private unsafe neural_network NN_init(int layer_num, int[] layer_neuron_num)
{
//Init NN
fixed(int *p = layer_neuron_num)
{
neural_network nn = new neural_network(layer_num, p);
return(nn);
}
}
unsafe static private void init_NN()
{
//Read NN parameter file
if (!File.Exists("NN_Parameter"))
{
MessageBox.Show("Open NN_Parameter fail!!");
Console.WriteLine("Open NN_Parameter fail!!");
return;
}
FileStream fsSource = new FileStream("NN_Parameter", FileMode.Open, FileAccess.Read);
byte[] temp_byte = new byte[8];
fsSource.Read(temp_byte, 0, 4);
int layer_num = BitConverter.ToInt32(temp_byte, 0);
//init nn layer parameter
int[] layer_neuron_num = new int[layer_num];
for (int i = 0; i < layer_num; i++)
{
fsSource.Read(temp_byte, 0, 4); //int is 4bytes
layer_neuron_num[i] = BitConverter.ToInt32(temp_byte, 0);
}
//init nn weight parameter
fsSource.Read(temp_byte, 0, 4); //int is 4bytes
int weight_num = BitConverter.ToInt32(temp_byte, 0);
double[] weight = new double[weight_num];
for (int i = 0; i < weight_num; i++)
{
fsSource.Read(temp_byte, 0, 8); //double is 8bytes
weight[i] = BitConverter.ToDouble(temp_byte, 0);
}
fsSource.Read(temp_byte, 0, 4); //int is 4bytes
int bias_num = BitConverter.ToInt32(temp_byte, 0);
double[] bias = new double[bias_num];
for (int i = 0; i < bias_num; i++)
{
fsSource.Read(temp_byte, 0, 8); //double is 8bytes
bias[i] = BitConverter.ToDouble(temp_byte, 0);
}
fsSource.Close();
//nn = NN_init(layer_num, layer_neuron_num, weight, bias);
//Init NN
fixed(int *p = layer_neuron_num)
{
nn = new neural_network(layer_num, p);
fixed(double *p1 = weight)
{
fixed(double *p2 = bias)
{
nn.set_parameter(p1, p2);
}
}
}
}
private unsafe void NN_Training()
{
int epoch = int.Parse(textBox1.Text);
int iteration = int.Parse(textBox2.Text);
double learning_factor = double.Parse(textBox3.Text);
double up_bound = double.Parse(textBox5.Text);
double lower_bound = double.Parse(textBox4.Text);
int[] layer_neuron_num = new int[4];
layer_neuron_num[0] = 784; //Input layer Num
layer_neuron_num[1] = Int32.Parse(textBox6.Text); // hidden_layer1 Num
layer_neuron_num[2] = Int32.Parse(textBox7.Text); // hidden_layer2 Num
layer_neuron_num[3] = 10; //Output_layer, Output layer Neuron 0 ~ 9 represent number 0 ~ 9
byte[] image_data = new byte[784]; //MNistread return image
Int16 label_data; //MNistread return label
double[] input_data = new double[784]; //NN_train input data
double[] target_data = new double[10]; //NN_train target data
bool check_data = false; //verify the nn_train function
float [] error_log = new float [iteration];
float nn_error = 1; //compara every epoch performmance
for (int epoch_count = 0; epoch_count < epoch; epoch_count++)
{
//init NN
neural_network nn = NN_init(4, layer_neuron_num);
//Change Form3 epoch label
f3.Change_Label(epoch_count, false, 100);
//Training NN
for (int iteration_count = 0; iteration_count < iteration; iteration_count++)
{
//init error log
error_log[iteration_count] = 0;
//Change Form3 iteration label
f3.Change_iteration_label(iteration_count);
//Get Test Label & Image to train NN
for (int Label_count = 0; Label_count < 60000; Label_count++) //MNIST Train dataset number of item is 60000
{
//Read data
mnist.ReadImg("train-labels.idx1-ubyte", "train-images.idx3-ubyte", Label_count, image_data);
label_data = mnist.label;
//transform image_data to input data
for (int input_count = 0; input_count < 784; input_count++)
{
input_data[input_count] = image_data[input_count];
}
//Set Target Output data
for (int output_count = 0; output_count < 10; output_count++)
{
target_data[output_count] = 0;
}
target_data[label_data] = 1; //expect NN Output
//train_nn
fixed(double *p1 = target_data)
{
fixed(double *p2 = input_data)
{
//if check output is wrong then train NN
check_data = nn.Classification_train(p1, p2, learning_factor, up_bound, lower_bound);
}
}
if (check_data)
{
error_log[iteration_count]++;
}
}
error_log[iteration_count] = error_log[iteration_count] / 60000; //compute error average
//Save the best performance data
if (nn_error > error_log[iteration_count])
{
nn_error = error_log[iteration_count];
this.nn = nn;
}
}
//show error log
f3.show_to_graph(error_log, epoch_count);
}
//Train end
f3.Change_Label(epoch - 1, true, nn_error);
}