def run_model(self, size, num_par_fd, method):
if MPI:
if MPI.COMM_WORLD.rank == 0:
mat = np.random.random(5 * size).reshape((5, size)) - 0.5
else:
mat = None
mat = MPI.COMM_WORLD.bcast(mat, root=0)
else:
mat = np.random.random(5 * size).reshape((5, size)) - 0.5
p = Problem()
model = p.model
model.add_subsystem('indep',
IndepVarComp('x', val=np.ones(mat.shape[1])))
comp = model.add_subsystem('comp',
MatMultComp(mat, approx_method=method))
comp.options['num_par_fd'] = num_par_fd
model.connect('indep.x', 'comp.x')
p.setup(mode='fwd', force_alloc_complex=(method == 'cs'))
p.run_model()
comp.num_computes = 0
J = p.compute_totals(of=['comp.y'], wrt=['indep.x'])
ncomputes = comp.num_computes if comp.comm.rank == 0 else 0
norm = np.linalg.norm(J['comp.y', 'indep.x'] - comp.mat)
self.assertLess(norm, 1.e-7)
if MPI:
self.assertEqual(MPI.COMM_WORLD.allreduce(ncomputes), size)
def test_fd_partials(self):
mat = np.arange(30, dtype=float).reshape(5, 6)
p = Problem()
model = p.model
model.add_subsystem('indep',
IndepVarComp('x', val=np.ones(mat.shape[1])))
comp = model.add_subsystem(
'comp', MatMultComp(mat, approx_method='fd', num_par_fd=3))
model.connect('indep.x', 'comp.x')
p.setup(mode='fwd')
p.run_model()
pre_count = comp.num_computes
# calling compute_totals will result in the computation of partials for comp
p.compute_totals(of=['comp.y'], wrt=['indep.x'])
# get the partial jacobian matrix
J = comp._jacobian['y', 'x']
post_count = comp.num_computes
# how many computes were used in this proc to compute the partial jacobian?
# Each proc should be doing 2 computes.
jac_count = post_count - pre_count
self.assertEqual(jac_count, 2)
# J and mat should be the same
self.assertLess(np.linalg.norm(J - mat), 1.e-7)
def test_fd_totals(self):
mat = np.arange(30, dtype=float).reshape(5, 6)
p = Problem(model=Group(num_par_fd=3))
model = p.model
model.approx_totals(method='fd')
model.add_subsystem('indep',
IndepVarComp('x', val=np.ones(mat.shape[1])))
comp = model.add_subsystem('comp', MatMultComp(mat))
model.connect('indep.x', 'comp.x')
p.setup(mode='fwd')
p.run_model()
pre_count = comp.num_computes
J = p.compute_totals(of=['comp.y'],
wrt=['indep.x'],
return_format='array')
post_count = comp.num_computes
# how many computes were used in this proc to compute the total jacobian?
# Each proc should be doing 2 computes.
jac_count = post_count - pre_count
self.assertEqual(jac_count, 2)
# J and mat should be the same
self.assertLess(np.linalg.norm(J - mat), 1.e-7)
def run_model(self, size, num_par_fd1, num_par_fd2, method, total=False):
if MPI:
if MPI.COMM_WORLD.rank == 0:
mat1 = np.random.random(5 * size).reshape((5, size)) - 0.5
else:
mat1 = None
mat1 = MPI.COMM_WORLD.bcast(mat1, root=0)
else:
mat1 = np.random.random(5 * size).reshape((5, size)) - 0.5
mat2 = mat1 * 5.0
p = Problem()
model = p.model
model.add_subsystem('indep',
IndepVarComp('x', val=np.ones(mat1.shape[1])))
par = model.add_subsystem('par', ParallelGroup())
if total:
meth = 'exact'
else:
meth = method
C1 = par.add_subsystem('C1', MatMultComp(mat1, approx_method=meth))
C2 = par.add_subsystem('C2', MatMultComp(mat2, approx_method=meth))
if total:
model.options['num_par_fd'] = num_par_fd1
model.approx_totals(method=method)
else:
C1.options['num_par_fd'] = num_par_fd1
C2.options['num_par_fd'] = num_par_fd2
model.connect('indep.x', 'par.C1.x')
model.connect('indep.x', 'par.C2.x')
p.setup(mode='fwd', force_alloc_complex=(method == 'cs'))
p.run_model()
J = p.compute_totals(of=['par.C1.y', 'par.C2.y'], wrt=['indep.x'])
norm = np.linalg.norm(J['par.C1.y', 'indep.x'] - C1.mat)
self.assertLess(norm, 1.e-7)
norm = np.linalg.norm(J['par.C2.y', 'indep.x'] - C2.mat)
self.assertLess(norm, 1.e-7)
def _setup_problem(mat, total_method='exact', partial_method='exact', total_num_par_fd=1,
partial_num_par_fd=1, approx_totals=False):
p = om.Problem(model=om.Group(num_par_fd=total_num_par_fd))
model = p.model
model.add_subsystem('indep', om.IndepVarComp('x', val=np.ones(mat.shape[1])))
model.add_subsystem('comp', MatMultComp(mat, approx_method=partial_method,
num_par_fd=partial_num_par_fd))
model.connect('indep.x', 'comp.x')
if approx_totals:
p.model.approx_totals()
p.setup(mode='fwd', force_alloc_complex='cs' in (total_method, partial_method))
return p
def test_par_fd(self, size, num_par_fd, method, solver):
if MPI:
if MPI.COMM_WORLD.rank == 0:
mat = np.random.random(5 * size).reshape((5, size)) - 0.5
else:
mat = None
mat = MPI.COMM_WORLD.bcast(mat, root=0)
else:
mat = np.random.random(5 * size).reshape((5, size)) - 0.5
p = om.Problem()
model = p.model
model.linear_solver = solver()
model.add_subsystem('indep',
om.IndepVarComp('x', val=np.ones(mat.shape[1])))
comp = model.add_subsystem(
'comp',
MatMultComp(mat, approx_method=method, num_par_fd=num_par_fd))
model.connect('indep.x', 'comp.x')
p.setup(mode='fwd', force_alloc_complex=(method == 'cs'))
p.run_model()
comp.num_computes = 0
J = p.compute_totals(of=['comp.y'], wrt=['indep.x'])
ncomputes = comp.num_computes if comp.comm.rank == 0 else 0
norm = np.linalg.norm(J['comp.y', 'indep.x'] - comp.mat)
self.assertLess(norm, 1.e-7)
if MPI:
self.assertEqual(MPI.COMM_WORLD.allreduce(ncomputes), size)
norm = np.linalg.norm(comp._jacobian['y', 'x'] - comp.mat)
self.assertLess(norm, 1.e-7)
# make sure check_partials works
data = p.check_partials(out_stream=None)
norm = np.linalg.norm(data['comp']['y', 'x']['J_fd'] - comp.mat)
self.assertLess(norm, 1.e-7)
def run_model(self, size, num_par_fd1, num_par_fd2, method, total=False):
if MPI:
if MPI.COMM_WORLD.rank == 0:
mat1 = np.random.random(5 * size).reshape((5, size)) - 0.5
else:
mat1 = None
mat1 = MPI.COMM_WORLD.bcast(mat1, root=0)
else:
mat1 = np.random.random(5 * size).reshape((5, size)) - 0.5
mat2 = mat1 * 5.0
if total:
grp = Group(num_par_fd=num_par_fd1)
else:
grp = Group()
p = Problem(model=grp)
model = p.model
model.add_subsystem('indep',
IndepVarComp('x', val=np.ones(mat1.shape[1])))
par = model.add_subsystem('par', ParallelGroup())
if total:
C1 = par.add_subsystem('C1',
MatMultComp(mat1, approx_method='exact'))
C2 = par.add_subsystem('C2',
MatMultComp(mat2, approx_method='exact'))
model.approx_totals(method=method)
else:
C1 = par.add_subsystem(
'C1',
MatMultComp(mat1, approx_method=method,
num_par_fd=num_par_fd1))
C2 = par.add_subsystem(
'C2',
MatMultComp(mat2, approx_method=method,
num_par_fd=num_par_fd2))
model.connect('indep.x', 'par.C1.x')
model.connect('indep.x', 'par.C2.x')
p.setup(mode='fwd', force_alloc_complex=(method == 'cs'))
p.run_model()
J = p.compute_totals(of=['par.C1.y', 'par.C2.y'], wrt=['indep.x'])
norm = np.linalg.norm(J['par.C1.y', 'indep.x'] - C1.mat)
self.assertLess(norm, 1.e-7)
norm = np.linalg.norm(J['par.C2.y', 'indep.x'] - C2.mat)
self.assertLess(norm, 1.e-7)
if not total:
# if total is True, the partials won't be computed during the compute_totals call
if C1 in par._subsystems_myproc:
norm = np.linalg.norm(C1._jacobian['y', 'x'] - C1.mat)
self.assertLess(norm, 1.e-7)
if C2 in par._subsystems_myproc:
norm = np.linalg.norm(C2._jacobian['y', 'x'] - C2.mat)
self.assertLess(norm, 1.e-7)
# make sure check_partials works
data = p.check_partials(out_stream=None)
if 'par.C1' in data:
norm = np.linalg.norm(data['par.C1']['y', 'x']['J_fd'] - C1.mat)
self.assertLess(norm, 1.e-7)
if 'par.C2' in data:
norm = np.linalg.norm(data['par.C2']['y', 'x']['J_fd'] - C2.mat)
self.assertLess(norm, 1.e-7)
