Logging

Logging in C

In SuperSCS, we may activate the logging of progress information using the property do_record_progress in ScsSettings.

In that case, the Info object that is updated using scs contains the following information:

• progress_iter an array of the iterations when the progress data were recorded. It is possible that progress data are stored with a stride.
• progress_time an array of timestamps (starting at 0) when the progress data were recorded
• progress_norm_fpr the values of the fixed-point residual at every iteration.
• progress_respri the recoded primal residuals
• progress_resdual the recorded dual residuals
• progress_relgap an array of the recoded values of the relative gap
• progress_pcost an array of the recoded values of the primal cost
• progress_dcost an array of the recoded values of the dual cost
• progress_ls the number of line-search iterations at each iteration of the algorithm
• progress_mode the type of step that was taken at every iteration; -1 corresponds to a nominal step, 0 is for a K0 step and 1 and 2 are for K1 and K2 steps respectively.

Note that the above arrays have length equal to the number of iterations.

Example:

scs_int i;
const scs_int column_size = 10;
printf("  i     P Cost    D Cost       Gap       FPR      PRes      DRes\n");
printf("----------------------------------------------------------------\n");
for (i = 0; i < info->iter; ++i) {
      printf("%*i ", 3, info->progress_iter[i]);
      printf("%*.2e", column_size, info->progress_pcost[i]);
      printf("%*.2e", column_size, info->progress_dcost[i]);
      printf("%*.2e", column_size, info->progress_relgap[i]);
      printf("%*.2e", column_size, info->progress_norm_fpr[i]);
      printf("%*.2e", column_size, info->progress_respri[i]);
      printf("%*.2e", column_size, info->progress_resdual[i]);
      printf("\n");
}

Logging in MATLAB

Similarly, we may record progress information in MATLAB when we call scs_direct.m, scs_indirect.m or scs.m.

All we have to do is to set do_record_progress to 1.

params.do_record_progress = 1; 
[x, y, s, info] = scs_direct(data, K, params);

This is the info structure we get when we run tests/matlab/cone_test.m:

info = 

                      iter: 32
                    status: 'Solved'
                      pobj: -16.3754
                      dobj: -16.3754
                    resPri: 6.7545e-12
                   resDual: 2.2441e-11
                 resInfeas: NaN
                  resUnbdd: 0.3220
                    relGap: 2.6977e-12
                 setupTime: 0.0507
                 solveTime: 1.2677
             progress_iter: [32x1 double]
           progress_relgap: [32x1 double]
           progress_respri: [32x1 double]
          progress_resdual: [32x1 double]
            progress_pcost: [32x1 double]
            progress_dcost: [32x1 double]
             progress_time: [32x1 double]
             progress_mode: [32x1 double]
               progress_ls: [32x1 double]
    allocated_memory_bytes: 14528

Logging via CVX

Likewise, when calling SuperSCS via CVX, set do_record_progress to 1 and dumpfile to a destination MAT file where the output is to be stored.

CVX cannot return info as an argument, but it can dump all details in a MAT file.

This contains:

• data: the sparse matrix A and the vectors b and c of the problem
• K: The cone
• output: What the solver prints while running
• pars: The parameters that we passed to the solver
• xx, yy and ss: The (x,y,s)-solution
• info: The info structure containing the progress information.

Here is a complete example:

rng('default');
rng(1);

d = 200;
p = 5;
A = sprandn(p,d,0.3,0.0001);
S = full(A'*A);

lam = 2;

cvx_begin sdp
    cvx_solver scs
    cvx_solver_settings('eps', 1e-3,...
        'verbose', 1,...
        'do_super_scs', 1, ...
        'direction', 100, ...
        'memory', 100,...
        'do_record_progress',1,...
        'dumpfile','pca.mat');
    variable X(d,d) symmetric
    minimize(-trace(S*X) + lam*norm(X,1))
    trace(X)==1
    X>=0
cvx_end  

We may then load the file pca.mat and plot the results

load('pca.mat')
semilogy(info.progress_time, info.progress_respri,'linewidth',2); hold on
semilogy(info.progress_time, info.progress_resdual,'linewidth',2);
legend('Primal Residual', 'Dual Residual');
xlabel('time [ms]'); ylabel('residual');
grid on

See Also

SuperSCS tutorial

Saving and loading problems

Useful cones and how to use them