C++ yvec_t::at方法代码示例

		double 
SvmSgd::my_evaluateEta(int imin, int imax, const xvec_t &xp, const yvec_t &yp, double eta00)
{
		SvmSgd clone(*this); // take a copy of the current state

		cout << "[my_evaluateEta: clone.wDivisor: ]" << setprecision(12) << clone.wDivisor << " clone.t: " << clone.t << " clone.eta0: " << clone.eta0 << endl; 
		cout << "Trying eta=" << eta00 ;

		assert(imin <= imax);
		double _t = 0;
		double eta = 0;
		for (int i=imin; i<=imax; i++){
				eta = eta00 / (1 + lambda * eta00 * _t);
				//cout << "[my_evaluateEta:] Eta: " << eta << endl;
				clone.trainOne(xp.at(i), yp.at(i), eta);
				_t++;
		}
		double loss = 0;
		double cost = 0;
		for (int i=imin; i<=imax; i++)
				clone.testOne(xp.at(i), yp.at(i), &loss, 0);
		loss = loss / (imax - imin + 1);
		cost = loss + 0.5 * lambda * clone.wnorm();
		cout <<" yields loss " << loss << endl;
		// cout << "Trying eta=" << eta << " yields cost " << cost << endl;
		return cost;
}

void 
SvmSgd::test(int imin, int imax, 
             const xvec_t &xp, const yvec_t &yp, 
             const char *prefix)

{
  cout << prefix << "Testing on [" << imin << ", " << imax << "]." << endl;
  assert(imin <= imax);
  int nerr = 0;
  double cost = 0;
  for (int i=imin; i<=imax; i++)
    {
      const SVector &x = xp.at(i);
      double y = yp.at(i);
      double wx = dot(w,x);
      double z = y * (wx + bias);
      if (z <= 0)
        nerr += 1;
#if LOSS < LOGLOSS
      if (z < 1)
#endif
        cost += loss(z);
    }
  int n = imax - imin + 1;
  double loss = cost / n;
  cost = loss + 0.5 * lambda * dot(w,w);
  cout << prefix << setprecision(4)
       << "Misclassification: " << (double)nerr * 100.0 / n << "%." << endl;
  cout << prefix << setprecision(12) 
       << "Cost: " << cost << "." << endl;
  cout << prefix << setprecision(12) 
       << "Loss: " << loss << "." << endl;
}

double
SvmAisgd::evaluateEta(int imin, int imax, const xvec_t &xp, const yvec_t &yp, double eta)
{
  SvmAisgd clone(*this); // take a copy of the current state
  assert(imin <= imax);
  for (int i=imin; i<=imax; i++)
    clone.trainOne(xp.at(i), yp.at(i), eta, 1.0);
  double loss = 0;
  double cost = 0;
  for (int i=imin; i<=imax; i++)
    clone.testOne(xp.at(i), yp.at(i), &loss, 0);
  loss = loss / (imax - imin + 1);
  cost = loss + 0.5 * lambda * clone.wnorm();
  // cout << "Trying eta=" << eta << " yields cost " << cost << endl;
  return cost;
}

double SvmSgdSJE::evaluate_objective(double *xp, const yvec_t yp, double *emb_mat, double *att_mat){
	double risk = 0.0;
	cout << "start to evaluate" << endl;
	for(int i = 0; i < nsamples; i++){
		double max_score = 0.0;
		int ni = yp.at(i) - 1;
		double norm_emb = 0.0;
		double xproj[emb_dim*dims];
		memset(xproj, 0, emb_dim*dims*sizeof(double));
		for(int j = 0; j < emb_dim; j++){
			for(int k = 0; k < dims; k++){
				xproj[j] += xp[dims*i+k] * emb_mat[dims*j+k];
				norm_emb += lambda * emb_mat[dims*j+k] * emb_mat[dims*j+k];
			}
		}
		for(int c = 0; c < nclass; c++){
			double score = 0.0;
			if(c != ni)
				score += 1;
			for(int j = 0; j < emb_dim; j++){
					score += xproj[j] * (att_mat[emb_dim*c+j] - att_mat[emb_dim*ni+j]); 	
			}
			if(max_score < score){
				max_score = score;
			}
		}
		if(max_score > 0){
			risk += max_score;
		}
		else cout << "The risk is below zero" << endl;
		risk += norm_emb;
	}
	return risk;
}

/// Perform a test pass
void
SvmAisgd::test(int imin, int imax, const xvec_t &xp, const yvec_t &yp, const char *prefix)
{
  cout << prefix << "Testing on [" << imin << ", " << imax << "]." << endl;
  assert(imin <= imax);
  double nerr = 0;
  double loss = 0;
  for (int i=imin; i<=imax; i++)
    testOne(xp.at(i), yp.at(i), &loss, &nerr);
  nerr = nerr / (imax - imin + 1);
  loss = loss / (imax - imin + 1);
  double cost = loss + 0.5 * lambda * anorm();
  cout << prefix
       << "Loss=" << setprecision(12) << loss
       << " Cost=" << setprecision(12) << cost
       << " Misclassification=" << setprecision(4) << 100 * nerr << "%."
       << endl;
}

/// Perform a SAG training epoch
void 
SvmSag::trainSag(int imin, int imax, const xvec_t &xp, const yvec_t &yp, const char *prefix)
{
  cout << prefix << "Training on [" << imin << ", " << imax << "]." << endl;
  assert(imin <= imax);
  assert(imin >= sdimin);
  assert(imax <= sdimax);
  assert(eta > 0);
  uniform_int_generator generator(imin, imax);
  for (int i=imin; i<=imax; i++)
    {
      int ii = generator(); 
      trainOne(xp.at(ii), yp.at(ii), eta, ii);
      t += 1;
    }
  cout << prefix << setprecision(6) << "wNorm=" << wnorm();
#if BIAS
  cout << " wBias=" << wBias;
#endif
  cout << endl;
}

/// Perform initial training epoch
void 
SvmSag::trainInit(int imin, int imax, const xvec_t &xp, const yvec_t &yp, const char *prefix)
{
  cout << prefix << "Training on [" << imin << ", " << imax << "]." << endl;
  assert(imin <= imax);
  assert(eta > 0);
  assert(m == 0);
  sd.resize(imax - imin + 1);
  sdimin = imin;
  sdimax = imax;
  for (int i=imin; i<=imax; i++)
    {
      m += 1;
      trainOne(xp.at(i), yp.at(i), eta, i);
      t += 1;
    }
  cout << prefix << setprecision(6) << "wNorm=" << wnorm();
#if BIAS
  cout << " wBias=" << wBias;
#endif
  cout << endl;
}

/// Perform a training epoch
void
SvmSgd::train(int imin, int imax, const xvec_t &xp, const yvec_t &yp, const char *prefix)
{
#if VERBOSE
  cout << prefix << "Training on [" << imin << ", " << imax << "]." << endl;
#endif
  assert(imin <= imax);
  assert(eta0 > 0);
  for (int i=imin; i<=imax; i++)
    {
      double eta = eta0 / (1 + lambda * eta0 * t);
      trainOne(xp.at(i), yp.at(i), eta);
      t += 1;
    }
#if VERBOSE
  cout << prefix << setprecision(6) << "wNorm=" << wnorm();
#if BIAS
  cout << " wBias=" << wBias;
#endif
  cout << endl;
#endif
}

/// Perform a training epoch
		void 
SvmSgd::train(int imin, int imax, const xvec_t &xp, const yvec_t &yp, const char *prefix)
{
		cout << prefix << "Training on [" << imin << ", " << imax << "]." << endl;
		assert(imin <= imax);
		assert(eta0 > 0);

		//cout << "wDivisor: " << wDivisor << "  wBias: " << wBias<< endl;
		for (int i=imin; i<=imax; i++)
		{
				double eta = eta0 / (1 + lambda * eta0 * t);
				//cout << "[my_evaluateEta:] Eta: " << eta << endl;
				trainOne(xp.at(i), yp.at(i), eta);
				t += 1;
		}

		//cout << "\nAfter training: \n  wDivisor: " << wDivisor << "  wBias: " << wBias<< endl;
		cout << prefix << setprecision(6) << "wNorm=" << wnorm();
#if BIAS
		cout << " wBias=" << wBias;
#endif
		cout << endl;
}

/// Testing
double SvmSgdSJE::test(int imin, int imax, double *xp, const yvec_t &yp, int dims,  
		double *emb_mat, int emb_dim, double *att_mat, const char* prefix)
{
  	cout << prefix << " Testing Multi-class for [" << imin << ", " << imax << "]." << endl;

  	assert(imin <= imax);
  	int nsamples = imax-imin+1;

  	double* scores = new double[nclass*nsamples];
  	int* conf_mat = new int[nclass*nclass];
  	memset(conf_mat,0,sizeof(int)*nclass*nclass);
	for(int i = 0; i < nclass*nsamples; i++)
		scores[i] = 0.0;
	double xproj[emb_dim];
  	for(int i = 0; i < nsamples; i++)
  	{
		// project test images onto label embedding space
		for(int iy = 0; iy < emb_dim; iy++)
			xproj[iy] = 0.0;
		// project training images onto label embedding space
		double xproj_norm = 0;
		for(int iy = 0; iy < emb_dim; iy++)
		{
			for(int ix = 0; ix < dims; ix++)
			{	
				xproj[iy] += xp[dims*i + ix] * emb_mat[dims*iy + ix];
			}
			xproj_norm += xproj[iy] * xproj[iy];
		}
		// normalize the projected vector
		xproj_norm = sqrt(xproj_norm);
		if(xproj_norm != 0)
		{
			// calculate the scores using dot product similarity using classifiers
			for(int iy = 0; iy < emb_dim; iy++)
				xproj[iy] = xproj[iy] / xproj_norm;	
		}
		for(int j = 0; j < nclass; j++)
  			{
				for(int iy = 0; iy < emb_dim; iy++)
					scores[nsamples*j + i] += xproj[iy] * att_mat[emb_dim*j + iy];
			}
			
  	}

  	double nerr = 0;
  	for(int i = 0; i < nsamples; i++)
  	{
		int true_class = int(yp.at(i)-1);

		double max_score = -1.0f;
		int predicted_class = -1;

		for(int c = 0; c < nclass; c++)
		{
			if ( scores[nsamples*c+i] > max_score )
			{
				predicted_class = c;
				max_score = scores[nsamples*c+i];
			}
		}
		
		conf_mat[nclass*true_class+predicted_class]++;
		//cout << true_class << " " << predicted_class << endl;
	
		if( true_class != predicted_class )
			nerr++;		
  	}

  	nerr = nerr / nsamples;
  	cout << " Per image accuracy = " << setprecision(4) << 100-(100 * nerr) << "%." << endl;

  	double sum_diag_conf=0;
  	double sum_each_line;
  	for(int i = 0; i < nclass; i++)
  	{
		sum_each_line=0;
		for(int j=0; j < nclass; j++)
		{
			sum_each_line = sum_each_line + conf_mat[i*nclass+j];
		}

		cout << " Class = " << i << " accuracy = " << setprecision(4) << double(conf_mat[i*nclass+i]/sum_each_line) << "%." << endl;
	
		sum_diag_conf = sum_diag_conf + double(double(conf_mat[i*nclass+i])/sum_each_line);
  	}
		
  	cout << " Per " << prefix << " class accuracy = " << setprecision(4) << 100 * double(sum_diag_conf / nclass) << "%." << endl;

 	double acc = double(sum_diag_conf / nclass);
  delete[] scores;
  delete[] conf_mat;

	return acc;	
}

/// Training the svms with SJE using ranking objective
void SvmSgdSJE::train(int imin, int imax, double *xp, const yvec_t &yp, int dims,
		 double *att_mat, int cls_dim, 
		 double *emb_mat, int emb_dim,
		 const char *prefix)
{
	cout << prefix << " Training SJE for lbd = " << lambda << ", eta = " << eta << " and " << nclass << " classes" << endl;

	assert(imin <= imax);

	double xproj[emb_dim];
	//int i = 0;
	for(int i = imin; i <= imax; i++)
	{
		//i = rand()%imax + imin;
		for(int iy = 0; iy < emb_dim; iy++)
			xproj[iy] = 0.0;
		// project training images onto label embedding space
		double xproj_norm = 0;
		for(int iy = 0; iy < emb_dim; iy++)
		{
			for(int ix = 0; ix < dims; ix++)
			{	
				xproj[iy] += xp[dims*i + ix] * emb_mat[dims*iy + ix];
			}
			xproj_norm += xproj[iy] * xproj[iy];
		}
		// normalize the projected vector
		xproj_norm = sqrt(xproj_norm);
		for(int iy = 0; iy < emb_dim; iy++)
			xproj[iy] = xproj[iy] / xproj_norm;	
		int best_index = -1;
		double best_score = 0.0;
		for(int j = 0; j < nclass; j++)
		{
		  double score = 0.0;
			for(int iy = 0; iy < emb_dim; iy++)
				score += xproj[iy] * att_mat[emb_dim*j + iy];
			
			//delta(y_n,y)= 1 if y_n != y
			if(j != yp.at(i) - 1)	
				score += 1;
			if(score > best_score)
			{
				best_score = score;
				best_index = j;
			}
		}
		//update the embedding matrix when the decision is wrong
		if(best_index != int(yp.at(i)-1) && best_index != -1)
		{
	//		memcpy(prev_emb_mat, emb_mat, emb_dim*dims*sizeof(double));
			int ni = int(yp.at(i) - 1);
			for(int iy = 0; iy < emb_dim; iy++)
			{
				for(int ix = 0; ix < dims; ix++)
				{
					emb_mat[dims*iy + ix] -= eta * (xp[dims*i + ix] * (att_mat[emb_dim*best_index + iy] - att_mat[emb_dim*ni + iy]) + lambda*emb_mat[dims*iy + ix]); 
				}
			}
		}
		t += 1;
	}
	//cout << "gradient norm=" << gradient_norm(g,dims*emb_dim) << endl;
}