MCPcopy Create free account
hub / github.com/davisking/dlib / test_loss_mmod

Function test_loss_mmod

dlib/test/dnn.cpp:5594–5672  ·  view source on GitHub ↗

This test really just checks if the mmod loss goes negative when a whole lot of overlapping truth rectangles are given.

Source from the content-addressed store, hash-verified

5592 // This test really just checks if the mmod loss goes negative when a whole lot of overlapping
5593 // truth rectangles are given.
5594 void test_loss_mmod()
5595 {
5596 print_spinner();
5597
5598 // Define input image size.
5599 constexpr int nc = 20;
5600 constexpr int nr = 20;
5601
5602 constexpr int margin = 3;
5603
5604 // Create a checkerboard pattern.
5605 std::deque<point> labeled_points;
5606 for (int y = margin; y < nr - margin; ++y)
5607 for (int x = margin + 1 - y % 2; x < nc - margin; x += 2)
5608 labeled_points.emplace_back(x, y);
5609
5610 // Create training data that follows the generated pattern.
5611 typedef matrix<float> input_image_type;
5612
5613 const auto generate_input_image = [&labeled_points, nr, nc]()
5614 {
5615 input_image_type sample(nr, nc);
5616 sample = -1.0;
5617
5618 for (const auto& point : labeled_points)
5619 sample(point.y(), point.x()) = 1.0;
5620
5621 return sample;
5622 };
5623
5624 const auto generate_labels = [&labeled_points]()
5625 {
5626 const auto point_to_rect = [](const point& point) {
5627 constexpr int rect_size = 5;
5628 return centered_rect(
5629 point.x(), point.y(),
5630 rect_size, rect_size
5631 );
5632 };
5633
5634 std::vector<mmod_rect> labels;
5635
5636 std::transform(
5637 labeled_points.begin(),
5638 labeled_points.end(),
5639 std::back_inserter(labels),
5640 point_to_rect
5641 );
5642
5643 return labels;
5644 };
5645
5646 const input_image_type input_image = generate_input_image();
5647 const std::vector<mmod_rect> labels = generate_labels();
5648
5649 mmod_options options(use_image_pyramid::no, { labels });
5650 options.be_quiet = true;
5651

Callers 1

run_testsMethod · 0.85

Calls 9

print_spinnerFunction · 0.85
centered_rectFunction · 0.85
emplace_backMethod · 0.80
train_one_stepMethod · 0.80
get_average_lossMethod · 0.80
sgdClass · 0.50
beginMethod · 0.45
endMethod · 0.45
sizeMethod · 0.45

Tested by

no test coverage detected