void softmax<T>::train(const Eigen::Ref<const EigenMat> &train,
const std::vector<int> &labels)
{
#ifdef OCV_TEST_SOFTMAX
gradient_check();
#endif
auto const UniqueLabels = get_unique_labels(labels);
auto const NumClass = UniqueLabels.size();
weight_ = EigenMat::Random(NumClass, train.rows());
grad_ = EigenMat::Zero(NumClass, train.rows());
auto const TrainCols = static_cast<int>(train.cols());
EigenMat const GroundTruth = get_ground_truth(static_cast<int>(NumClass),
TrainCols,
UniqueLabels,
labels);
std::random_device rd;
std::default_random_engine re(rd());
int const Batch = (get_batch_size(TrainCols));
int const RandomSize = TrainCols != Batch ?
TrainCols - Batch - 1 : 0;
std::uniform_int_distribution<int>
uni_int(0, RandomSize);
for(size_t i = 0; i != params_.max_iter_; ++i){
auto const Cols = uni_int(re);
auto const &TrainBlock =
train.block(0, Cols, train.rows(), Batch);
auto const >Block =
GroundTruth.block(0, Cols, NumClass, Batch);
auto const Cost = compute_cost(TrainBlock, weight_, GTBlock);
if(std::abs(params_.cost_ - Cost) < params_.epsillon_ ||
Cost < 0){
break;
}
params_.cost_ = Cost;
compute_gradient(TrainBlock, weight_, GTBlock);
weight_.array() -= grad_.array() * params_.lrate_;//*/
}
}
void train(Eigen::MatrixBase<Derived> const &input)
{
if(!reuse_layer_){
layers_.clear();
}
std::random_device rd;
std::default_random_engine re(rd());
int const Batch = get_batch_size(input.cols());
int const RandomSize = input.cols() != Batch ?
input.cols() - Batch - 1 : 0;
std::uniform_int_distribution<int>
uni_int(0, RandomSize);
#ifdef OCV_TEST_AUTOENCODER
gradient_check();
#endif
for(size_t i = 0; i < params_.hidden_size_.size(); ++i){
Eigen::MatrixBase<Derived> const &TmpInput =
i == 0 ? input
: eactivation_;
if(!reuse_layer_){
layer es(TmpInput.rows(), params_.hidden_size_[i]);
reduce_cost(uni_int, re, Batch, TmpInput, es);
generate_activation(es, TmpInput,
i==0?true:false);
layers_.push_back(es);
}else{
if(layers_.size() <= i){
layers_.emplace_back(static_cast<int>(TmpInput.rows()),
params_.hidden_size_[i]);
}
reduce_cost(uni_int, re, Batch, TmpInput, layers_[i]);
generate_activation(layers_[i], TmpInput,
i==0?true:false);
}
}
act_.clear();
buffer_.clear();//*/
}
void clustered_ba::connect(const vertices_size_type &idx)
{
pagmo_assert(get_number_of_vertices() > 0);
const vertices_size_type prev_size = get_number_of_vertices() - 1;
if (prev_size < m_m0) {
// If we had not built the initial m0 nodes, do it.
// We want to connect the newcomer island with high probability, and make sure that
// at least one connection exists (otherwise the island stays isolated).
// NOTE: is it worth to make it a user-tunable parameter?
const double prob = 0.0;
// Flag indicating if at least 1 connection was added.
bool connection_added = false;
// Main loop.
for (std::pair<v_iterator,v_iterator> vertices = get_vertices(); vertices.first != vertices.second; ++vertices.first) {
// Do not consider the new vertex itself.
if (*vertices.first != idx) {
if (m_drng() < prob) {
connection_added = true;
// Add the connections
add_edge(*vertices.first,idx);
add_edge(idx,*vertices.first);
}
}
}
// If no connections were established and this is not the first island being inserted,
// establish at least one connection with a random island other than n.
if ((!connection_added) && (prev_size != 0)) {
// Get a random vertex index between 0 and n_vertices - 1. Keep on repeating the procedure if by
// chance we end up on idx again.
boost::uniform_int<vertices_size_type> uni_int(0,get_number_of_vertices() - 1);
vertices_size_type rnd;
do {
rnd = uni_int(m_urng);
} while (rnd == idx);
// Add connections to the random vertex.
add_edge(rnd,idx);
add_edge(idx,rnd);
}
} else {
// Now we need to add j edges, choosing the nodes with a probability
// proportional to their number of connections. We keep track of the
// connection established in order to avoid connecting twice to the same
// node.
// j is a random integer in the range 1 to m.
boost::uniform_int<edges_size_type> uni_int2(1,m_m);
std::size_t i = 0;
std::size_t j = uni_int2(m_urng);
std::pair<v_iterator,v_iterator> vertices;
std::pair<a_iterator,a_iterator> adj_vertices;
while (i < j) {
// Let's find the current total number of edges.
const edges_size_type n_edges = get_number_of_edges();
pagmo_assert(n_edges > 0);
boost::uniform_int<edges_size_type> uni_int(0,n_edges - 1 - i);
// Here we choose a random number between 0 and n_edges - 1 - i.
const edges_size_type rn = uni_int(m_urng);
edges_size_type n = 0;
// Iterate over all vertices and accumulate the number of edges for each of them. Stop when the accumulated number of edges is greater
// than rn. This is equivalent to giving a chance of connection to vertex v directly proportional to the number of edges departing from v.
// You can think of this process as selecting a random edge among all the existing edges and connecting to the vertex from which the
// selected edge departs.
vertices = get_vertices();
for (; vertices.first != vertices.second; ++vertices.first) {
// Do not consider it_n.
if (*vertices.first != idx) {
adj_vertices = get_adjacent_vertices(*vertices.first);
n += boost::numeric_cast<edges_size_type>(std::distance(adj_vertices.first,adj_vertices.second));
if (n > rn) {
break;
}
}
}
pagmo_assert(vertices.first != vertices.second);
// If the candidate was not already connected, then add it.
if (!are_adjacent(idx,*vertices.first)) {
// Connect to nodes that are already adjacent to idx with probability p.
// This step increases clustering in the network.
adj_vertices = get_adjacent_vertices(idx);
for(;adj_vertices.first != adj_vertices.second; ++adj_vertices.first) {
if(m_drng() < m_p && *adj_vertices.first != *vertices.first && !are_adjacent(*adj_vertices.first,*vertices.first)) {
add_edge(*adj_vertices.first, *vertices.first);
add_edge(*vertices.first, *adj_vertices.first);
}
}
// Connect to idx
add_edge(*vertices.first,idx);
add_edge(idx,*vertices.first);
++i;
}
}
}
}
void ihs::evolve(population &pop) const
{
// Let's store some useful variables.
const problem::base &prob = pop.problem();
const problem::base::size_type prob_dimension = prob.get_dimension(), prob_i_dimension = prob.get_i_dimension();
const decision_vector &lb = prob.get_lb(), &ub = prob.get_ub();
const population::size_type pop_size = pop.size();
// Get out if there is nothing to do.
if (pop_size == 0 || m_gen == 0) {
return;
}
decision_vector lu_diff(prob_dimension);
for (problem::base::size_type i = 0; i < prob_dimension; ++i) {
lu_diff[i] = ub[i] - lb[i];
}
// Int distribution to be used when picking random individuals.
boost::uniform_int<population::size_type> uni_int(0,pop_size - 1);
const double c = std::log(m_bw_min/m_bw_max) / m_gen;
// Temporary individual used during evolution.
population::individual_type tmp;
tmp.cur_x.resize(prob_dimension);
tmp.cur_f.resize(prob.get_f_dimension());
tmp.cur_c.resize(prob.get_c_dimension());
for (std::size_t g = 0; g < m_gen; ++g) {
const double ppar_cur = m_ppar_min + ((m_ppar_max - m_ppar_min) * g) / m_gen, bw_cur = m_bw_max * std::exp(c * g);
// Continuous part.
for (problem::base::size_type i = 0; i < prob_dimension - prob_i_dimension; ++i) {
if (m_drng() < m_phmcr) {
// tmp's i-th chromosome element is the one from a randomly chosen individual.
tmp.cur_x[i] = pop.get_individual(uni_int(m_urng)).cur_x[i];
// Do pitch adjustment with ppar_cur probability.
if (m_drng() < ppar_cur) {
// Randomly, add or subtract pitch from the current chromosome element.
if (m_drng() > .5) {
tmp.cur_x[i] += m_drng() * bw_cur * lu_diff[i];
} else {
tmp.cur_x[i] -= m_drng() * bw_cur * lu_diff[i];
}
// Handle the case in which we added or subtracted too much and ended up out
// of boundaries.
if (tmp.cur_x[i] > ub[i]) {
tmp.cur_x[i] = boost::uniform_real<double>(lb[i],ub[i])(m_drng);
} else if (tmp.cur_x[i] < lb[i]) {
tmp.cur_x[i] = boost::uniform_real<double>(lb[i],ub[i])(m_drng);
}
}
} else {
// Pick randomly within the bounds.
tmp.cur_x[i] = boost::uniform_real<double>(lb[i],ub[i])(m_drng);
}
}
//Integer Part
for (problem::base::size_type i = prob_dimension - prob_i_dimension; i < prob_dimension; ++i) {
if (m_drng() < m_phmcr) {
tmp.cur_x[i] = pop.get_individual(uni_int(m_urng)).cur_x[i];
if (m_drng() < ppar_cur) {
if (m_drng() > .5) {
tmp.cur_x[i] += double_to_int::convert(m_drng() * bw_cur * lu_diff[i]);
} else {
tmp.cur_x[i] -= double_to_int::convert(m_drng() * bw_cur * lu_diff[i]);
}
// Wrap over in case we went past the bounds.
if (tmp.cur_x[i] > ub[i]) {
tmp.cur_x[i] = lb[i] + double_to_int::convert(tmp.cur_x[i] - ub[i]) % static_cast<int>(lu_diff[i]);
} else if (tmp.cur_x[i] < lb[i]) {
tmp.cur_x[i] = ub[i] - double_to_int::convert(lb[i] - tmp.cur_x[i]) % static_cast<int>(lu_diff[i]);
}
}
} else {
// Pick randomly within the bounds.
tmp.cur_x[i] = boost::uniform_int<int>(lb[i],ub[i])(m_urng);
}
}
// And we push him back
pop.push_back(tmp.cur_x);
// We locate the worst individual.
const population::size_type worst_idx = pop.get_worst_idx();
// And we get rid of him :)
pop.erase(worst_idx);
}
}
