Using Gurobi in ICLR experiments.

[zhe-tao]

Request

It would be nice if the interface provides the option to use Gurobi. For now, the implementation is hard-coded to use scipy.optimize.linprog. For example:

REASSURE/ICLR/tools/build_PNN.py

Lines 121 to 124 in 4781c8e

	tem1 = linprog(np.matmul(P[i], f1), A, b, bounds=(None, None)).fun
	# tem1 = solve_LP(np.matmul(P[i], f1), A, b, bounds=None)
	tem2 = -linprog(-np.matmul(P[i], f1), A, b, bounds=(None, None)).fun
	# tem2 = solve_LP(np.matmul(P[i], f1), A, b, bounds=None, is_minimum=False)

REASSURE/ICLR/tools/build_PNN.py

Lines 61 to 64 in 4781c8e

	out_lowerbound = linprog(c[i], bound_A, bound_b, bounds=[None, None]).fun + d[i]
	# out_lowerbound = solve_LP(c[i].detach().numpy(), bound_A, bound_b, bounds=None)+d[i]
	out_upperbound = -linprog(-c[i], bound_A, bound_b, bounds=[None, None]).fun + d[i]
	# out_upperbound = solve_LP(c[i].detach().numpy(), bound_A, bound_b, is_minimum=False, bounds=None)+d[i]

REASSURE/ICLR/tools/gurobi_solver.py

Lines 14 to 48 in 4781c8e

	def solve_LP(c: np.array, A: np.array, b: np.array, bounds, is_minimum=True, DualReductions=1):
	# assert len(c) == len(A[0])
	# assert len(A) == len(b)
	# x_dim = len(c)
	# m = Model()
	# blockPrint()
	# m.Params.DualReductions = DualReductions
	# m.Params.LogToConsole = 0
	# enablePrint()
	# if bounds:
	# x = m.addMVar(x_dim, lb=bounds[0], ub=bounds[1])
	# else:
	# x = m.addMVar(x_dim, lb=-GRB.INFINITY, ub=GRB.INFINITY)
	# if is_minimum:
	# m.setObjective(c @ x, GRB.MINIMIZE)
	# else:
	# m.setObjective(c @ x, GRB.MAXIMIZE)
	# m.addConstr(A @ x <= b, name='Ab')
	# m.optimize()
	# if m.status == GRB.UNBOUNDED:
	# if is_minimum:
	# return -np.inf
	# else:
	# return np.inf
	# elif m.status == GRB.OPTIMAL:
	# obj = m.getObjective()
	# return obj.getValue()
	# elif DualReductions == 1:
	# return solve_LP(c, A, b, bounds, is_minimum, DualReductions=0)
	# else:
	# print('error')
	if is_minimum:
	return linprog(c, A, b, bounds).fun
	else:
	return -linprog(-c, A, b, bounds).fun

Possible Solution

Perhaps the solve_LP function could be the interface to decide whether to use Gurobi or not.

Comments

Gurobi by default uses up to 32 threads. This restriction can be lifted by setting Model.Params.Threads=os.cpu_count().