cvxr / TFOCS

A MATLAB toolbox for building first-order solvers for convex models.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

TFOCS continuation bug

epsilon-net opened this issue · comments

I am reporting the problem below after I didn't get any response on CVX Forum.

I want to solve a linear program with the tfocs_SCD routine of TFOCS (tried v1.3 and v1.4), that generally looks like

However. after the first continuation iteration TFOCS returns an error message. Below is an example of the code followed by the error message.

SAMPLE CODE

opts = [];
opts.continuation = true;
opts.stopCrit = 4;
contopts.maxIts = 6;
mu = 1e-1;

A=randn(10,2);
a=randn(2,1);
x = tfocs_SCD(smooth_linear(a),{A, zeros(10,1)},prox_l1,mu,[],[],opts,contopts)

ERROR MESSAGE
--- Continuation step 1 of 6, mu: 1.0e-01
Auslender & Teboulle's single-projection method
Iter Objective |dx|/|x| step
----+----------------------------------
62 | -3.81899e-01 6.38e-04 1.43e-02
Finished: Step size tolerance reached
Continuation statistics: mu/2||x-x_0||^2: 9.6e-03
--- Continuation step 2 of 6, mu: 1.0e-01
Error using tfocs_initialize (line 519)
Could not determine the dimensions of the problem. Please supply an explicit value for x0.

Error in tfocs_AT (line 13)
tfocs_initialize

Error in tfocs (line 105)
[ varargout{1:max(nargout,1)} ] = feval( [ 'tfocs_', alg ], smoothF, affineF, projectorF, x0, opts );

Error in tfocs_SCD>solver (line 193)
[varargout{1:max(nargout,2)}] = tfocs( smoothF, affineF, dualproxF, z0, opts );

Error in tfocs_SCD>@(mu,x0,z0,opts)solver(objectiveF,affineF,dualproxF,mu,x0,z0,opts) (line 118)
continuation_solver=@(mu,x0,z0,opts)solver(objectiveF,affineF,dualproxF, mu,x0,z0,opts);

Error in continuation (line 129)
[x, odata, optsOut ] = fcn( mu, x0, z0, optsTemp );

Error in tfocs_SCD (line 119)
[ x, odata, opts ] = continuation( continuation_solver, mu, x0, z0, opts,contOpts );

I think the issue is only caused when the size compatibility for the affine operator is checked in tfocs_initialize. One of the methods to check the mentioned sizes appears to have a bug (I guess the second method). I managed to make my code run without any issue just by commenting line 519 of tfocs_initialize which spits the error message. I am looking forward to a more reasonable resolution for this bug.