SCS without gpu give solutions however SCS with gpu=true return unbounded

ghost commented 5 years ago

I run into a situation that for the same problem, the behavior of SCS is different with GPU on or OFF..

With GPU = True here is the output:

    SCS v2.0.2 - Splitting Conic Solver
    (c) Brendan O'Donoghue, Stanford University, 2012-2017

Lin-sys: sparse-indirect GPU, nnz in A = 5038575, CG tol ~ 1/iter^(2.00) eps = 1.00e-05, alpha = 1.50, max_iters = 5000, normalize = 1, scale = 1.00 acceleration_lookback = 20, rho_x = 1.00e-03 Variables n = 30100, constraints m = 31100 Cones: primal zero / dual free vars: 11000 sd vars: 20100, sd blks: 1 Setup time: 2.57e+00s

Iter | pri res | dua res | rel gap | pri obj | dua obj | kap/tau | time (s)

 0| 9.18e-01  8.76e-01  2.03e-01 -3.94e+00 -6.20e+00  2.56e-16  5.25e-01 
20| 9.18e-01  8.76e-01  2.58e-01 -3.10e+00 -1.62e+00  1.47e+00  6.97e-01

Status: Unbounded Timing: Solve time: 6.98e-01s Lin-sys: avg # CG iterations: 0.00, avg solve time: 2.41e-02s Cones: avg projection time: 2.34e-04s Acceleration: avg step time: 6.71e-03s

Certificate of dual infeasibility: dist(s, K) = 1.1314e-10 |Ax + s|_2 * |c|_2 = 6.5790e-09 c'x = -1.0000

with GPU false here is the output:

    SCS v2.0.2 - Splitting Conic Solver
    (c) Brendan O'Donoghue, Stanford University, 2012-2017

Lin-sys: sparse-indirect, nnz in A = 5041955, CG tol ~ 1/iter^(2.00) eps = 1.00e-05, alpha = 1.50, max_iters = 5000, normalize = 1, scale = 1.00 acceleration_lookback = 20, rho_x = 1.00e-03 Variables n = 30100, constraints m = 31100 Cones: primal zero / dual free vars: 11000 sd vars: 20100, sd blks: 1 Setup time: 2.60e-01s

Iter | pri res | dua res | rel gap | pri obj | dua obj | kap/tau | time (s)

 0| 1.09e+21  2.12e+21  1.00e+00 -4.89e+21  1.98e+21  9.28e+20  7.11e-01

100| 1.09e-03 3.43e-03 4.30e-04 1.00e+00 1.00e+00 3.75e-16 2.81e+01 160| 5.45e-06 8.68e-06 1.73e-06 1.00e+00 1.00e+00 1.26e-16 3.98e+01

Status: Solved Timing: Solve time: 3.98e+01s Lin-sys: avg # CG iterations: 8.27, avg solve time: 2.14e-01s Cones: avg projection time: 1.97e-02s Acceleration: avg step time: 1.12e-02s

Error metrics: dist(s, K) = 1.2674e-09, dist(y, K*) = 6.9758e-10, s'y/|s||y| = -5.9677e-11 primal res: |Ax + s - b|_2 / (1 + |b|_2) = 5.4544e-06 dual res: |A'y + c|_2 / (1 + |c|_2) = 8.6825e-06 rel gap: |c'x + b'y| / (1 + |c'x| + |b'y|) = 1.7304e-06

c'x = 1.0000, -b'y = 1.0000

Does anyone has similar problems.

The problem is just a nuclear norm minimization problem with constraints (equivalent SDP with constraints)

ghost commented 5 years ago

An update:

I tried several other types problems. It is the same issue. So the problem is not due to the SDP. I also tried to run the included GPU test in SCS, it work.

So the problem is that when using GPU through cvxpy.

bodono commented 5 years ago

My guess is that your GPU is running out of memory, SCS tries to catch that but if the problem is almost able to fit in memory then weird things happen. If you run nvidia-smi at the command line you can see how much memory your GPU has and how much is being used.

cynnjjs commented 5 years ago

I'm having the same problem here. When I run problem.solve with gpu=False, I get status = solved, but with gpu=True I get Unbounded. I'm simply running the following testing program:

m = 30 n = 20 np.random.seed(1) A = np.random.randn(m, n) b = np.random.randn(m)
x = cp.Variable(n) objective = cp.Minimize(cp.sum_squares(A*x - b)) constraints = [0 <= x, x <= 1] prob = cp.Problem(objective, constraints)
result = prob.solve("SCS", gpu=False, verbose=True)

I found this post: (https://stackoverflow.com/questions/52579531/how-to-install-cvxpy-scs-to-use-with-gpu) and tried the following setup, but it did not help:

python3 setup.py install --scs --gpu --int

bodono commented 5 years ago

Should be fixed now. When installing from source remember to delete all previous build artifacts, ie, something like rm -rf build dist scs.egg-info.

cvxgrp / scs