def test_one_pt(self):
surrogate = ResponseSurface()
x = array([[0.0]])
y = array([[1.0]])
surrogate.train(x, y)
assert_rel_error(self, surrogate.betas, array([[1.0], [0.0], [0.0]]), 1e-9)
def test_one_pt(self):
surrogate = ResponseSurface()
x = array([[0.]])
y = array([[1.]])
surrogate.train(x, y)
assert_rel_error(self, surrogate.betas, array([[1.], [0.], [0.]]), 1e-9)
def test_vector_derivs(self):
surrogate = ResponseSurface()
x = array([[a, b] for a, b in itertools.product(linspace(0, 1, 10), repeat=2)])
y = array([[a + b, a - b] for a, b in x])
surrogate.train(x, y)
jac = surrogate.jacobian(array([[0.5, 0.5]]))
assert_rel_error(self, jac, array([[1, 1], [1, -1]]), 1e-5)
def test_no_training_data(self):
surrogate = ResponseSurface()
try:
surrogate.predict([0.0, 1.0])
except RuntimeError as err:
self.assertEqual(str(err), "ResponseSurface has not been trained, so no prediction can be made.")
else:
self.fail("RuntimeError Expected")
def test_1d_ill_conditioned(self):
# Test for least squares solver utilization when ill-conditioned
x = array([[case] for case in linspace(0.0, 1.0, 40)])
y = sin(x)
surrogate = ResponseSurface()
surrogate.train(x, y)
new_x = array([0.5])
mu = surrogate.predict(new_x)
assert_rel_error(self, mu, sin(0.5), 1e-3)
def test_1d_training(self):
x = array([[0.0], [2.0], [3.0]])
y = array([[branin_1d(case)] for case in x])
surrogate = ResponseSurface()
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_rel_error(self, mu, y0, 1e-9)
def test_scalar_derivs(self):
surrogate = ResponseSurface()
x = array([[0.0], [1.0], [2.0], [3.0]])
y = x.copy()
surrogate.train(x, y)
jac = surrogate.jacobian(array([[0.0]]))
assert_rel_error(self, jac[0][0], 1.0, 1e-3)
def test_1d_training(self):
x = array([[0.0], [2.0], [3.0]])
y = array([[branin_1d(case)] for case in x])
surrogate = ResponseSurface()
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_rel_error(self, mu, y0, 1e-9)
def test_scalar_derivs(self):
surrogate = ResponseSurface()
x = array([[0.], [1.], [2.], [3.]])
y = x.copy()
surrogate.train(x, y)
jac = surrogate.linearize(array([[0.]]))
assert_rel_error(self, jac[0][0], 1., 1e-3)
def test_vector_derivs(self):
surrogate = ResponseSurface()
x = array([[a, b] for a, b in
itertools.product(linspace(0, 1, 10), repeat=2)])
y = array([[a + b, a - b] for a, b in x])
surrogate.train(x, y)
jac = surrogate.linearize(array([[0.5, 0.5]]))
assert_rel_error(self, jac, array([[1, 1], [1, -1]]), 1e-5)
def test_1d_ill_conditioned(self):
# Test for least squares solver utilization when ill-conditioned
x = array([[case] for case in linspace(0., 1., 40)])
y = sin(x)
surrogate = ResponseSurface()
surrogate.train(x, y)
new_x = array([0.5])
mu = surrogate.predict(new_x)
assert_rel_error(self, mu, sin(0.5), 1e-3)
def test_scalar_derivs(self):
surrogate = ResponseSurface()
x = array([[0.], [1.], [2.], [3.]])
y = x.copy()
surrogate.train(x, y)
jac = surrogate.linearize(array([[0.]]))
assert_rel_error(self, jac[0][0], 1., 1e-3)
def test_no_training_data(self):
surrogate = ResponseSurface()
try:
surrogate.predict([0., 1.])
except RuntimeError as err:
self.assertEqual(str(err),
"ResponseSurface has not been trained, so no prediction can be made.")
else:
self.fail("RuntimeError Expected")
def test_vector_input(self):
surrogate = ResponseSurface()
x = array([[0.0, 0.0, 0.0], [1.0, 1.0, 1.0]])
y = array([[0.0], [3.0]])
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_rel_error(self, mu, y0, 1e-9)
def test_vector_input(self):
surrogate = ResponseSurface()
x = array([[0., 0., 0.], [1., 1., 1.]])
y = array([[0.], [3.]])
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_rel_error(self, mu, y0, 1e-9)
def test_1d_predictor(self):
x = array([[0.0], [2.0], [3.0], [4.0], [6.0]])
y = array([[branin_1d(case)] for case in x])
surrogate = ResponseSurface()
surrogate.train(x, y)
new_x = array([pi])
mu = surrogate.predict(new_x)
assert_rel_error(self, mu, 1.73114, 1e-4)
def test_vector_output(self):
surrogate = ResponseSurface()
x = array([[0.], [2.], [4.]])
y = array([[0., 0.], [1., 1.], [2., 0.]])
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_near_equal(mu, y0, 1e-9)
def test_1d_predictor(self):
x = array([[0.0], [2.0], [3.0], [4.0], [6.0]])
y = array([[branin_1d(case)] for case in x])
surrogate = ResponseSurface()
surrogate.train(x, y)
new_x = array([pi])
mu = surrogate.predict(new_x)
assert_rel_error(self, mu, 1.73114, 1e-4)
def test_2d(self):
x = array([[-2.0, 0.0], [-0.5, 1.5], [1.0, 1.0], [0.0, 0.25], [0.25, 0.0], [0.66, 0.33]])
y = array([[branin(case)] for case in x])
surrogate = ResponseSurface()
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_rel_error(self, mu, y0, 1e-9)
mu = surrogate.predict(array([0.5, 0.5]))
assert_rel_error(self, mu, branin([0.5, 0.5]), 1e-1)
def test_basics(self):
# create a metamodel component
mm = MetaModelUnStructuredComp()
mm.add_input('x1', 0.)
mm.add_input('x2', 0.)
mm.add_output('y1', 0.)
mm.add_output('y2', 0., surrogate=FloatKrigingSurrogate())
mm.options['default_surrogate'] = ResponseSurface()
# add metamodel to a problem
prob = Problem(model=Group())
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
# check that surrogates were properly assigned
surrogate = mm._metadata('y1').get('surrogate')
self.assertTrue(isinstance(surrogate, ResponseSurface))
surrogate = mm._metadata('y2').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate))
# populate training data
mm.options['train:x1'] = [1.0, 2.0, 3.0]
mm.options['train:x2'] = [1.0, 3.0, 4.0]
mm.options['train:y1'] = [3.0, 2.0, 1.0]
mm.options['train:y2'] = [1.0, 4.0, 7.0]
# run problem for provided data point and check prediction
prob['mm.x1'] = 2.0
prob['mm.x2'] = 3.0
self.assertTrue(mm.train) # training will occur before 1st run
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 2.0, .00001)
assert_rel_error(self, prob['mm.y2'], 4.0, .00001)
# run problem for interpolated data point and check prediction
prob['mm.x1'] = 2.5
prob['mm.x2'] = 3.5
self.assertFalse(mm.train) # training will not occur before 2nd run
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 1.5934, .001)
# change default surrogate, re-setup and check that metamodel re-trains
mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob.setup(check=False)
surrogate = mm._metadata('y1').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate))
self.assertTrue(mm.train) # training will occur after re-setup
def test_warning_bug(self):
# Make sure we don't warn that we are doing FD when the surrogate has analytic derivs.
x_train = np.arange(0., 10.)
y_train = np.arange(10., 20.)
z_train = x_train**2 + y_train**2
p = Problem()
p.model = m = Group()
params = IndepVarComp()
params.add_output('x', val=0.)
params.add_output('y', val=0.)
m.add_subsystem('params', params, promotes=['*'])
sm = MetaModelUnStructuredComp(default_surrogate=ResponseSurface())
sm.add_input('x', val=0.)
sm.add_input('y', val=0.)
sm.add_output('z', val=0.)
sm.options['train:x'] = x_train
sm.options['train:y'] = y_train
sm.options['train:z'] = z_train
# With or without the line below does not matter
# Only when method is set to fd, then RuntimeWarning disappears
sm.declare_partials('*', '*', method='exact')
m.add_subsystem('sm', sm, promotes=['*'])
m.add_design_var('x', lower=0., upper=10.)
m.add_design_var('y', lower=0., upper=10.)
m.add_objective('z')
p.setup(check=True)
stderr = sys.stderr
str_err = StringIO()
sys.stderr = str_err
try:
p.final_setup()
finally:
sys.stderr = stderr
output = str_err.getvalue()
self.assertTrue('finite difference' not in output)
def test_warm_start(self):
# create metamodel with warm_restart = True
mm = MetaModelUnStructured()
mm.add_input('x1', 0.)
mm.add_input('x2', 0.)
mm.add_output('y1', 0.)
mm.add_output('y2', 0.)
mm.default_surrogate = ResponseSurface()
mm.warm_restart = True
# add to problem
prob = Problem()
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
# provide initial training data
mm.metadata['train:x1'] = [1.0, 3.0]
mm.metadata['train:x2'] = [1.0, 4.0]
mm.metadata['train:y1'] = [3.0, 1.0]
mm.metadata['train:y2'] = [1.0, 7.0]
# run against a data point and check result
prob['mm.x1'] = 2.0
prob['mm.x2'] = 3.0
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 1.9085, .001)
assert_rel_error(self, prob['mm.y2'], 3.9203, .001)
# Add 3rd training point, moves the estimate for that point
# back to where it should be.
mm.metadata['train:x1'] = [2.0]
mm.metadata['train:x2'] = [3.0]
mm.metadata['train:y1'] = [2.0]
mm.metadata['train:y2'] = [4.0]
mm.train = True # currently need to tell meta to re-train
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 2.0, .00001)
assert_rel_error(self, prob['mm.y2'], 4.0, .00001)
def test_warm_start(self):
# create metamodel with warm_restart = True
meta = MetaModel()
meta.add_param('x1', 0.)
meta.add_param('x2', 0.)
meta.add_output('y1', 0.)
meta.add_output('y2', 0.)
meta.default_surrogate = ResponseSurface()
meta.warm_restart = True
# add to problem
prob = Problem(Group())
prob.root.add('meta', meta)
prob.setup(check=False)
# provide initial training data
prob['meta.train:x1'] = [1.0, 3.0]
prob['meta.train:x2'] = [1.0, 4.0]
prob['meta.train:y1'] = [3.0, 1.0]
prob['meta.train:y2'] = [1.0, 7.0]
# run against a data point and check result
prob['meta.x1'] = 2.0
prob['meta.x2'] = 3.0
prob.run()
assert_rel_error(self, prob['meta.y1'], 1.9085, .001)
assert_rel_error(self, prob['meta.y2'], 3.9203, .001)
# Add 3rd training point, moves the estimate for that point
# back to where it should be.
prob['meta.train:x1'] = [2.0]
prob['meta.train:x2'] = [3.0]
prob['meta.train:y1'] = [2.0]
prob['meta.train:y2'] = [4.0]
meta.train = True # currently need to tell meta to re-train
prob.run()
assert_rel_error(self, prob['meta.y1'], 2.0, .00001)
assert_rel_error(self, prob['meta.y2'], 4.0, .00001)
def test_2d(self):
x = array([[-2., 0.], [-0.5, 1.5], [1., 1.], [0., .25], [.25, 0.], [.66, .33]])
y = array([[branin(case)] for case in x])
surrogate = ResponseSurface()
surrogate.train(x, y)
for x0, y0 in zip(x, y):
mu = surrogate.predict(x0)
assert_rel_error(self, mu, y0, 1e-9)
mu = surrogate.predict(array([.5, .5]))
assert_rel_error(self, mu, branin([.5, .5]), 1e-1)
def test_metamodel_deprecated(self):
# run same test as above, only with the deprecated component,
# to ensure we get the warning and the correct answer.
# self-contained, to be removed when class name goes away.
from openmdao.components.meta_model_unstructured_comp import MetaModel # deprecated
msg = "'MetaModel' has been deprecated. Use 'MetaModelUnStructuredComp' instead."
with assert_warning(DeprecationWarning, msg):
mm = MetaModel()
mm.add_input('x1', 0.)
mm.add_input('x2', 0.)
mm.add_output('y1', 0.)
mm.add_output('y2', 0., surrogate=FloatKrigingSurrogate())
msg = "The 'default_surrogate' attribute provides backwards compatibility with " \
"earlier version of OpenMDAO; use options['default_surrogate'] instead."
with assert_warning(DeprecationWarning, msg):
mm.default_surrogate = ResponseSurface()
# add metamodel to a problem
prob = Problem(model=Group())
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
# check that surrogates were properly assigned
surrogate = mm._metadata('y1').get('surrogate')
self.assertTrue(isinstance(surrogate, ResponseSurface))
surrogate = mm._metadata('y2').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate))
# populate training data
msg = "The 'metadata' attribute provides backwards compatibility " \
"with earlier version of OpenMDAO; use 'options' instead."
with assert_warning(DeprecationWarning, msg):
mm.metadata['train:x1'] = [1.0, 2.0, 3.0]
mm.metadata['train:x2'] = [1.0, 3.0, 4.0]
mm.metadata['train:y1'] = [3.0, 2.0, 1.0]
mm.metadata['train:y2'] = [1.0, 4.0, 7.0]
# run problem for provided data point and check prediction
prob['mm.x1'] = 2.0
prob['mm.x2'] = 3.0
self.assertTrue(mm.train) # training will occur before 1st run
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 2.0, .00001)
assert_rel_error(self, prob['mm.y2'], 4.0, .00001)
# run problem for interpolated data point and check prediction
prob['mm.x1'] = 2.5
prob['mm.x2'] = 3.5
self.assertFalse(mm.train) # training will not occur before 2nd run
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 1.5934, .001)
# change default surrogate, re-setup and check that metamodel re-trains
msg = "The 'default_surrogate' attribute provides backwards compatibility with " \
"earlier version of OpenMDAO; use options['default_surrogate'] instead."
with assert_warning(DeprecationWarning, msg):
mm.default_surrogate = FloatKrigingSurrogate()
prob.setup(check=False)
surrogate = mm._metadata('y1').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate))
self.assertTrue(mm.train) # training will occur after re-setup
prob['mm.x1'] = 2.5
prob['mm.x2'] = 3.5
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 1.5, 1e-2)
def test_basics(self):
# create a metamodel component
mm = MetaModel()
mm.add_param('x1', 0.)
mm.add_param('x2', 0.)
mm.add_output('y1', 0.)
mm.add_output('y2', 0., surrogate=FloatKrigingSurrogate())
mm.default_surrogate = ResponseSurface()
# add metamodel to a problem
prob = Problem(root=Group())
prob.root.add('mm', mm)
prob.setup(check=False)
# check that surrogates were properly assigned
surrogate = prob.root.unknowns.metadata('mm.y1').get('surrogate')
self.assertTrue(isinstance(surrogate, ResponseSurface))
surrogate = prob.root.unknowns.metadata('mm.y2').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate))
# populate training data
prob['mm.train:x1'] = [1.0, 2.0, 3.0]
prob['mm.train:x2'] = [1.0, 3.0, 4.0]
prob['mm.train:y1'] = [3.0, 2.0, 1.0]
prob['mm.train:y2'] = [1.0, 4.0, 7.0]
# run problem for provided data point and check prediction
prob['mm.x1'] = 2.0
prob['mm.x2'] = 3.0
self.assertTrue(mm.train) # training will occur before 1st run
prob.run()
assert_rel_error(self, prob['mm.y1'], 2.0, .00001)
assert_rel_error(self, prob['mm.y2'], 4.0, .00001)
# run problem for interpolated data point and check prediction
prob['mm.x1'] = 2.5
prob['mm.x2'] = 3.5
self.assertFalse(mm.train) # training will not occur before 2nd run
prob.run()
assert_rel_error(self, prob['mm.y1'], 1.5934, .001)
# change default surrogate, re-setup and check that metamodel re-trains
mm.default_surrogate = FloatKrigingSurrogate()
prob.setup(check=False)
surrogate = prob.root.unknowns.metadata('mm.y1').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate))
self.assertTrue(mm.train) # training will occur after re-setup
mm.warm_restart = True # use existing training data
prob['mm.x1'] = 2.5
prob['mm.x2'] = 3.5
prob.run()
assert_rel_error(self, prob['mm.y1'], 2., 1e-2)
