// run the whole learning process of RBM
void rbm::run(std::string model_weight_file, std::string model_visible_bias_file,
std::string model_hidden_bias_file, std::string learned_features_file, std::string file){
load_data(file.c_str());
// randomly initialize weights and biases
float ** weights = new float*[num_hidden_units];
float * hidden_bias = new float[num_hidden_units];
float * visible_bias = new float[num_visible_units];
for (int i=0; i<num_hidden_units; i++){
weights[i] = new float[num_visible_units];
}
init_paras(weights, visible_bias, hidden_bias);
// std::string in_weights = "init_weights";
// std::string in_hidden = "init_hidden";
// std::string in_visible = "init_visible";
// init_paras(weights, visible_bias, hidden_bias, in_weights.c_str(), in_hidden.c_str(), in_visible.c_str());
// train rbm
train(weights, visible_bias, hidden_bias);
// save model
save_model(weights, visible_bias, hidden_bias, model_weight_file.c_str(),
model_visible_bias_file.c_str(), model_hidden_bias_file.c_str());
for (int i=0; i<num_hidden_units; i++){
delete[] weights[i];
}
delete[] weights;
delete[] visible_bias;
delete[] hidden_bias;
}
// run the whole learning process of DNN
void dnn::run(std::string model_weight_file, std::string model_bias_file)
{
// assign id to threads
if (!thread_id.get()) {
thread_id.reset(new int(thread_counter++));
}
// get access to tables
petuum::PSTableGroup::RegisterThread();
mat *weights= new mat[num_layers-1];
mat *biases=new mat[num_layers-1];
for(int i=0;i<num_layers-1;i++){
weights[i]=petuum::PSTableGroup::GetTableOrDie<float>(i);
biases[i]=petuum::PSTableGroup::GetTableOrDie<float>(i+num_layers-1);
}
// Run additional iterations to let stale values finish propagating
for (int iter = 0; iter < staleness; ++iter) {
petuum::PSTableGroup::Clock();
}
// initialize parameters
if (client_id==0&&(*thread_id) == 0){
std::cout<<"init parameters"<<std::endl;
init_paras(weights, biases);
std::cout<<"init parameters done"<<std::endl;
}
process_barrier->wait();
// do DNN training
if (client_id==0&&(*thread_id) == 0)
std::cout<<"training starts"<<std::endl;
train(weights, biases);
// Run additional iterations to let stale values finish propagating
for (int iter = 0; iter < staleness; ++iter) {
petuum::PSTableGroup::Clock();
}
//save model
if(client_id==0&&(*thread_id)==0)
{
save_model(weights, biases, model_weight_file.c_str(), model_bias_file.c_str());
}
delete[]weights;
delete[]biases;
petuum::PSTableGroup::DeregisterThread();
}
virtual void solve() {
auto lines = paracel_load(input);
local_parser(item_vects, lines);
std::cout << "parser done" << std::endl;
if(learning_method == "default") {
auto all_lines = paracel_loadall(input);
local_parser(all_item_vects, all_lines);
std::cout << "loadall done" << std::endl;
normalize(item_vects);
normalize(all_item_vects);
std::cout << "normalize done" << std::endl;
sync();
learning();
} else if(learning_method == "limit_storage") {
normalize(item_vects); // normalize here to reduce calculation
init_paras();
sync();
mls_learning();
} else {}
}
void solve() {
init_paras();
paracel_sync();
learning();
}
