void add_state_transition(TrainingData& td, char from, char to)
{
int f_idx = statechar2index(from);
int t_idx = statechar2index(to);
int count = get(td.state_transitions, f_idx, t_idx);
set(td.state_transitions, f_idx, t_idx, count + 1);
}
void TransitionParameters::add_transition_observation(char state_from, char state_to)
{
int f_idx = statechar2index(state_from);
int t_idx = statechar2index(state_to);
int count = get(training_data.state_transitions, f_idx, t_idx);
set(training_data.state_transitions, f_idx, t_idx, count + 1);
}
void KHMMParameters::train()
{
TrainingData& td = training_data;
//
// Profile HMM transitions
//
fprintf(stderr, "TRANSITIONS\n");
size_t sum_m_not_k = get(td.state_transitions, statechar2index('M'), statechar2index('M')) +
get(td.state_transitions, statechar2index('M'), statechar2index('E'));
size_t me = get(td.state_transitions, statechar2index('M'), statechar2index('E'));
double p_me_not_k = (double)me / sum_m_not_k;
fprintf(stderr, "M->E|not_k: %lf\n", p_me_not_k);
size_t sum_e = 0;
for(int j = 0; j < td.state_transitions.n_cols; ++j) {
sum_e += get(td.state_transitions, statechar2index('E'), j);
}
size_t ee = get(td.state_transitions, statechar2index('E'), statechar2index('E'));
double p_ee = (double)ee / sum_e;
fprintf(stderr, "E->E: %lf\n", p_ee);
for(int i = 0; i < td.state_transitions.n_rows; ++i) {
fprintf(stderr, "\t%c: ", "MEK"[i]);
for(int j = 0; j < td.state_transitions.n_cols; ++j) {
fprintf(stderr, "%d ", get(td.state_transitions, i, j));
}
fprintf(stderr, "\n");
}
if(sum_e == 0 || sum_m_not_k == 0) {
// insufficient data to train, use defaults
return;
}
trans_m_to_e_not_k = p_me_not_k;
trans_e_to_e = p_ee;
//
// Signal-dependent skip probability
//
// Initialize observations with pseudocounts from the current model
size_t num_bins = skip_probabilities.size();
uint32_t pseudocount = 100;
std::vector<double> total_observations(num_bins, 0.0f);
std::vector<double> skip_observations(num_bins, 0.0f);
for(size_t bin = 0; bin < num_bins; bin++) {
skip_observations[bin] = skip_probabilities[bin] * pseudocount;
total_observations[bin] = pseudocount;
}
for(size_t oi = 0; oi < td.kmer_transitions.size(); ++oi) {
const KmerTransitionObservation& to = td.kmer_transitions[oi];
bool is_skip = to.state == 'K';
size_t bin = get_bin(*this, to.level_1, to.level_2);
skip_observations[bin] += is_skip;
total_observations[bin] += 1;
}
// Update probabilities
for(size_t bin = 0; bin < num_bins; bin++) {
skip_probabilities[bin] = skip_observations[bin] / total_observations[bin];
fprintf(stderr, "SKIPLEARN -- bin[%zu] %.3lf %.3lf %.3lf\n", bin, skip_observations[bin], total_observations[bin], skip_probabilities[bin]);
}
}
