| 5333 | // ---------------------------------------------------------------------------------------- |
| 5334 | |
| 5335 | void test_loss_dot() |
| 5336 | { |
| 5337 | print_spinner(); |
| 5338 | |
| 5339 | std::vector<matrix<float,0,1>> samples; |
| 5340 | std::vector<matrix<float,0,1>> labels; |
| 5341 | |
| 5342 | const matrix<float> proj = matrix_cast<float>(randm(2,3)); |
| 5343 | for (int i = 0; i < 128; ++i) |
| 5344 | { |
| 5345 | // The task is going to be to learn the matrix proj. So we make our |
| 5346 | // training data thusly: |
| 5347 | matrix<float,0,1> x = matrix_cast<float>(randm(3,1)); |
| 5348 | matrix<float,0,1> y = normalize(proj*x); |
| 5349 | samples.push_back(x); |
| 5350 | labels.push_back(y); |
| 5351 | } |
| 5352 | |
| 5353 | using net_type = loss_dot< |
| 5354 | l2normalize<fc_no_bias<2, |
| 5355 | input<matrix<float,0,1>> |
| 5356 | >>>; |
| 5357 | |
| 5358 | net_type net; |
| 5359 | dnn_trainer<net_type> trainer(net, sgd(1e-4, 0.9)); |
| 5360 | trainer.set_learning_rate(0.01); |
| 5361 | trainer.set_min_learning_rate(0.0000001); |
| 5362 | trainer.set_mini_batch_size(128); |
| 5363 | trainer.set_max_num_epochs(50000); |
| 5364 | trainer.train(samples, labels); |
| 5365 | |
| 5366 | |
| 5367 | for (size_t i = 0; i < samples.size(); ++i) |
| 5368 | { |
| 5369 | DLIB_TEST(std::abs(1-dot(net(samples[i]),labels[i])) < 0.001); |
| 5370 | } |
| 5371 | } |
| 5372 | |
| 5373 | // ---------------------------------------------------------------------------------------- |
| 5374 |
no test coverage detected