void mmEM::training(param myParam)
{
vector_2str wordX;
vector_2str wordY;
bool stillTrain = true;
// reading file //
readFileXY(myParam, myParam.inputFilename, &wordX, &wordY);
// initialization //
cout << "Initialization ... ";
initialization(myParam, wordX, wordY);
// maximization //
cout << "Maximization ... " << endl;
maximization(myParam);
cout << endl;
int atIter = 0;
// still training //
while (stillTrain)
{
atIter++;
cout << "Iteration " << atIter << endl;
// for each x and y pair //
cout << "Expectation ... " ;
for (int i = 0; i < wordX.size(); i++)
{
// expectation //
expectation(myParam, wordX[i], wordY[i]);
}
cout << "Maximization ... ";
long double totalChange = maximization(myParam);
cout << "Total probability change = " << totalChange << endl;
// stop by the probability change condition //
if ((totalChange <= myParam.cutOff) && (myParam.cutOff < 1))
{
stillTrain = false;
}
// stop by the number of iteration condition //
if ((myParam.cutOff >= 1) && (atIter >= myParam.cutOff))
{
stillTrain = false;
}
}
cout << endl;
}
void IBM1::em(int T)
{
int total_counts;
for(int i = 0; i < T; ++i){
cout << "Doing EM iteration " << i+1 << " for IBM1." << endl;
expectation();
maximization();
}
}
void em(char* dataset, int k, char* start, char* dir)
{
FILE* lhood_fptr;
char string[100];
int iteration;
double convergence = 1, lhood = 0, lhood_old = 0;
corpus* corpus;
llna_model *model;
llna_ss* ss;
time_t t1,t2;
double avg_niter, converged_pct, old_conv = 0;
gsl_matrix *corpus_lambda, *corpus_nu, *corpus_phi_sum;
short reset_var = 1;
// read the data and make the directory
corpus = read_data(dataset);
mkdir(dir, S_IRUSR|S_IWUSR|S_IXUSR);
// set up the log likelihood log file
sprintf(string, "%s/likelihood.dat", dir);
lhood_fptr = fopen(string, "w");
// run em
model = em_initial_model(k, corpus, start);
ss = new_llna_ss(model);
corpus_lambda = gsl_matrix_alloc(corpus->ndocs, model->k);
corpus_nu = gsl_matrix_alloc(corpus->ndocs, model->k);
corpus_phi_sum = gsl_matrix_alloc(corpus->ndocs, model->k);
time(&t1);
init_temp_vectors(model->k-1); // !!! hacky
iteration = 0;
sprintf(string, "%s/%03d", dir, iteration);
write_llna_model(model, string);
do
{
printf("***** EM ITERATION %d *****\n", iteration);
expectation(corpus, model, ss, &avg_niter, &lhood,
corpus_lambda, corpus_nu, corpus_phi_sum,
reset_var, &converged_pct);
time(&t2);
convergence = (lhood_old - lhood) / lhood_old;
fprintf(lhood_fptr, "%d %5.5e %5.5e %5ld %5.5f %1.5f\n",
iteration, lhood, convergence, (int) t2 - t1, avg_niter, converged_pct);
if (((iteration % PARAMS.lag)==0) || isnan(lhood))
{
sprintf(string, "%s/%03d", dir, iteration);
write_llna_model(model, string);
sprintf(string, "%s/%03d-lambda.dat", dir, iteration);
printf_matrix(string, corpus_lambda);
sprintf(string, "%s/%03d-nu.dat", dir, iteration);
printf_matrix(string, corpus_nu);
}
time(&t1);
if (convergence < 0)
{
reset_var = 0;
if (PARAMS.var_max_iter > 0)
PARAMS.var_max_iter += 10;
else PARAMS.var_convergence /= 10;
}
else
{
maximization(model, ss);
lhood_old = lhood;
reset_var = 1;
iteration++;
}
fflush(lhood_fptr);
reset_llna_ss(ss);
old_conv = convergence;
}
while ((iteration < PARAMS.em_max_iter) &&
((convergence > PARAMS.em_convergence) || (convergence < 0)));
sprintf(string, "%s/final", dir);
write_llna_model(model, string);
sprintf(string, "%s/final-lambda.dat", dir);
printf_matrix(string, corpus_lambda);
sprintf(string, "%s/final-nu.dat", dir);
printf_matrix(string, corpus_nu);
fclose(lhood_fptr);
}
int EMclustering::EM(int k, int *IDX, bool spatial, bool att)
{
clusternum = k;
MatrixXf x;
/*if(spatial)
{
if(att)
{
x.resize(4,dataSize);
for(int i=0;i<dataSize;i++)
{
x(0,i) = dataPos[i][0];
x(1,i) = dataPos[i][1];
x(2,i) = dataPos[i][2];
x(3,i) = dataDen[i];
}
}
else
{
x.resize(6,dataSize);
for(int i=0;i<dataSize;i++)
{
x(0,i) = dataPos[i][0];
x(1,i) = dataPos[i][1];
x(2,i) = dataPos[i][2];
x(3,i) = dataVel[i][0];
x(4,i) = dataVel[i][1];
x(5,i) = dataVel[i][2];
}
}
}
else
{*/
if(att)
{
x.resize(1,dataDen.size());
for(int i=0;i<dataDen.size();i++)
{
x(0,i) = dataDen[i];
}
//cerr<<x;
//cerr<<endl;
if(k>dataDen.size())
return -1;
}
else
{
x.resize(3,dataSize);
for(int i=0;i<dataSize;i++)
{
x(0,i) = dataVel[i][0];//fabs(cos(-PI/4)*dataVel[i][0] - sin(-PI/4)*dataVel[i][1]);
x(1,i) = dataVel[i][1];//fabs(sin(-PI/4)*dataVel[i][0] + cos(-PI/4)*dataVel[i][1]);
x(2,i) = dataVel[i][2];
}
if(k>dataSize)
return -1;
}
//}
//cout<<"EM for Gaussian mixture: running ... "<<endl;
//cerr<<x<<endl;
MatrixXf r =initialization(x,k);// kmeans(x,k);//
//cerr<<"Initialization is Done"<<endl;//cerr<<r<<endl;
VectorXi label(r.rows());
for(int i=0;i<r.rows();i++)
{
int index;
float tmp1 = r.row(i).maxCoeff(&index);
label(i) = index;
}//cerr<<label<<endl;
VectorXi tmpp(label.size());
VectorXi tmp2 = unique(label,tmpp);
int tmpd = tmp2.size(); //cerr<<tmpd<<endl;
MatrixXf tmpr(r.rows(),tmpd);
for(int i=0;i<tmpd;i++)
{
tmpr.col(i) = r.col(tmp2(i));
}//cerr<<"done1"<<endl;
r.resize(r.rows(),tmpd);
r = tmpr;//cerr<<r.cols()<<endl;
float tol = 1e-10;
int max = 300;
double llh = -9e+9;
bool converged = false;
int t = 1;
//cerr<<"done1"<<endl;
//gaussian_model model;
int clusternum_error;
MatrixXf tmpmodel;
while(!converged&&t<max)
{
t = t + 1;
gaussian_model model = maximization(x,r);//cerr<<t<<" "<<"max"<<endl;
float tmpllh = llh;
r = expectation(x,model,llh);//cerr<<t<<" "<<"exp"<<endl;
for(int i=0;i<r.rows();i++)
{
int index;
float tmp1 = r.row(i).maxCoeff(&index);
label(i) = index;
}
VectorXi u = unique(label,tmpp);//cerr<<t<<" "<<u.size()<<" "<<r.cols()<<" "<<r.rows()<<endl;
clusternum_error = clusternum - u.size();
if(r.cols()!=u.size())
{
/* tmpr.resize(r.rows(),u.size());
for(int i=0;i<u.size();i++)
{
tmpr.col(i) = r.col(u(i));
}
r.resize(r.rows(),u.size());
r = tmpr;//cerr<<"r"<<endl;*/
}
else
{
if((llh - tmpllh)<tol*abs(llh))
converged = true;
else
converged = false;
}
//cerr<<"t"<<t<<endl;
//return_model = model;
tmpmodel.resize(model.mu.rows(),model.mu.cols());
//return_model = model.mu;
tmpmodel = model.mu;
u.resize(0);
//cerr<<tmpmodel<<endl;
}
/*ofstream off2("rr");
off2<<r.row(0)<<endl;
for(int i=1;i<r.rows();i++)
if(r.row(i)!=r.row(i-1))
{off2<<x.col(i)<<" ";
off2<<r.row(i)<<endl;}
off2.close();*/
cerr<<clusternum_error<<endl;
return_model = tmpmodel;
//cerr<<label<<endl;
if (converged)
cerr<<"Converged in "<<t-1<<endl;
else
cerr<<max<<endl;
//cerr<<t-1<<endl;
for(int i=0;i<label.size();i++)
{
IDX[i] = label(i);
//cerr<<IDX[i]<<" ";
}//cerr<<endl;
//cerr<<label.size()<<endl;
x.resize(0,0);
r.resize(0,0);
tmpr.resize(0,0);
tmpmodel.resize(0,0);
label.resize(0);
tmpp.resize(0);
tmp2.resize(0);
return clusternum_error;
}
void mmEM::training(param myParam)
{
vector_2str wordX;
vector_2str wordY;
bool stillTrain = true;
// reading file //
readFileXY(myParam, myParam.inputFilename, &wordX, &wordY);
// initialization //
cout << "Initialization ... ";
initialization(myParam, wordX, wordY);
// maximization //
cout << "Maximization ... " << endl;
maximization(myParam); // lhuang: M-step after initialize (count = 1)
cout << endl;
int atIter = 0;
// still training //
global_index--;
vector<double> probs_vector(global_index), counts_vector(global_index), newprobs_vector(global_index),alpha_vector(global_index);
while (stillTrain)
{
atIter++;
cout << "Iteration " << atIter << endl;
// for each x and y pair //
cout << "Expectation ... " ;
for (int i = 0; i < wordX.size(); i++)
{
// expectation //
expectation(myParam, wordX[i], wordY[i]);
}
if (!myParam.ashish)
{
cout << "Maximization ... ";
long double totalChange = maximization(myParam); // lhuang: real M-step
cout << "Total probability change = " << totalChange << endl;
// stop by the probability change condition //
if ((totalChange <= myParam.cutOff) && (myParam.cutOff < 1))
{
stillTrain = false;
}
}
else
{
cout << "Ashish" ;
double sum_probs = 0;
for(hash_2StrDouble::iterator pos = index.begin(); pos != index.end(); pos++)
{
for (hash_StrDouble::iterator pos2 = (pos->second).begin(); pos2 != (pos->second).end(); pos2++)
{
counts_vector[index[pos->first][pos2->first]-1] = counts[pos->first][pos2->first];
// cout << pos->first << " "<< pos2->first << " "
// << counts_vector[index[pos->first][pos2->first]-1] << endl;
probs_vector[index[pos->first][pos2->first]-1] = probs[pos->first][pos2->first];
alpha_vector[index[pos->first][pos2->first]-1] = alphas[pos->first][pos2->first];
sum_probs += probs_vector[index[pos->first][pos2->first]-1];
}
}
cout << "sum " << sum_probs <<endl;
// getchar();
sum_probs = 0;
for (int i=0; i<newprobs_vector.size(); i++)
{
// cout << counts_vector[i] << " " << probs_vector[i] << " ";
sum_probs += probs_vector[i];
}
cout << "sum vector " << sum_probs <<endl;
// getchar();
projectedGradientDescentWithArmijoRule(counts_vector, probs_vector, newprobs_vector,alpha_vector,myParam.beta,myParam.pgdIter);
for(hash_2StrDouble::iterator pos = index.begin(); pos != index.end(); pos++)
{
for (hash_StrDouble::iterator pos2 = (pos->second).begin(); pos2 != (pos->second).end(); pos2++)
{
// cout << pos->first << " "<< pos2->first << " "
// << probs[pos->first][pos2->first] << " "
// << counts[pos->first][pos2->first] << " "
// << newprobs_vector[index[pos->first][pos2->first]-1] << endl;
probs[pos->first][pos2->first] = newprobs_vector[index[pos->first][pos2->first]-1];
}
}
// for (int i=0; i<newprobs_vector.size(); i++)
// {
// cout << newprobs_vector[i] << " " ;
// }
cout<< endl;
// getchar();
// clean counts //
counts.clear();
}
// stop by the number of iteration condition //
if ((myParam.cutOff >= 1) && (atIter >= myParam.cutOff))
{
stillTrain = false;
}
}
cout << endl;
}
