fix: scipy time limit raises SolverError (#2080)

* fix: scipy time limit raises SolverError

* fix: only allow optimal_inaccurate from scipy if solution is present

* chore: lint fix

cvxpy/reductions/solvers/conic_solvers/scipy_conif.py

@@ -42,7 +42,7 @@ class SCIPY(ConicSolver):
 
     # Map of SciPy linprog status
     STATUS_MAP = {0: s.OPTIMAL,  # Optimal
-                  1: s.SOLVER_ERROR,  # Iteration limit reached
+                  1: s.OPTIMAL_INACCURATE,  # Iteration/time limit reached
                   2: s.INFEASIBLE,  # Infeasible
                   3: s.UNBOUNDED,  # Unbounded
                   4: s.SOLVER_ERROR  # Numerical difficulties encountered
@@ -217,7 +217,12 @@ def _log_scipy_method_warning(self, meth):
     def invert(self, solution, inverse_data):
         """Returns the solution to the original problem given the inverse_data.
         """
-        status = self.STATUS_MAP[solution['status']]
+        status = self.STATUS_MAP[solution["status"]]
+
+        # Sometimes when the solver's time limit is reached, the solver doesn't return a solution.
+        # In these situations we correct the status from s.OPTIMAL_INACCURATE to s.SOLVER_ERROR
+        if (status == s.OPTIMAL_INACCURATE) and (solution.x is None):
+            status = s.SOLVER_ERROR
 
         primal_vars = None
         dual_vars = None

cvxpy/tests/solver_test_helpers.py

@@ -753,6 +753,32 @@ def mi_lp_5() -> SolverTestHelper:
     return sth
 
 
+def mi_lp_7() -> SolverTestHelper:
+    """Problem that takes significant time to solve - for testing time/iteration limits"""
+    np.random.seed(0)
+    n = 24 * 8
+    c = cp.Variable((n,), pos=True)
+    d = cp.Variable((n,), pos=True)
+    c_or_d = cp.Variable((n,), boolean=True)
+    big = 1e3
+    s = cp.cumsum(c * 0.9 - d)
+    p = np.random.random(n)
+    objective = cp.Maximize(p @ (d - c))
+    constraints = [
+        d <= 1,
+        c <= 1,
+        s >= 0,
+        s <= 1,
+        c <= c_or_d * big,
+        d <= (1 - c_or_d) * big,
+    ]
+    return SolverTestHelper(
+        (objective, None),
+        [(c, None,), (d, None), (c_or_d, None)],
+        [(con, None) for con in constraints]
+    )
+
+
 def mi_socp_1() -> SolverTestHelper:
     """
     Formulate the following mixed-integer SOCP with cvxpy

cvxpy/tests/test_conic_solvers.py

@@ -24,6 +24,7 @@
 
 import cvxpy as cp
 import cvxpy.tests.solver_test_helpers as sths
+from cvxpy import SolverError
 from cvxpy.reductions.solvers.defines import (
     INSTALLED_MI_SOLVERS,
     INSTALLED_SOLVERS,
@@ -1994,6 +1995,19 @@ def test_scipy_mi_lp_4(self) -> None:
     def test_scipy_mi_lp_5(self) -> None:
         StandardTestLPs.test_mi_lp_5(solver='SCIPY')
 
+    @unittest.skipUnless('SCIPY' in INSTALLED_MI_SOLVERS, 'SCIPY version cannot solve MILPs')
+    def test_scipy_mi_time_limit_reached(self) -> None:
+        sth = sths.mi_lp_7()
+
+        # run without enough time to find optimum
+        sth.solve(solver='SCIPY', scipy_options={"time_limit": 0.01})
+        assert sth.prob.status == cp.OPTIMAL_INACCURATE
+        assert sth.objective.value > 0
+
+        # run without enough time to do anything
+        with pytest.raises(SolverError):
+            sth.solve(solver='SCIPY', scipy_options={"time_limit": 0.})
+
 
 @unittest.skipUnless('COPT' in INSTALLED_SOLVERS, 'COPT is not installed.')
 class TestCOPT(unittest.TestCase):