int
main(int argc, const char **argv)
{
parse(argc, argv);
// load training set
load(trainfile.c_str(), xtrain, ytrain);
cout << "Number of features " << dim << "." << endl;
int imin = 0;
int imax = xtrain.size() - 1;
if (maxtrain > 0 && imax >= maxtrain)
imax = imin + maxtrain -1;
// prepare svm
SvmSgd svm(dim, lambda);
Timer timer;
// load testing set
if (! testfile.empty())
load(testfile.c_str(), xtest, ytest);
int tmin = 0;
int tmax = xtest.size() - 1;
for(int i=0; i<epochs; i++)
{
cout << "--------- Epoch " << i+1 << "." << endl;
timer.start();
svm.train(imin, imax, xtrain, ytrain);
timer.stop();
cout << "Total training time " << setprecision(6)
<< timer.elapsed() << " secs." << endl;
svm.test(imin, imax, xtrain, ytrain, "train: ");
if (tmax >= tmin)
svm.test(tmin, tmax, xtest, ytest, "test: ");
}
}
void
load(const char *fname, xvec_t &xp, yvec_t &yp)
{
cout << "Loading " << fname << "." << endl;
igzstream f;
f.open(fname);
if (! f.good())
{
cerr << "ERROR: cannot open " << fname << "." << endl;
exit(10);
}
int pcount = 0;
int ncount = 0;
bool binary;
string suffix = fname;
if (suffix.size() >= 7)
suffix = suffix.substr(suffix.size() - 7);
if (suffix == ".dat.gz")
binary = false;
else if (suffix == ".bin.gz")
binary = true;
else
{
cerr << "ERROR: filename should end with .bin.gz or .dat.gz" << endl;
exit(10);
}
while (f.good())
{
SVector x;
double y;
if (binary)
{
y = (f.get()) ? +1 : -1;
x.load(f);
}
else
{
f >> y >> x;
}
if (f.good())
{
assert(y == +1 || y == -1);
xp.push_back(x);
yp.push_back(y);
if (y > 0)
pcount += 1;
else
ncount += 1;
if (x.size() > dim)
dim = x.size();
}
if (trainsize > 0 && xp.size() > (unsigned int)trainsize)
break;
}
cout << "Read " << pcount << "+" << ncount
<< "=" << pcount + ncount << " examples." << endl;
}
int main(int argc, const char **argv)
{
parse(argc, argv);
config(argv[0]);
if (trainfile)
load_datafile(trainfile, xtrain, ytrain, dims, normalize, maxtrain);
if (testfile)
load_datafile(testfile, xtest, ytest, dims, normalize);
cout << "# Number of features " << dims << "." << endl;
// prepare svm
int imin = 0;
int imax = xtrain.size() - 1;
int tmin = 0;
int tmax = xtest.size() - 1;
// heuristic determination of averaging start point
int avgfrom = fabs(avgstart) * (imax - imin + 1);
avgfrom = (avgstart < 0 && dims < avgfrom) ? dims : avgfrom;
// create
SvmAisgd svm(dims, lambda, avgfrom);
Timer timer;
// determine eta0 using sample
int smin = 0;
int smax = imin + min(1000, imax);
timer.start();
svm.determineEta0(smin, smax, xtrain, ytrain);
timer.stop();
// train
for(int i=0; i<epochs; i++)
{
cout << "--------- Epoch " << i+1 << "." << endl;
timer.start();
svm.train(imin, imax, xtrain, ytrain);
timer.stop();
cout << "Total training time " << setprecision(6)
<< timer.elapsed() << " secs." << endl;
svm.test(imin, imax, xtrain, ytrain, "train: ");
if (tmax >= tmin)
svm.test(tmin, tmax, xtest, ytest, "test: ");
}
svm.renorm();
// Linear classifier is in svm.a and svm.aBias
return 0;
}
int main(int argc, const char **argv)
{
parse(argc, argv);
config(argv[0]);
if (trainfile)
load_datafile(trainfile, xtrain, ytrain, dims, normalize, maxtrain);
if (testfile)
load_datafile(testfile, xtest, ytest, dims, normalize);
cout << "# Number of features " << dims << "." << endl;
// prepare svm
int imin = 0;
int imax = xtrain.size() - 1;
int tmin = 0;
int tmax = xtest.size() - 1;
SvmSag svm(dims, lambda);
Timer timer;
// determine eta0 using sample
int smin = 0;
int smax = imin + min(1000, imax);
timer.start();
if (eta > 0)
svm.setEta(eta);
else
svm.determineEta(smin, smax, xtrain, ytrain);
timer.stop();
// train
for(int i=0; i<epochs; i++)
{
cout << "--------- Epoch " << i+1 << "." << endl;
timer.start();
if (i == 0)
svm.trainInit(imin, imax, xtrain, ytrain);
else
svm.trainSag(imin, imax, xtrain, ytrain);
timer.stop();
cout << "Total training time " << setprecision(6)
<< timer.elapsed() << " secs." << endl;
svm.test(imin, imax, xtrain, ytrain, "train: ");
if (tmax >= tmin)
svm.test(tmin, tmax, xtest, ytest, "test: ");
}
return 0;
}
int main(int argc, const char **argv)
{
parse(argc, argv);
config(argv[0]);
if (trainfile)
load_datafile(trainfile, xtrain, ytrain, dims, normalize, maxtrain);
if (testfile)
load_datafile(testfile, xtest, ytest, dims, normalize);
cout << "# Number of features " << dims << "." << endl;
// prepare svm
int imin = 0;
int imax = xtrain.size() - 1;
int tmin = 0;
int tmax = xtest.size() - 1;
SvmSgd svm(dims, lambda);
Timer timer;
// determine eta0 using sample
int smin = 0;
int smax = imin + min(1000, imax);
// train
Timer totalTimer, overheadtimer, exetimer;
totalTimer.start();
//winnie, initial wDivisor and wBias
timeval t1, t4, t5, t6, t7, t8;
overheadtimer.start();
if(sample_file){
int sample_size = 500;
int bin_num = 20;
int num_compare = 49;
int dimension = dims - 1; //The first feature is the classification, does not count in sampling.
gettimeofday(&t1, NULL);
Sampling<double> selector(imax, 0 ,
dimension, sample_size, bin_num, num_compare);
selector.do_sampling(sample_file);
gettimeofday(&t4, NULL);
selector.calc_ecdf();
gettimeofday(&t5, NULL);
// std::cout << "test init kmeans " << std::endl;
//step3a, do database search
if(num_compare <= 0){
cerr << "Number of comparison is less than or equal to 0" << endl;
return -1;
}
else{
//TODO: make the comparison choose the second best result, when we use data base that contains dataset itself
//winnie, prerun several iterations, and log some information
int prerun_iters = 3;
FVector old_w(dimension);
svm.determineEta0(smin, smax, xtrain, ytrain);
for(int i = 0; i < prerun_iters; i ++ ){
svm.get_w(old_w);
svm.train(imin, imax, xtrain, ytrain);
//svm.test(imin, imax, xtrain, ytrain, "train: ");
}
//get the idx of dimensions
map <double, int> delta_w;
svm.get_delta_w(old_w, delta_w);
//int reducedDimNum[8] = {1, 2, 3, 4, 8, 16, 32, 50};
int reducedDimNum[8] = {1, 3, 8, 16, 32, 50, 64, 128};
int selected_id[9];
multimap <double, int> error_dim; //error value and dimension
for(int iout = 0; iout < 8; iout++){
vector<int> reducedDimIdx(reducedDimNum[iout]);
map<double, int>::reverse_iterator rmit = delta_w.rbegin();
double sum_value = 0;
for(int i = 0; i < reducedDimNum[iout]; i ++){
reducedDimIdx[i] = rmit->second;
sum_value += rmit->first;
//cout << "reduceDim: " << reducedDimIdx[i] << " value " << rmit->first << endl;
rmit++;
}
cout << "sum_value: " << sum_value << " with dim num: " << reducedDimNum[iout] << endl;
//cout << "distancetype: " << distancetype << endl;
selected_id[iout] = selector.search_database('b', database_file, distancetype, reducedDimIdx);
gettimeofday(&t6, NULL);
if(selected_id[iout] < 0)
return -1;
else{
std::cout << "selected id: " << selected_id[iout] << std::endl;
//Do data customization, use norm.cpp program,
stringstream ss;//create a stringstream
ss << setfill('0') << setw(2) << selected_id[iout];
string filename = string(database_dir) + "/" + ss.str() + ".txt";
// double value1, value2;
// selector.load_memo(filename.c_str(), &value1, &value2);
cout << "Init with file: " << filename << endl;
svm.init_w(filename.c_str(), dims);
gettimeofday(&t7, NULL);
//run one iter use current w
double pre_loss_1; //, pre_loss_2;
//double pre_loss_delta_1; //, pre_loss_delta_2;
//svm.train(imin, imax, xtrain, ytrain);
svm.test(tmin, tmax, xtest, ytest, "test: ");
pre_loss_1 = svm.get_cost();
/*
svm.train(imin, imax, xtrain, ytrain);
svm.test(tmin, tmax, xtest, ytest, "test: ");
pre_loss_2 = svm.get_cost();
pre_loss_delta_1 = pre_loss_1 - pre_loss_2; */
double pre_loss_delta_ratio = pre_loss_1; //method A
//double pre_loss_delta_ratio = pre_loss_delta_1; //method B
//double pre_loss_delta_ratio = -svm.t; //method C
cout << "-svm.t: " << -svm.t << endl;
//svm.train(imin, imax, xtrain, ytrain);
//svm.test(tmin, tmax, xtest, ytest, "test: ");
//pre_loss_3 = svm.get_cost();
//pre_loss_delta_2 = pre_loss_2 - pre_loss_3;
/*
double pre_loss_delta_ratio;
if(pre_loss_delta_1 < 0 || pre_loss_delta_2 < 0){
pre_loss_delta_ratio = 1000;
}
else
pre_loss_delta_ratio = predict_iter(pre_loss_delta_1, pre_loss_delta_2, pre_loss_2);
cout << "pre_loss_1: " << pre_loss_1 << endl;
cout << "pre_loss_2: " << pre_loss_2 << endl;
cout << "pre_loss_3: " << pre_loss_3 << endl;
cout << "pre_loss_delta_1: " << pre_loss_delta_1 << endl;
cout << "pre_loss_delta_2: " << pre_loss_delta_2 << endl;
*/
cout << "insert pre_loss_ratio " << pre_loss_delta_ratio << " at dim " << reducedDimNum[iout] << endl;
error_dim.insert(pair<double, int>(pre_loss_delta_ratio, iout));
//error_dim.insert(pair<double, int>(fabs(pre_loss_new - pre_loss_old), iout));
}
}
stringstream ss;//create a stringstream
ss << setfill('0') << setw(2) << selected_id[error_dim.begin()->second];
string filename = string(database_dir) + "/" + ss.str() + ".txt";
cout << "Init with: " << filename << endl;
svm.init_w(filename.c_str(), dims);
cout << "chosen dim num: " << reducedDimNum[error_dim.begin()->second] << " id: " << selected_id[error_dim.begin()->second] << endl;
cout << "error of each dim_reduction: " << endl;
cout << error_dim.size() << "in total\n" ;
for(map<double, int>::iterator mit = error_dim.begin(); mit != error_dim.end(); mit++){
cout << mit->second << " " << mit->first << endl;
}
//winnie, end of prerun
}
}
else{
if(!initpara_file){
cerr << "initpara_file does not exist\n";
}
else
svm.init_w(initpara_file, dims);
}
overheadtimer.stop();
//winnie, end of initial wDivisor and wBias
exetimer.start();
timer.start();
//double temp_t = 1;
//Original version
//svm.determineEta0(imin, imax, xtrain, ytrain);
//svm.determineEta0(smin, smax, xtrain, ytrain);
//svm.t = 0;
//Version 3 --------------------------------------------------------
svm.my_determineEta0_v3(imin, imax, xtrain, ytrain);
//svm.my_determineEta0_v3(smin, smax, xtrain, ytrain);
//svm.t = 0;
//svm.t = 500000;
//Version 2 --------------------------------------------------------
//svm.my_determineEta0_v2(smin, smax, xtrain, ytrain, svm.t, 4);
//svm.my_determineEta0_v2(imin, imax, xtrain, ytrain, svm.t, 4);
//svm.my_determineEta0(imin, imax, xtrain, ytrain, svm.t);
timer.stop();
//svm.eta0 = svm.eta0 / 10;
//svm.t = 0;
cout << "New eta0: " << svm.eta0 << " t: " << svm.t << endl;
cout << "eta: " << setprecision(12) << svm.eta0 / (1 + lambda * svm.eta0 * svm.t) << endl;
//winnie, modify code end condition from number of epochs to misclassification rate
double loss_new = 1, loss_old = 1;
int i = 0;
// for(int i=0; i<epochs; i++)
if(revalue == 0){
svm.test(imin, imax, xtrain, ytrain);
loss_new = svm.get_cost();
while(loss_old - loss_new > 1e-5 || loss_new - loss_old > 1e-5)
{
cout << "--------- Epoch " << ++i << endl;
//timer.start();
svm.train(imin, imax, xtrain, ytrain);
//timer.stop();
//cout << "Total training time " << setprecision(6)
// << timer.elapsed() << " secs." << endl;
svm.test(imin, imax, xtrain, ytrain, "train: ");
if (tmax >= tmin)
svm.test(tmin, tmax, xtest, ytest, "test: ");
loss_old = loss_new;
loss_new = svm.get_cost();
}
}
else{
cout<< "[Main: svm.wDivisor: ]" << setprecision(12) <<svm.wDivisor << " svm.t: " << svm.t << "svm.eta0: " << svm.eta0 << endl;
while(loss_new > revalue)
{
cout << "--------- Epoch " << ++i << endl;
//timer.start();
svm.train(imin, imax, xtrain, ytrain);
//timer.stop();
//cout << "Total training time " << setprecision(6)
// << timer.elapsed() << " secs." << endl;
svm.test(imin, imax, xtrain, ytrain, "train: ");
loss_old = loss_new;
loss_new = svm.get_cost();
if (tmax >= tmin)
svm.test(tmin, tmax, xtest, ytest, "test: ");
if(i >= 200){
cout << "Exit because reach 2 epochs\n";
break;
}
}
}
cout << "Loss_new: " << loss_new << " Loss_old:" << loss_old << endl;
exetimer.stop();
totalTimer.stop();
cout << "Total time: " << setprecision(6)
<< totalTimer.elapsed() << " secs." << endl;
cout << "Total exe time: " << setprecision(6)
<< exetimer.elapsed() << " secs." << endl;
cout << "Total overhead time: " << setprecision(6)
<< overheadtimer.elapsed() << " secs." << endl;
if(sample_file){
double elapsedTime;
elapsedTime =(t4.tv_sec - t1.tv_sec) * 1000.0 + (t4.tv_usec - t1.tv_usec) / 1000.0; // sec and us to ms
std::cout <<"Overhead s1: "<< elapsedTime << " ms.\n";
elapsedTime =(t5.tv_sec - t4.tv_sec) * 1000.0 + (t5.tv_usec - t4.tv_usec) / 1000.0; // sec and us to ms
std::cout <<"Overhead s2: "<< elapsedTime << " ms.\n";
elapsedTime =(t6.tv_sec - t5.tv_sec) * 1000.0 + (t6.tv_usec - t5.tv_usec) / 1000.0; // sec and us to ms
std::cout <<"Overhead s3: "<< elapsedTime << " ms.\n";
elapsedTime =(t7.tv_sec - t6.tv_sec) * 1000.0 + (t7.tv_usec - t6.tv_usec) / 1000.0; // sec and us to ms
std::cout <<"Overhead s4: "<< elapsedTime << " ms.\n";
elapsedTime =(t8.tv_sec - t7.tv_sec) * 1000.0 + (t8.tv_usec - t7.tv_usec) / 1000.0; // sec and us to ms
std::cout <<"Overhead s5: "<< elapsedTime << " ms.\n";
}
//winnie, output parameter wDivisor and wBias:
cout << "wDivisor: " << svm.get_wDivisor() << endl;
cout << "wBias: " << svm.get_wBias() << endl;
svm.output_w(outputfile);
return 0;
}
