int main(int argc, char** argv){
double learning_rate;
int batch_size;
vector<int> layers;
string train_fname;
int max_epoch;
float valid_ratio=0.1;
string output_model;
string structure;
Net d;
Net bestModel;
//Set parameters
float early_stop_thres = 0.03;
//ex: ./run 0.01 5-4-3 300 train_file_name output_model_name model_name
if(argc < 7){
printf("Usage:\n");
printf("./train learning_rate(0.01) batch_size(10) structure(5-4-3) max_epoch(100) train_file output_model [load_model]\n");
return 0;
}else{
learning_rate = atof(argv[1]);
batch_size = atoi(argv[2]);
string lyr(argv[3]);
structure = lyr.c_str();
vector<string> x = split(lyr,"-");
for(int i=0;i<x.size();i++){
layers.push_back(atoi(x[i].c_str()));
}
max_epoch = atoi(argv[4]);
train_fname.assign(argv[5]);
output_model.assign(argv[6]);
}
//Initialize neural network
srand(time(NULL));
if(argc == 7){
d.load_model(layers);
}else if(argc == 8){
string m_name(argv[7]);
d.load_model(m_name);
}else{
printf("wrong parameters\n");
return 0;
}
d.learning_rate = learning_rate;
d.batch_size = batch_size;
//Loading data
puts("Loading training data...");
d.load_train_data(train_fname,d.data,d.label,d.index);
vector<int> valid_index(d.index.begin(), d.index.begin()+d.index.size()*valid_ratio);
vector<int> train_index(d.index.begin()+d.index.size()*valid_ratio, d.index.end());
//Pre-training
for (int layer = 0; layer < layers.size() - 1; layer++)
{
d.initDeltaRBM(layer);
for (int epoch = 0; epoch<max_epoch; epoch++){
random_shuffle(train_index.begin(), train_index.end());
int j = 0;
for (vector<int>::iterator it = train_index.begin(); it != train_index.end(); ++it, ++j){
d.gibbSample(layer, d.data[*it]);
if ((j % d.batch_size == 0) && j != 0)
d.updateRBM(layer);
}
}
}
//Training
puts("Start training...");
float best_error_rate = 1;
int flat_count = 0;
for(int epoch=0;epoch<max_epoch;epoch++){
random_shuffle(train_index.begin(), train_index.end());
int j=0;
for(vector<int>::iterator it=train_index.begin();it!=train_index.end();++it,++j){
mat y = zeros<mat>(layers.back(),1);
d.feedforward(d.data[*it]);
y(d.label[*it],0) = 1;
d.backprop(y);
if((j % d.batch_size == 0) && j != 0)
d.update();
}
float train_err = d.report_error_rate(d.data,d.label, train_index);
float valid_err = d.report_error_rate(d.data,d.label, valid_index);
printf("epoch %d\ttrain err:%f\tvalid err:%f\n", epoch,train_err,valid_err);
d.save_model(output_model, structure);
if (valid_err < best_error_rate)
{
best_error_rate = valid_err;
bestModel = d;
}
if (valid_err - best_error_rate > early_stop_thres)
{
bestModel.save_model(output_model, structure);
printf("early stop here, and save the best model\n");
break;
}
}
}
