using System;

using System.Diagnostics;

using Numpy;

namespace NeuralNetworkExample

{

class Program

{

static void Main(string[] args)

{

Console.WriteLine("Fitting random data with a two layer Neural Network ...");

// N is batch size; D_in is input dimension;

// H is hidden dimension; D_out is output dimension.

var (N, D_in, H, D_out) = (64, 1000, 100, 10);

// Create random input and output data

Console.WriteLine("\tcreating random data");

var x = np.random.randn(N, D_in);

var y = np.random.randn(N, D_out);

Console.WriteLine("\tlearning");

var stopwatch = Stopwatch.StartNew();

// Randomly initialize weights

var w1 = np.random.randn(D_in, H);

var w2 = np.random.randn(H, D_out);

var learning_rate = 1.0e-6;

double loss=double.MaxValue;

for (int t = 0; t < 500; t++)

{

// Forward pass: compute predicted y

var h = x.dot(w1);

var h_relu = np.maximum(h, (NDarray)0);

var y_pred = h_relu.dot(w2);

// Compute and print loss

loss = (double)(np.square(y_pred - y).sum());

if (t%20==0)

Console.WriteLine($"\tstep: {t} loss: {loss}");

// Backprop to compute gradients of w1 and w2 with respect to loss

var grad_y_pred = 2.0 * (y_pred - y);

var grad_w2 = h_relu.T.dot(grad_y_pred);

var grad_h_relu = grad_y_pred.dot(w2.T);

var grad_h = grad_h_relu.copy();

grad_h[h < 0] = (NDarray)0;

var grad_w1 = x.T.dot(grad_h);

// Update weights

w1.isub( learning_rate * grad_w1); // inplace substraction is faster than -=

w2.isub( learning_rate * grad_w2);

}

stopwatch.Stop();

Console.WriteLine($"\tstep: 500, final loss: {loss}, elapsed time: {stopwatch.Elapsed.TotalSeconds:F3} seconds\n");

Console.WriteLine("Hit any key to exit.");

Console.ReadKey();

}

}

}