void
TempoTrackV2::calculateBeatPeriod(const vector<double> &df,
vector<double> &beat_period,
vector<double> &tempi)
{
// to follow matlab.. split into 512 sample frames with a 128 hop size
// calculate the acf,
// then the rcf.. and then stick the rcfs as columns of a matrix
// then call viterbi decoding with weight vector and transition matrix
// and get best path
unsigned int wv_len = 128;
double rayparam = 43.;
// make rayleigh weighting curve
d_vec_t wv(wv_len);
for (unsigned int i=0; i<wv.size(); i++)
{
wv[i] = (static_cast<double> (i) / pow(rayparam,2.)) * exp((-1.*pow(-static_cast<double> (i),2.)) / (2.*pow(rayparam,2.)));
}
// beat tracking frame size (roughly 6 seconds) and hop (1.5 seconds)
unsigned int winlen = 512;
unsigned int step = 128;
// matrix to store output of comb filter bank, increment column of matrix at each frame
d_mat_t rcfmat;
int col_counter = -1;
// main loop for beat period calculation
for (unsigned int i=0; i+winlen<df.size(); i+=step)
{
// get dfframe
d_vec_t dfframe(winlen);
for (unsigned int k=0; k<winlen; k++)
{
dfframe[k] = df[i+k];
}
// get rcf vector for current frame
d_vec_t rcf(wv_len);
get_rcf(dfframe,wv,rcf);
rcfmat.push_back( d_vec_t() ); // adds a new column
col_counter++;
for (unsigned int j=0; j<rcf.size(); j++)
{
rcfmat[col_counter].push_back( rcf[j] );
}
}
// now call viterbi decoding function
viterbi_decode(rcfmat,wv,beat_period,tempi);
}
// Open Media file and return elapsed time in millseconds
int Waveform::OpenfileMedia(AudioManager* media, wxString& error)
{
_media = media;
views.clear();
if (_media != NULL)
{
float samplesPerLine = GetSamplesPerLineFromZoomLevel(mZoomLevel);
WaveView wv(mZoomLevel, samplesPerLine, media);
views.push_back(wv);
mCurrentWaveView = 0;
return media->LengthMS();
}
else
{
return 0;
}
}
void Waveform::SetZoomLevel(int level)
{
mZoomLevel = level;
if(!mIsInitialized){return;}
mCurrentWaveView = NO_WAVE_VIEW_SELECTED;
for(int i=0;i<views.size();i++)
{
if(views[i].GetZoomLevel() == mZoomLevel)
{
mCurrentWaveView = i;
}
}
if (mCurrentWaveView == NO_WAVE_VIEW_SELECTED)
{
float samplesPerLine = GetSamplesPerLineFromZoomLevel(mZoomLevel);
WaveView wv(mZoomLevel,samplesPerLine,m_left_data,mMediaTrackSize);
views.push_back(wv);
mCurrentWaveView = views.size()-1;
}
}
// Open Media file and return elapsed time in millseconds
int Waveform::OpenfileMediaFile(const wxString &filename, wxString& error)
{
mpg123_handle *mh = NULL;
int err;
size_t buffer_size;
int channels, encoding;
long rate;
err = mpg123_init();
if(err != MPG123_OK || (mh = mpg123_new(NULL, &err)) == NULL)
{
error = wxString::Format("Basic setup goes wrong: %s", mpg123_plain_strerror(err));
if (mh != NULL) {
cleanup(mh);
}
return -1;
}
/* open the file and get the decoding format */
if( mpg123_open(mh, filename.c_str()) != MPG123_OK ||
mpg123_getformat(mh, &rate, &channels, &encoding) != MPG123_OK )
{
error = wxString::Format("Trouble with mpg123: %s", mpg123_strerror(mh));
cleanup(mh);
return -1;
}
if( encoding != MPG123_ENC_SIGNED_16 )
{
error = "Encoding unsupported. Must be signed 16 bit.";
cleanup(mh);
return -2;
}
/* set the output format and open the output device */
m_bits = mpg123_encsize(encoding);
m_rate = rate;
m_channels = channels;
/* Get Track Size */
mMediaTrackSize = GetTrackSize(mh,m_bits,m_channels);
buffer_size = mpg123_outblock(mh);
int size = (mMediaTrackSize+buffer_size)*m_bits*m_channels;
char * trackData = (char*)malloc(size);
LoadTrackData(mh,trackData, size);
// Split data into left and right and normalize -1 to 1
m_left_data = (float*)malloc(sizeof(float)*mMediaTrackSize);
if( m_channels == 2 )
{
m_right_data = (float*)malloc(sizeof(float)*mMediaTrackSize);
SplitTrackDataAndNormalize((signed short*)trackData,mMediaTrackSize,m_left_data,m_right_data);
}
else if( m_channels == 1 )
{
NormalizeMonoTrackData((signed short*)trackData,mMediaTrackSize,m_left_data);
}
else
{
error = "More than 2 audio channels is not supported yet.";
}
views.clear();
float samplesPerLine = GetSamplesPerLineFromZoomLevel(mZoomLevel);
WaveView wv(mZoomLevel,samplesPerLine,m_left_data,mMediaTrackSize);
views.push_back(wv);
mCurrentWaveView = 0;
mpg123_close(mh);
mpg123_delete(mh);
mpg123_exit();
free(trackData);
float seconds = (float)mMediaTrackSize * ((float)1/(float)rate);
mAudioIsLoaded = true;
return (int)(seconds * (float)1000);
}
arma::field<arma::field<arma::mat> > model_select(arma::vec& data,
const std::set<std::vector<std::string > >& models,
const std::vector< std::string >& full_model,
std::string model_type,
bool bs_optimism,
double alpha,
std::string compute_v,
unsigned int K, unsigned int H, unsigned int G,
bool robust, double eff, unsigned int seed){
// Number of data points
unsigned int N = data.n_rows;
// Number of models
unsigned int num_models = models.size();
// Store output from models
arma::field<arma::field<arma::mat> > model_results(num_models);
// Make an iterator to iterator through it
std::set<std::vector<std::string > > ::const_iterator iter;
iter = models.begin();
// Get the first model
std::vector<std::string> desc = full_model;
// Find where the results should be input. (No protection needed, we know it is in the matrix)
unsigned int full_model_index = std::distance(models.begin(), models.find(full_model));
// Build the fields off of the first model's description
arma::vec theta = model_theta(desc);
arma::field<arma::vec> objdesc = model_objdesc(desc);
// Build matrix to store results
arma::mat results(num_models, 4);
Rcpp::Rcout << "Processing model 1 out of " << num_models << std::endl;
set_seed(seed);
// Obtain the largest models information
arma::field<arma::mat> master = gmwm_master_cpp(data,
theta,
desc,
objdesc,
model_type,
true, //starting
alpha,
"fast", // compute V
K, H,
G,
robust, eff);
// Theta update
theta = master(0);
// Define WV Empirical
arma::vec wv_empir = master(2);
// Create WV Matrix
arma::mat wv(wv_empir.n_elem,3);
wv.col(0) = wv_empir;
wv.col(1) = master(3);
wv.col(2) = master(4);
// Get the original "FAST" matrix
arma::mat orgV = master(6); // Original V
// Take the inverse
arma::mat omega = inv(orgV); // Original V => Omega
// Get expect_diff for guesses with DR
double expect_diff = arma::as_scalar(master(7));
// Obtain the theoretical WV
arma::vec theo = master(8);
// Obtain the obj_value of the function
double obj_value = arma::as_scalar(master(10));
// Calculate the values of the Scales
arma::vec scales = scales_cpp(floor(log2(N)));
double dr_slope = arma::as_scalar(master(12));
// ------------------------------------
// Store output from default GMWM object
arma::field<arma::mat> mod_output(13);
mod_output(0) = theta;
mod_output(1) = master(1);
mod_output(2) = wv_empir;
mod_output(3) = master(3);
mod_output(4) = master(4);
mod_output(5) = master(5);
mod_output(6) = orgV;
mod_output(7) = expect_diff;
mod_output(8) = theo;
mod_output(9) = master(9);
mod_output(10) = obj_value;
mod_output(11) = omega;
mod_output(12) = dr_slope;
// ------------------------------------
// Here we set up specifics that are used not in a specific mode.
// Asymptotic ----
// Get bootstrapped V
arma::mat V;
// Bootstrap ----
// Hold optimism result
arma::mat cov_nu_nu_theta;
// Hold the bootstrapped obj values
arma::vec bs_obj_values;
if(bs_optimism){
results.row(full_model_index) = bs_optim_calc(theta, desc, objdesc, model_type, scales, omega, N,
obj_value, alpha, compute_v, K, H, G, robust, eff);
}else{
Rcpp::Rcout << "Bootstrapping the covariance matrix... Please stand by." << std::endl;
V = cov_bootstrapper(theta, desc, objdesc, N, robust, eff, H, false); // Bootstrapped V (largest model)
// Calculate the model score according to model selection criteria paper
results.row(full_model_index) = asympt_calc(theta, desc, objdesc, model_type, scales, V, omega, wv_empir, theo, obj_value);
}
// Custom GMWM Update Obj
arma::field<arma::mat> model_update_info(6);
model_update_info(0) = master(0); // theta
model_update_info(1) = master(1); // guessed_theta
model_update_info(2) = master(5); // V
model_update_info(3) = master(8); // theo
model_update_info(4) = master(9); // decomp_theo
model_update_info(5) = master(10); // objective value
model_results(full_model_index) = model_update_info;
// Initialize counter to keep track of values
unsigned int count = 0;
unsigned int countModels = 1;
while(iter != models.end()){
if(full_model_index != count){
countModels++;
// Set guessing seed
set_seed(seed);
Rcpp::Rcout << "Processing model " << countModels << " out of " << num_models << std::endl;
// Get the first model
desc = *iter;
// Build the fields off of the first model's description
theta = model_theta(desc);
objdesc = model_objdesc(desc);
// Run the update version of the GMWM
arma::field<arma::mat> update = gmwm_update_cpp(theta, desc, objdesc, model_type,
N, expect_diff, dr_slope,
orgV, scales, wv,
true, //starting
"fast",
K,H,G,
robust,eff);
// Theta update
theta = update(0);
// Update theo
theo = update(3);
// Update objective function
obj_value = arma::as_scalar(update(5));
if(bs_optimism){
results.row(count) = bs_optim_calc(theta, desc, objdesc, model_type, scales, omega, N,
obj_value, alpha, compute_v, K, H, G, robust, eff);
}else{
// Calculate the model score according to model selection criteria paper
results.row(count) = asympt_calc(theta, desc, objdesc, model_type, scales, V, omega, wv_empir, theo, obj_value);
}
model_results(count) = update;
}
// end if
// Increment iterator
iter++;
// Increase count
count++;
}
// Only run if in asymptotic mode
if(!bs_optimism){
// std::set<std::vector<std::string > > ::const_iterator iter2;
//
// iter = models.begin();
//
// /*
// * Iterate through all models
// * If i != j, then
// * IF (Complex_i > Complex_j && Obj_i > Obj_j){
// * IF(Crit_i < Crit_J)
// * }
// */
// while(iter != models.end()){
// iter2 = models.begin();
// while(iter2 != models.end()){
// if(iter != iter2){
// double m1_index = std::distance(models.begin(),iter);
// double m2_index = std::distance(models.begin(),iter2);
//
// if(count_params(*iter) > count_params(*iter2) && results(m1_index,0) > results(m2_index,0)){
// if(results(m1_index,2) < results(m2_index,2)){
// results(m1_index,1) = results(m2_index,1) + fabs(R::rnorm(0, sqrt(results(m2_index,0)/10) ));
// }
// }
// }
// iter2++;
// }
// iter++;
// }
}
results.col(2) = results.col(0) + results.col(1);
// Sort matrix
arma::uvec sort_ind = sort_index(results.col(2));
// Pick the "best" model
unsigned int best_model_id = sort_ind(0);
// Get the sort column index
arma::vec ex_sort = arma::conv_to<arma::vec>::from(sort_ind);
// Sort the result matrix by Criterion
arma::mat sorted = sort_mat(results, 2);
// ---------
// Set up output feed
arma::field< arma::field<arma::mat> > out(2);
// Build the results matrix
arma::field<arma::mat> ms(2);
ms(0) = sorted;
ms(1) = ex_sort + 1; // Sorted vector (so R can know desc IDs) ALSO change index
// Create m
// If the output object is not the full model, let's inject the new results.
if(best_model_id != full_model_index){
arma::field<arma::mat> m;
m = model_results(best_model_id);
mod_output(0) = m(0);
mod_output(1) = m(1);
mod_output(5) = m(2);
mod_output(8) = m(3);
mod_output(9) = m(4);
mod_output(10) = m(5);
}
// Output to me!
out(0) = ms;
out(1) = mod_output;
return out;
}
