lisyarus/psemek
1
0
Fork 1
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Remove obsolete 'optimize' mode from neural animation example

Browse source
This commit is contained in:
Nikita Lisitsa 2021-01-15 22:37:53 +03:00
parent f73f536d8d
commit 25e0480805
1 changed files with 3 additions and 75 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

78
examples/animation_2d.cpp
View file

@ -669,13 +669,12 @@ struct animation_2d_app
enum class mode
{
train,
optimize,
test,
} mode = mode::train;
std::vector<controller> population;
std::size_t const population_size = 256;
std::size_t const max_train_frames = 60.f / physics.dt;
std::size_t const population_size = 16;
std::size_t const max_train_frames = 10.f / physics.dt;
std::size_t const max_train_variations = 1;
float const position_variation_amplitude = 0.f;
float const angle_variation_amplitude = 0.f;
@ -686,10 +685,6 @@ struct animation_2d_app
static constexpr auto mutation_amplitude = [](float t){ return 10.f * geom::lerp(1.f, 0.01f, t); };
// static constexpr auto mutation_amplitude = [](float){ return 0.00001f; };
std::size_t optimize_iterations = 0;
std::size_t const max_optimize_iterations = 1;
float const optimize_amplitude = 0.1f;
float best_score = 0.f;
bool const warm_start = false;
bool const enable_testing = true;
@ -820,18 +815,6 @@ void animation_2d_app::update()
{
do_train();
if (train_iterations >= max_train_iterations)
{
mode = mode::optimize;
}
else if (train_iterations >= max_train_iterations / 2 && best_score < -1000.f)
{
// train_iterations = 0;
}
}
else if (mode == mode::optimize)
{
do_optimize();
if (optimize_iterations >= max_optimize_iterations)
{
mode = mode::test;
frame_clock.restart();
@ -1243,61 +1226,6 @@ void animation_2d_app::do_train()
best_score = scores.front().first;
}
void animation_2d_app::do_optimize()
{
auto & c = population[0];
auto rng = this->rng;
float current_score = eval_score(c, rng);
float eps = 0.1f;
std::vector<float> scores(c.param_count, 0.f);
std::atomic<std::size_t> dispatched{0};
for (std::size_t i = 0; i < c.param_count; ++i)
{
bg.dispatch([&, i, rng = rng]() mutable {
controller cc = c;
cc.params()[i] += eps;
scores[i] = eval_score(cc, rng);
++dispatched;
});
}
// bg.wait();
while (dispatched.load() < scores.size());
for (auto & s : scores)
s = (s - current_score) / eps;
float m = 0.f;
for (auto s : scores)
m = std::max(m, std::abs(s));
log::info() << "Max gradient: " << m;
for (std::size_t k = 0; k < 30; ++k)
{
auto cc = c;
for (std::size_t i = 0; i < c.param_count; ++i)
cc.params()[i] += optimize_amplitude * scores[i];
float new_score = eval_score(cc, rng);
if (new_score > current_score)
{
log::info() << "k = " << k;
log::info() << "Score change: " << current_score << " -> " << new_score;
c = cc;
break;
}
for (auto & s : scores)
s /= 2.f;
}
++optimize_iterations;
}
void animation_2d_app::do_test()
{
std::optional<geom::point<float, 2>> m;
@ -1466,7 +1394,7 @@ void animation_2d_app::present()
opts.x = gfx::painter::x_align::left;
opts.y = gfx::painter::y_align::top;
opts.scale = 2.f;
painter.text({40.f, 40.f}, util::to_string(train_iterations, "/", max_train_iterations, " ", optimize_iterations, "/", max_optimize_iterations), opts);
painter.text({40.f, 40.f}, util::to_string(train_iterations, "/", max_train_iterations), opts);
painter.text({40.f, 64.f}, util::to_string("Best score: ", std::setprecision(10), best_score), opts);
painter.text({40.f, 88.f}, util::to_string("Model: ", test_id, "/", population.size(), ", gen ", population[test_id].generation), opts);
painter.text({40.f, 112.f}, util::to_string(util::pretty(test_clock.duration(), std::chrono::milliseconds{1})), opts);