trainer_result_t batch_trainer_t::train(
const task_t& task, const fold_t& fold, const loss_t& loss, size_t nthreads, const string_t& criterion,
model_t& model) const
{
if (fold.second != protocol::train)
{
log_error() << "batch trainer: can only train models with training samples!";
return trainer_result_t();
}
// initialize the model
model.resize(task, true);
model.random_params();
// prune training & validation data
sampler_t tsampler(task);
tsampler.setup(fold).setup(sampler_t::atype::annotated);
sampler_t vsampler(task);
tsampler.split(80, vsampler);
if (tsampler.empty() || vsampler.empty())
{
log_error() << "batch trainer: no annotated training samples!";
return trainer_result_t();
}
// parameters
const size_t iterations = math::clamp(text::from_params<size_t>(configuration(), "iters", 1024), 4, 4096);
const scalar_t epsilon = math::clamp(text::from_params<scalar_t>(configuration(), "eps", 1e-4), 1e-8, 1e-3);
const optim::batch_optimizer optimizer = text::from_string<optim::batch_optimizer>
(text::from_params<string_t>(configuration(), "opt", "lbfgs"));
// train the model
const trainer_result_t result = ncv::batch_train(
model, task, tsampler, vsampler, nthreads,
loss, criterion, optimizer, iterations, epsilon);
const trainer_state_t state = result.optimum_state();
log_info() << "optimum [train = " << state.m_tvalue << "/" << state.m_terror_avg
<< ", valid = " << state.m_vvalue << "/" << state.m_verror_avg
<< ", epoch = " << result.optimum_epoch()
<< ", config = " << text::concatenate(result.optimum_config(), "/")
<< "].";
// OK
if (result.valid())
{
model.load_params(result.optimum_params());
}
return result;
}
// advection
void Fluid2::fluidAdvection( const float dt )
{
// ink
{
Array2<float> inkcopy( ink );
CellSampler inksampler( grid, inkcopy );
const Index2& size = ink.getSize();
for( unsigned int i = 0; i < size.x; ++i )
for( unsigned int j = 0; j < size.y; ++j )
{
const Index2 id( i, j );
const Vec2 pos( grid.getCellPos( id ) );
const Vec2 vel( ( velocityX[ id ] + velocityX[ Index2( i+1, j ) ] ) * 0.5f,
( velocityY[ id ] + velocityY[ Index2( i, j+1 ) ] ) * 0.5f );
const Vec2 endpos( pos - dt * vel );
ink[ id ] = inksampler.getValue( endpos );;
}
}
// velocity
{
Array2< float > ucopy( velocityX );
Array2< float > vcopy( velocityY );
FaceSampler usampler( grid, ucopy, 0 );
FaceSampler vsampler( grid, vcopy, 1 );
const Index2& sizeu = velocityX.getSize();
const Index2& sizev = velocityY.getSize();
for( unsigned int i = 0; i < sizeu.x; ++i )
for( unsigned int j = 0; j < sizeu.y; ++j )
{
const Index2 id( i, j );
const Index2 idv1( clamp( i-1, 0, sizev.x-1 ), clamp( j , 0, sizev.y-1 ) );
const Index2 idv2( clamp( i , 0, sizev.x-1 ), clamp( j , 0, sizev.y-1 ) );
const Index2 idv3( clamp( i-1, 0, sizev.x-1 ), clamp( j+1, 0, sizev.y-1 ) );
const Index2 idv4( clamp( i , 0, sizev.x-1 ), clamp( j+1, 0, sizev.y-1 ) );
const Vec2 pos( grid.getFaceXPos( id ) );
const Vec2 vel( ucopy[ id ], ( vcopy[ idv1 ] + vcopy[ idv2 ] + vcopy[ idv3 ] + vcopy[ idv4 ] ) * 0.25f );
const Vec2 endpos( pos - dt * vel );
velocityX[ id ] = usampler.getValue( endpos );
}
for( unsigned int i = 0; i < sizev.x; ++i )
for( unsigned int j = 0; j < sizev.y; ++j )
{
const Index2 id( i, j );
const Index2 idu1( clamp( i , 0, sizeu.x-1 ), clamp( j-1, 0, sizeu.y-1 ) );
const Index2 idu2( clamp( i , 0, sizeu.x-1 ), clamp( j , 0, sizeu.y-1 ) );
const Index2 idu3( clamp( i+1, 0, sizeu.x-1 ), clamp( j-1, 0, sizeu.y-1 ) );
const Index2 idu4( clamp( i+1, 0, sizeu.x-1 ), clamp( j , 0, sizeu.y-1 ) );
const Vec2 pos( grid.getFaceYPos( id ) );
const Vec2 vel( ( ucopy[ idu1 ] + ucopy[ idu2 ] + ucopy[ idu3 ] + ucopy[ idu4 ] ) * 0.25f, vcopy[ id ] );
const Vec2 endpos( pos - dt * vel );
velocityY[ id ] = vsampler.getValue( endpos );
}
}
}
