(0), Z1.GetLength(1)];
for (int i = 0; i < reluZ1.GetLength(0); i++)
{
for (int j = 0; j < reluZ1.GetLength(1);  j++)
{
reluZ1[i, j] = ReLU_deriv(Z1[i, j]);
}
}
dZ1 = MProduct(MTranspose(W2), dZ2);
for (int i = 0; i < dZ1.GetLength(0); i++)
{
for (int j = 0; j < dZ1.GetLength(1); j++)
{
dZ1[i, j] *= reluZ1[i, j];
}
}
double[,] dW1 = new double[10, 784];
double[,] dZ1dotX = MProduct(dZ1, X_train);
for (int i = 0; i < dW1.GetLength(0); i++)
{
for (int j = 0; j < dW1.GetLength(1); j++)
{
dW1[i, j] = (double)1 / m * dZ1dotX[i, j];
}
}
double[] dB1 = new double[10];
for (int i = 0; i < dB1.Length; i++)
{
dB1[i] = (double)1 / m * Enumerable.Range(0, m).Select(x => dZ1[i, x]).Sum();//dZ1[i, j];
}

//update params

for (int i = 0; i < W1.GetLength(0); i++)
{
for (int j = 0; j < W1.GetLength(1); j++)
{
W1[i, j] = W1[i,j] - learningRate * dW1[i,j];
}
}
for (int i = 0; i < W2.GetLength(0); i++)
{
for (int j = 0; j < W2.GetLength(1); j++)
{
W2[i, j] = W2[i, j] - learningRate * dW2[i, j];
}
}
for (int i = 0; i < b1.Length; i++)
{
b1[i] = b1[i] - learningRate * dB1[i];
}
for (int i = 0; i < b2.Length; i++)
{
b2[i] = b2[i] - learningRate * dB2[i];
}
}
void gradientDescent(int[] Y, int iterations)
{
init_params();
for (int i = 0; i < iterations; i++)
{
updateIteration(i.ToString());
forward_prop();
backward_prop(Y,0.01f);
testAll(i);
Application.DoEvents();
}
}
double[] softmax(double[] Z)
{
double[] A = new double[Z.Length];
double[] exp = new double[10];
for (int i = 0; i < Z.Length; i++)
{
exp[i] = Math.Exp(Z[i] - Z.Max());
}
for (int i = 0; i < Z.Length; i++)
{
A[i] = exp[i] / exp.Sum();
}
return A;
}
void forward_prop()
{
double[,] W1dotX = MProduct(W1, MTranspose(X_train));
for (int i = 0; i < Z1.GetLength(0); i++)
{
for (int j = 0; j < Z1.GetLength(1); j++)
{
Z1[i, j] = W1dotX[i, j] + b1[i];
}
}
for (int i = 0; i < A1.GetLength(0); i++)
{
for (int j = 0; j < A1.GetLength(1); j++)
{
A1[i, j] = ReLU(Z1[i, j]);
}
}
double[,] W2dotA1 = MProduct(W2, A1);
for (int i = 0; i < Z2.GetLength(0); i++)
{
for (int j = 0; j < Z2.GetLength(1); j++)
{
Z2[i, j] = W2dotA1[i, j] + b2[i];
}
}
for (int i = 0; i < Z2.GetLength(0); i++)
{
double[] softZ2 = softmax(Enumerable.Range(0, Z2.GetLength(0)).Select(x => Z2[x, i]).ToArray());
for (int j = 0; j < Z2.GetLength(0); j++)
{
A2[i, j] = softZ2[j];
}
}
}
double[,] MProduct(double[,] M1, double[,] M2)
{
double[,] M3 = new double[M1.GetLength(0), M2.GetLength(1)];
for (int i = 0; i < M1.GetLength(0); i++)
{
for (int j = 0; j < M2.GetLength(1); j++)
{
double sum = 0d;
for (int k = 0; k < M1.GetLength(1);  k++)
{
sum += M1[i, k] * M2[k, j];
}
M3[i, j] = sum;
}
}
return M3;
}
double[,] MTranspose(double[,] M1)
{
double[,] M2 = new double[M1.GetLength(1), M1.GetLength(0)];

for (int i = 0; i < M1.GetLength(0); i++)
{
for (int j = 0; j < M1.GetLength(1); j++)
{
M2[j, i] = M1[i, j];
}
}

return M2;
}

private void trainButton_Click(object sender, EventArgs e)
{
int[] Y = new int[trainingData.Length];
for (int i = 0; i < Y.Length; i++)
{
Y[i] = int.Parse(trainingData[i][0].ToString());
}
gradientDescent(Y,250);
MessageBox.Show("done");
}

private void testButton_Click(object sender, EventArgs e)
{
int num;
double[] O = new double[10];
if (testTextBox.Text == "*")
{
testAll(1);
}
else
{
try
{
num = int.Parse(testTextBox.Text);
}
catch
{
return;
}
showImg(testingData[num]);
O = testExample(testingData[num][2..].Split(',').Select(x => double.Parse(x) / 255f).ToArray());
MessageBox.Show(String.Join(',', O) + " : " + Array.IndexOf(O, O.Max()));
}
}
private void testAll(int iteration)
{
int num;
double[] O = new double[10];
int numCorrect = 0;
for (int i = 0; i < testingData.Length; i++)
{
O = testExample(testingData[i][2..].Split(',').Select(x => double.Parse(x) / 255f).ToArray());
if (Array.IndexOf(O, O.Max()) == int.Parse(testingData[i][0].ToString()))
{
numCorrect++;
}
}
iterationLabel.Text = iteration.ToString() + " : " + ((double)numCorrect / testingData.Length * 100).ToString() + "% correct";
}
double[] testExample(double[] example)
{
double[] H = new double[10];
double [] O = new double[10];

for (int i = 0; i < W1.GetLength(0); i++)
{
double sum = 0f;
for (int j = 0; j < W1.GetLength(1); j++)
{
sum += W1[i,j] * example[j];
}
H[i] = sum + b1[i];
}
for (int i = 0; i < W2.GetLength(0); i++)
{
double sum = 0f;
for (int j = 0; j < W2.GetLength(1); j++)
{
sum += W2[i, j] * H[j];
}
O[i] = sum + b2[i];
}
return O;
}
async void showImg(string line)
{
//MessageBox.Show(trainingData[0]);
Bitmap bmp = new Bitmap(28, 28);
List pixelBrightnesses = new List();

pixelBrightnesses = line.Split(',').Select(int.Parse).ToList();

for (int i = 0; i < 28; i++)
{
for (int j = 0; j < 28; j++)
{
int colour = int.Parse(pixelBrightnesses[j * 28 + i + 1].ToString());
bmp.SetPixel(i, j, Color.FromArgb(colour, colour, colour));
}
}

pictureBox1.Image = bmp;
await Task.Delay(500);
}
async void updateIteration(string iteration)
{
//MessageBox.Show(iteration.ToString());
iterationLabel.Text = "iteration: " + iteration;
await Task.Delay(500);
}
}
}

Я использую формы Windows C#, заранее спасибо.
В своем текущем состоянии он «обучается» с точностью примерно от 9% до 15. -20%, а затем упадет до 10%, хотя в примере на Python, который я использую, с этим проблем не было.

Подробнее здесь: https://stackoverflow.com/questions/786 ... ls-back-to