int main(void) {
DBN mdbn;
mdbn.load("mixed_hmm2.dbn");
std::vector<Node*> nodes = mdbn.getNodes0();
Node *h1 = nodes[0];
Node *o1 = nodes[1];
cout << "*** LOADED MODEL WITH PARAMETERS ***" << endl;
cout << "h1: " << *h1 << endl;
cout << "o1: " << *o1 << endl;
return EXIT_SUCCESS;
}
int main(void) {
mocapy_seed((uint)5556575);
// Number of trainining sequences
int N = 500;
// Sequence lengths
int T = 100;
// Gibbs sampling parameters
int MCMC_BURN_IN = 10;
// HMM hidden and observed node sizes
uint H_SIZE=2;
bool init_random=false;
CPD th0_cpd;
th0_cpd.set_shape(2); th0_cpd.set_values(vec(0.1, 0.9));
CPD th1_cpd;
th1_cpd.set_shape(2,2); th1_cpd.set_values( vec(0.95, 0.05, 0.1, 0.9));
// The target DBN (This DBN generates the data)
Node* th0 = NodeFactory::new_discrete_node(H_SIZE, "th0", init_random, th0_cpd);
Node* th1 = NodeFactory::new_discrete_node(H_SIZE, "th1", init_random, th1_cpd);
Node* to0 = NodeFactory::new_kent_node("to0");
DBN tdbn;
tdbn.set_slices(vec(th0, to0), vec(th1, to0));
tdbn.add_intra("th0", "to0");
tdbn.add_inter("th0", "th1");
tdbn.construct();
// The model DBN (this DBN will be trained)
// For mh0, get the CPD from th0 and fix parameters
Node* mh0 = NodeFactory::new_discrete_node(H_SIZE, "mh0", init_random, CPD(), th0, true );
Node* mh1 = NodeFactory::new_discrete_node(H_SIZE, "mh1", init_random);
Node* mo0 = NodeFactory::new_kent_node("mo0");
DBN mdbn;
mdbn.set_slices(vec(mh0, mo0), vec(mh1, mo0));
mdbn.add_intra("mh0", "mo0");
mdbn.add_inter("mh0", "mh1");
mdbn.construct();
cout << "*** TARGET ***" << endl;
cout << *th0 << endl;
cout << *th1 << endl;
cout << *to0 << endl;
cout << "*** MODEL ***" << endl;
cout << *mh0 << endl;
cout << *mh1 << endl;
cout << *mo0 << endl;
vector<Sequence> seq_list;
vector< MDArray<eMISMASK> > mismask_list;
cout << "Generating data" << endl;
MDArray<eMISMASK> mismask;
mismask.repeat(T, vec(MOCAPY_HIDDEN, MOCAPY_OBSERVED));
// Generate the data
double sum_LL(0);
for (int i=0; i<N; i++) {
pair<Sequence, double> seq_ll = tdbn.sample_sequence(T);
sum_LL += seq_ll.second;
seq_list.push_back(seq_ll.first);
mismask_list.push_back(mismask);
}
cout << "Average LL: " << sum_LL/N << endl;
GibbsRandom mcmc = GibbsRandom(&mdbn);
EMEngine em = EMEngine(&mdbn, &mcmc, &seq_list, &mismask_list);
cout << "Starting EM loop" << endl;
double bestLL=-1000;
uint it_no_improvement(0);
uint i(0);
// Start EM loop
while (it_no_improvement<10) {
em.do_E_step(1, MCMC_BURN_IN, true);
double ll = em.get_loglik();
cout << "LL= " << ll;
if (ll > bestLL) {
cout << " * saving model *" << endl;
mdbn.save("kent_hmm.dbn");
bestLL = ll;
it_no_improvement=0;
}
else { it_no_improvement++; cout << endl; }
i++;
em.do_M_step();
}
cout << "DONE" << endl;
mdbn.load("kent_hmm.dbn");
cout << "*** TARGET ***" << endl;
cout << *th0 << endl;
cout << *th1 << endl;
cout << *to0 << endl;
cout << "*** MODEL ***" << endl;
cout << *mh0 << endl;
cout << *mh1 << endl;
cout << *mo0 << endl;
delete th0;
delete th1;
delete to0;
delete mh0;
delete mh1;
delete mo0;
return EXIT_SUCCESS;
}
void execute(DBN& dbn, task& task, const std::vector<std::string>& actions) {
print_title("Network");
dbn.display();
using dbn_t = std::decay_t<DBN>;
//Execute all the actions sequentially
for (auto& action : actions) {
if (action == "pretrain") {
print_title("Pretraining");
if (task.pretraining.samples.empty()) {
std::cout << "dllp: error: pretrain is not possible without a pretraining input" << std::endl;
return;
}
std::vector<Container> pt_samples;
//Try to read the samples
if (!read_samples<Three>(task.pretraining.samples, pt_samples)) {
std::cout << "dllp: error: failed to read the pretraining samples" << std::endl;
return;
}
if (task.pt_desc.denoising) {
std::vector<Container> clean_samples;
//Try to read the samples
if (!read_samples<Three>(task.pretraining_clean.samples, clean_samples)) {
std::cout << "dllp: error: failed to read the clean samples" << std::endl;
return;
}
//Pretrain the network
cpp::static_if<dbn_t::layers_t::is_denoising>([&](auto f) {
f(dbn).pretrain_denoising(pt_samples.begin(), pt_samples.end(), clean_samples.begin(), clean_samples.end(), task.pt_desc.epochs);
});
} else {
//Pretrain the network
dbn.pretrain(pt_samples.begin(), pt_samples.end(), task.pt_desc.epochs);
}
} else if (action == "train") {
print_title("Training");
if (task.training.samples.empty() || task.training.labels.empty()) {
std::cout << "dllp: error: train is not possible without samples and labels" << std::endl;
return;
}
std::vector<Container> ft_samples;
std::vector<std::size_t> ft_labels;
//Try to read the samples
if (!read_samples<Three>(task.training.samples, ft_samples)) {
std::cout << "dllp: error: failed to read the training samples" << std::endl;
return;
}
//Try to read the labels
if (!read_labels(task.training.labels, ft_labels)) {
std::cout << "dllp: error: failed to read the training labels" << std::endl;
return;
}
using last_layer = typename dbn_t::template layer_type<dbn_t::layers - 1>;
//Train the network
cpp::static_if<sgd_possible<last_layer>::value>([&](auto f) {
auto ft_error = f(dbn).fine_tune(ft_samples, ft_labels, task.ft_desc.epochs);
std::cout << "Train Classification Error:" << ft_error << std::endl;
});
} else if (action == "test") {
print_title("Testing");
if (task.testing.samples.empty() || task.testing.labels.empty()) {
std::cout << "dllp: error: test is not possible without samples and labels" << std::endl;
return;
}
std::vector<Container> test_samples;
std::vector<std::size_t> test_labels;
//Try to read the samples
if (!read_samples<Three>(task.testing.samples, test_samples)) {
std::cout << "dllp: error: failed to read the test samples" << std::endl;
return;
}
//Try to read the labels
if (!read_labels(task.testing.labels, test_labels)) {
std::cout << "dllp: error: failed to read the test labels" << std::endl;
return;
}
auto classes = dbn_t::output_size();
etl::dyn_matrix<std::size_t, 2> conf(classes, classes, 0.0);
std::size_t n = test_samples.size();
std::size_t tp = 0;
for (std::size_t i = 0; i < test_samples.size(); ++i) {
auto sample = test_samples[i];
auto label = test_labels[i];
auto predicted = dbn.predict(sample);
if (predicted == label) {
++tp;
}
++conf(label, predicted);
}
double test_error = (n - tp) / double(n);
std::cout << "Error rate: " << test_error << std::endl;
std::cout << "Accuracy: " << (1.0 - test_error) << std::endl
<< std::endl;
std::cout << "Results per class" << std::endl;
double overall = 0.0;
std::cout << " | Accuracy | Error rate |" << std::endl;
for (std::size_t l = 0; l < classes; ++l) {
std::size_t total = etl::sum(conf(l));
double acc = (total - conf(l, l)) / double(total);
std::cout << std::setw(3) << l;
std::cout << "|" << std::setw(10) << (1.0 - acc) << "|" << std::setw(12) << acc << "|" << std::endl;
overall += acc;
}
std::cout << std::endl;
std::cout << "Overall Error rate: " << overall / classes << std::endl;
std::cout << "Overall Accuracy: " << 1.0 - (overall / classes) << std::endl
<< std::endl;
std::cout << "Confusion Matrix (%)" << std::endl
<< std::endl;
std::cout << " ";
for (std::size_t l = 0; l < classes; ++l) {
std::cout << std::setw(5) << l << " ";
}
std::cout << std::endl;
for (std::size_t l = 0; l < classes; ++l) {
std::size_t total = etl::sum(conf(l));
std::cout << std::setw(3) << l << "|";
for (std::size_t p = 0; p < classes; ++p) {
std::cout << std::setw(5) << std::setprecision(2) << 100.0 * (conf(l, p) / double(total)) << "|";
}
std::cout << std::endl;
}
std::cout << std::endl;
} else if (action == "save") {
print_title("Save Weights");
dbn.store(task.w_desc.file);
std::cout << "Weights saved" << std::endl;
} else if (action == "load") {
print_title("Load Weights");
dbn.load(task.w_desc.file);
std::cout << "Weights loaded" << std::endl;
} else {
std::cout << "dllp: error: Invalid action: " << action << std::endl;
}
}
//TODO
}
