| 4054 | // ---------------------------------------------------------------------------------------- |
| 4055 | |
| 4056 | void test_multioutput_linear_regression() |
| 4057 | { |
| 4058 | const int num_outputs = 2; |
| 4059 | const int num_samples = 1000; |
| 4060 | ::std::vector<matrix<double>> x(num_samples); |
| 4061 | ::std::vector<matrix<float>> y(num_samples); |
| 4062 | ::std::default_random_engine generator(16); |
| 4063 | ::std::normal_distribution<float> distribution(0,0.1); |
| 4064 | ::std::normal_distribution<float> slope_distribution(10,5); |
| 4065 | ::std::normal_distribution<float> intercept_distribution(50,10); |
| 4066 | ::std::vector<float> true_intercepts(num_outputs); |
| 4067 | ::std::vector<float> true_slopes(num_outputs); |
| 4068 | for ( int jj = 0; jj < num_outputs; ++jj ) |
| 4069 | { |
| 4070 | true_slopes[jj] = slope_distribution(generator); |
| 4071 | true_intercepts[jj] = intercept_distribution(generator); |
| 4072 | } |
| 4073 | matrix<float> ytmp(num_outputs, 1); |
| 4074 | for ( int ii = 0; ii < num_samples; ++ii ) |
| 4075 | { |
| 4076 | const double val = static_cast<double>(ii)/10; |
| 4077 | matrix<double> tmp(1,1); |
| 4078 | tmp = val; |
| 4079 | x[ii] = tmp; |
| 4080 | for ( int jj = 0; jj < num_outputs; ++jj ) |
| 4081 | ytmp(jj, 0) = (true_intercepts[jj] + true_slopes[jj]*static_cast<float>(val) + distribution(generator)); |
| 4082 | |
| 4083 | y[ii] = ytmp; |
| 4084 | } |
| 4085 | |
| 4086 | using net_type = loss_mean_squared_multioutput<fc<num_outputs, input<matrix<double>>>>; |
| 4087 | net_type net; |
| 4088 | layer<1>(net).layer_details().set_bias_learning_rate_multiplier(900); |
| 4089 | sgd defsolver(0,0.9); |
| 4090 | dnn_trainer<net_type> trainer(net, defsolver); |
| 4091 | trainer.set_learning_rate(1e-5); |
| 4092 | trainer.set_min_learning_rate(1e-6); |
| 4093 | trainer.set_mini_batch_size(50); |
| 4094 | trainer.set_max_num_epochs(170); |
| 4095 | trainer.train(x, y); |
| 4096 | |
| 4097 | float slope_error = 0.0; |
| 4098 | float intercept_error = 0.0; |
| 4099 | const float eps_slope = 0.05, eps_intercept = 0.1; |
| 4100 | |
| 4101 | for ( int jj = 0; jj < num_outputs; ++jj ) |
| 4102 | { |
| 4103 | slope_error += abs(layer<1>(net).layer_details().get_weights().host()[jj] - true_slopes[jj]); |
| 4104 | intercept_error += abs(layer<1>(net).layer_details().get_biases().host()[jj] - true_intercepts[jj]); |
| 4105 | } |
| 4106 | |
| 4107 | slope_error /= float(num_outputs); |
| 4108 | intercept_error /= float(num_outputs); |
| 4109 | |
| 4110 | DLIB_TEST_MSG(slope_error <= eps_slope, |
| 4111 | "Average absolute slope error = " << slope_error << " Error limit = " << eps_slope); |
| 4112 | DLIB_TEST_MSG(intercept_error <= eps_intercept, |
| 4113 | "Average absolute intercept error = " << intercept_error << " Error limit = " << eps_intercept); |
no test coverage detected