| 4117 | // ---------------------------------------------------------------------------------------- |
| 4118 | |
| 4119 | void test_simple_autoencoder() |
| 4120 | { |
| 4121 | print_spinner(); |
| 4122 | |
| 4123 | srand(1234); |
| 4124 | |
| 4125 | const int output_width = 7; |
| 4126 | const int output_height = 7; |
| 4127 | const int num_samples = 100; |
| 4128 | ::std::vector<matrix<float>> x(num_samples); |
| 4129 | |
| 4130 | matrix<float> tmp(output_width, output_height); |
| 4131 | for (int i = 0; i < num_samples; ++i) |
| 4132 | { |
| 4133 | const int model = i % 4; |
| 4134 | |
| 4135 | for (int r = 0; r < output_height; ++r) |
| 4136 | for (int c = 0; c < output_width; ++c) |
| 4137 | switch (model) { |
| 4138 | case 0: tmp(r, c) = r / output_height; break; |
| 4139 | case 1: tmp(r, c) = c / output_width; break; |
| 4140 | case 2: tmp(r, c) = 1.0 - r / output_height; break; |
| 4141 | case 3: tmp(r, c) = 1.0 - c / output_width; break; |
| 4142 | default: DLIB_TEST_MSG(false, "Invalid model: " << model << " (should be between 0 and 3)"); |
| 4143 | } |
| 4144 | |
| 4145 | x[i] = tmp; |
| 4146 | } |
| 4147 | |
| 4148 | using net_type = loss_mean_squared_per_pixel< |
| 4149 | cont<1,output_height,output_width,2,2, |
| 4150 | relu<con<4,output_height,output_width,2,2, |
| 4151 | input<matrix<float>>>>>>; |
| 4152 | net_type net; |
| 4153 | |
| 4154 | const auto autoencoder_error = [&x, &net, &output_height, &output_width]() |
| 4155 | { |
| 4156 | const auto y = net(x); |
| 4157 | double error = 0.0; |
| 4158 | for (size_t i = 0; i < x.size(); ++i) |
| 4159 | for (int r = 0; r < output_height; ++r) |
| 4160 | for (int c = 0; c < output_width; ++c) |
| 4161 | error += fabs(y[i](r, c) - x[i](r, c)); |
| 4162 | |
| 4163 | return error / (x.size() * output_height * output_width); |
| 4164 | }; |
| 4165 | |
| 4166 | // The autoencoder can't be very good before it's been trained |
| 4167 | // (or at least the probability of the reconstruction error |
| 4168 | // being small should be super low; in fact, the error ought to |
| 4169 | // be much higher than 0.01, however since the initialization |
| 4170 | // is random, putting the limit below too high could make the |
| 4171 | // tests fail when other, unrelated tests are added into the |
| 4172 | // sequence) |
| 4173 | const double error_before = autoencoder_error(); |
| 4174 | DLIB_TEST_MSG(error_before > 0.01, "Autoencoder error before training = " << error_before); |
| 4175 | |
| 4176 | // Make sure there's an information bottleneck, as intended |
no test coverage detected