public static void test_matrix_to_csv_and_back()
{
var result = new result_state();
var x = ml_funcs.matrix_create(2, 3);
var file_name = "test_matrix_to_csv_and_back.txt";
double[][] y;
x[0] = new double[] { 1.123, 2.12, 3.123 };
x[1] = new double[] { 4.123, 5.123, 6.123 };
result = ml_funcs.matrix_to_csv(file_name, x);
Console.Write("matrix_to_csv_and_back");
if (result.has_errors())
{
Console.WriteLine(" .. FAILED");
Console.WriteLine(result.all_errors_to_string());
return;
}
result = ml_funcs.matrix_from_csv(file_name, out y);
if (result.has_errors())
{
Console.WriteLine(" .. FAILED");
Console.WriteLine(result.all_errors_to_string());
return;
}
if (ml_funcs.matrix_compare_deep(x, y))
{
Console.WriteLine(" .. OK");
}
else
{
Console.WriteLine(" .. FAILED");
}
File.Delete(file_name);
}
public static void test_nn()
{
Console.Write("two_layer_nn ");
double lambda = 0;
const int input_layer_size = 400;
const int hidden_layer_size = 25;
const int output_layer_size = 10;
int training_examples = 0;
double cost;
double[] result_data, unrolled_theta, trained_theta, cost_list; // y
double[][] train_data, theta_1, theta_2; // X - train_data
double[][] theta1_gradient, theta2_gradient;
result_state result;
// Loading X and Y
{
string X_file_content;
result = file_utils.read_file("./data/ex4_data_X.txt", out X_file_content);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
train_data = string_utils.string_to_matrix(X_file_content, " ");
if (train_data.Length != 5000 || train_data[0].Length != 400)
{
Console.WriteLine(" .. FAILED. Should have 5000 training examples and 400 features");
return;
}
training_examples = train_data.Length;
result_data = new double[training_examples];
int y = 0, label = 10;
for (var i = 0; i < training_examples; i++)
{
result_data[i] = label;
if (++y == 500)
{
if (label == 10)
{
label = 1;
}
else
{
label++;
}
y = 0;
}
}
}
// Load weights
{
string file_content_1, file_content_2;
result = file_utils.read_file("./data/ex4_theta_1.txt", out file_content_1);
result.combine_errors(utils.file_utils.read_file("./data/ex4_theta_2.txt", out file_content_2));
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
theta_1 = string_utils.string_to_matrix(file_content_1, " ");
theta_2 = string_utils.string_to_matrix(file_content_2, " ");
unrolled_theta = matrix_flatten_two(theta_1, theta_2, flatten_direction.by_column);
if (unrolled_theta.Length != 10285)
{
Console.WriteLine(".. FAILED. Incorrect unrolled theta parameter count");
return;
}
}
// testing cost with initial thetas
if (1 == 0)
{
lambda = 0;
result = nn_cost_two_layer(train_data, result_data, matrix_transpose(theta_1), matrix_transpose(theta_2), output_layer_size, lambda, out cost, out theta1_gradient, out theta2_gradient);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
if (Math.Round(cost, 6) != 0.287629)
{
Console.WriteLine(".. FAILED. Cost with regularization off (labda 0) is incorrect");
return;
}
lambda = 1;
result = nn_cost_two_layer(train_data, result_data, matrix_transpose(theta_1), matrix_transpose(theta_2), output_layer_size, lambda, out cost, out theta1_gradient, out theta2_gradient);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
if (Math.Round(cost, 6) != 0.383770)
{
Console.WriteLine(".. FAILED. Cost with regularization off (labda 1) is incorrect");
return;
}
}
// training nn
{
theta_1 = nn_random_weights(input_layer_size + 1, hidden_layer_size);
theta_2 = nn_random_weights(hidden_layer_size + 1, output_layer_size);
lambda = 1;
result = nn_cost_two_layer(train_data, result_data, matrix_transpose(theta_1), matrix_transpose(theta_2), output_layer_size, lambda, out cost, out theta1_gradient, out theta2_gradient);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
cost_delegate nn_cost_delegate = (double[][] train_data, double[] result_data, double[] theta, double lambda, out double cost, out double[] gradient) => {
var result = new result_state();
cost = 0;
gradient = null;
// 1. convert theta back to neural network layers
var theta1 = matrix_unflatten(theta, hidden_layer_size, 0, (input_layer_size + 1) * hidden_layer_size - 1);
var theta2 = matrix_unflatten(theta, output_layer_size, (input_layer_size + 1) * hidden_layer_size);
// 2. pass to nn_cost_two_layer neural network thetas
result = nn_cost_two_layer(train_data, result_data, theta1, theta2, output_layer_size, lambda, out cost, out theta1_gradient, out theta2_gradient);
Console.WriteLine($"Cost: {cost}");
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return(result);
}
// 3. we get theta1, theta2 gradients then we flattan them into gradient
gradient = matrix_flatten_two(theta1_gradient, theta2_gradient);
return(result);
};
result = rasmussen(train_data, result_data, unrolled_theta, lambda, max_iterations: 3, nn_cost_delegate, out cost_list, out trained_theta);
}
Console.WriteLine(".. OK");
}
public static void test_rasmussen(bool force_training = false)
{
var result = new result_state();
const string trained_theta_file_path = "./data/ex3_trained_theta.txt";
const string cost_progression_file_path = "./data/ex3_cost_progression.txt";
const double lambda = 0.1d;
const int max_iterations = 10;
var file_path = "./data/ex3data1.txt";
double[][] train_data, trained_theta, cost_progression;
double[] result_data, label_result_data, initial_theta;
int i, labels, label_train_count;
int[] predict_indices;
var parse_result = file_utils.parse_file(file_path, out train_data, out result_data);
if (parse_result.has_errors())
{
Console.WriteLine(parse_result.all_errors_to_string());
return;
}
label_train_count = result_data.Length / 10; // how many training examples we have for each label (numbers from 0 to 9)
trained_theta = ml_funcs.matrix_create(10, train_data[0].Length);
cost_progression = ml_funcs.matrix_create(10, 10);
// don't retrain
// delete file if you want to retrain
if (!File.Exists(trained_theta_file_path) || force_training)
{
for (labels = 0; labels < 10; labels++)
{
Console.WriteLine($"Training label: {labels}");
label_result_data = new double[train_data.Length];
for (i = 0; i < label_train_count; i++)
{
label_result_data[label_train_count * labels + i] = 1;
}
initial_theta = new double[train_data[0].Length + 1];
result = ml_funcs.rasmussen(train_data, label_result_data, initial_theta, lambda, max_iterations, ml_funcs.cost_logistic_regression_regularized, out cost_progression[labels], out trained_theta[labels]);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
}
result = ml.ml_funcs.matrix_to_csv(trained_theta_file_path, trained_theta);
var write_vector_result = ml_funcs.matrix_to_csv(cost_progression_file_path, cost_progression);
result.combine_errors(write_vector_result);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
}
else
{
result = ml_funcs.matrix_from_csv(trained_theta_file_path, out trained_theta);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
}
result = ml_funcs.predict_one_vs_all(trained_theta, train_data, out predict_indices);
if (result.has_errors())
{
Console.WriteLine(result.all_errors_to_string());
return;
}
var correct_predictions = 0;
for (i = 0; i < result_data.Length; i++)
{
if (result_data[i] == predict_indices[i])
{
correct_predictions++;
}
}
double training_accuracy = (double)correct_predictions / (double)result_data.Length * 100;
Console.Write("rasmussen");
if (training_accuracy > 90d)
{
Console.WriteLine(" .. OK");
}
else
{
Console.WriteLine(" .. FAILED");
}
// Console.WriteLine($"Training accuracy with {max_iterations} iterations is {training_accuracy}%");
}
