def test_2darray_outputs(self):
mm = MetaModelUnStructuredComp()
mm.add_input('x', np.zeros((2, 2)))
mm.add_output('y', np.zeros((2, 2)))
mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob = Problem()
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
mm.options['train:x'] = [[[1.0, 1.0], [1.0, 1.0]],
[[2.0, 1.0], [1.0, 1.0]],
[[1.0, 2.0], [1.0, 1.0]],
[[1.0, 1.0], [2.0, 1.0]],
[[1.0, 1.0], [1.0, 2.0]]]
mm.options['train:y'] = [[[3.0, 1.0], [3.0, 1.0]],
[[2.0, 4.0], [2.0, 4.0]],
[[1.0, 7.0], [1.0, 7.0]],
[[6.0, -3.0], [6.0, -3.0]],
[[-2.0, 3.0], [-2.0, 3.0]]]
prob['mm.x'] = [[1.0, 2.0], [1.0, 1.0]]
prob.run_model()
assert_rel_error(self, prob['mm.y'], np.array([[1.0, 7.0],
[1.0, 7.0]]), .00001)
def test_metamodel_feature_vector(self):
# Like simple sine example, but with input of length n instead of scalar
# The expected behavior is that the output is also of length n, with
# each one being an independent prediction.
# Its as if you stamped out n copies of metamodel, ran n scalars
# through its input, then muxed all those outputs into one contiguous
# array but you skip all the n-copies thing and do it all as an array
import numpy as np
from openmdao.api import Problem, MetaModelUnStructuredComp, FloatKrigingSurrogate
size = 3
# create a vectorized MetaModelUnStructuredComp for sine
trig = MetaModelUnStructuredComp(
vec_size=size, default_surrogate=FloatKrigingSurrogate())
trig.add_input('x', np.zeros(size))
trig.add_output('y', np.zeros(size))
# add it to a Problem
prob = Problem()
prob.model.add_subsystem('trig', trig)
prob.setup(check=False)
# provide training data
trig.options['train:x'] = np.linspace(0, 10, 20)
trig.options['train:y'] = .5 * np.sin(trig.options['train:x'])
# train the surrogate and check predicted value
prob['trig.x'] = np.array([2.1, 3.2, 4.3])
prob.run_model()
assert_rel_error(self, prob['trig.y'],
np.array(.5 * np.sin(prob['trig.x'])), 1e-4)
def test_metamodel_feature_vector2d(self):
# similar to previous example, but processes 3 inputs/outputs at a time
import numpy as np
from openmdao.api import Problem, MetaModelUnStructuredComp, FloatKrigingSurrogate
size = 3
# create a vectorized MetaModelUnStructuredComp for sine and cosine
trig = MetaModelUnStructuredComp(
vec_size=size, default_surrogate=FloatKrigingSurrogate())
trig.add_input('x', np.zeros(size))
trig.add_output('y', np.zeros((size, 2)))
# add it to a Problem
prob = Problem()
prob.model.add_subsystem('trig', trig)
prob.setup(check=False)
# provide training data
trig.options['train:x'] = np.linspace(0, 10, 20)
trig.options['train:y'] = np.column_stack(
(.5 * np.sin(trig.options['train:x']),
.5 * np.cos(trig.options['train:x'])))
# train the surrogate and check predicted value
prob['trig.x'] = np.array([2.1, 3.2, 4.3])
prob.run_model()
assert_rel_error(
self, prob['trig.y'],
np.column_stack(
(.5 * np.sin(prob['trig.x']), .5 * np.cos(prob['trig.x']))),
1e-4)
def test_vectorized(self):
size = 3
# create a vectorized MetaModelUnStructuredComp for sine
trig = MetaModelUnStructuredComp(
vec_size=size, default_surrogate=FloatKrigingSurrogate())
trig.add_input('x', np.zeros(size))
trig.add_output('y', np.zeros(size))
# add it to a Problem
prob = Problem()
prob.model.add_subsystem('trig', trig)
prob.setup(check=False)
# provide training data
trig.options['train:x'] = np.linspace(0, 10, 20)
trig.options['train:y'] = .5 * np.sin(trig.options['train:x'])
# train the surrogate and check predicted value
prob['trig.x'] = np.array([2.1, 3.2, 4.3])
prob.run_model()
assert_rel_error(self, prob['trig.y'],
np.array(.5 * np.sin(prob['trig.x'])), 1e-4)
data = prob.check_partials(out_stream=None)
abs_errors = data['trig'][('y', 'x')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-6)
def test_vectorized_kriging(self):
# Test for coverage (handling the rmse)
size = 3
# create a vectorized MetaModelUnStructuredComp for sine
trig = MetaModelUnStructuredComp(
vec_size=size, default_surrogate=KrigingSurrogate(eval_rmse=True))
trig.add_input('x', np.zeros(size))
trig.add_output('y', np.zeros(size))
# add it to a Problem
prob = Problem()
prob.model.add_subsystem('trig', trig)
prob.setup(check=False)
# provide training data
trig.options['train:x'] = np.linspace(0, 10, 20)
trig.options['train:y'] = .5 * np.sin(trig.options['train:x'])
# train the surrogate and check predicted value
prob['trig.x'] = np.array([2.1, 3.2, 4.3])
prob.run_model()
assert_rel_error(self, prob['trig.y'],
np.array(.5 * np.sin(prob['trig.x'])), 1e-4)
self.assertEqual(len(prob.model.trig._metadata('y')['rmse']), 3)
def test_metamodel_feature2d(self):
# similar to previous example, but output is 2d
import numpy as np
from openmdao.api import Problem, MetaModelUnStructuredComp, FloatKrigingSurrogate
# create a MetaModelUnStructuredComp that predicts sine and cosine as an array
trig = MetaModelUnStructuredComp(
default_surrogate=FloatKrigingSurrogate())
trig.add_input('x', 0)
trig.add_output('y', np.zeros(2))
# add it to a Problem
prob = Problem()
prob.model.add_subsystem('trig', trig)
prob.setup(check=False)
# provide training data
trig.options['train:x'] = np.linspace(0, 10, 20)
trig.options['train:y'] = np.column_stack(
(.5 * np.sin(trig.options['train:x']),
.5 * np.cos(trig.options['train:x'])))
# train the surrogate and check predicted value
prob['trig.x'] = 2.1
prob.run_model()
assert_rel_error(
self, prob['trig.y'],
np.append(.5 * np.sin(prob['trig.x']),
.5 * np.cos(prob['trig.x'])), 1e-4)
def test_metamodel_feature(self):
# create a MetaModelUnStructuredComp, specifying surrogates for the outputs
import numpy as np
from openmdao.api import Problem, MetaModelUnStructuredComp, FloatKrigingSurrogate
trig = MetaModelUnStructuredComp()
x_train = np.linspace(0, 10, 20)
trig.add_input('x', 0., training_data=x_train)
trig.add_output('sin_x',
0.,
training_data=.5 * np.sin(x_train),
surrogate=FloatKrigingSurrogate())
trig.add_output('cos_x', 0., training_data=.5 * np.cos(x_train))
trig.options['default_surrogate'] = FloatKrigingSurrogate()
# add it to a Problem, run and check the predicted values
prob = Problem()
prob.model.add_subsystem('trig', trig)
prob.setup(check=False)
prob['trig.x'] = 2.1
prob.run_model()
assert_rel_error(self, prob['trig.sin_x'], .5 * np.sin(prob['trig.x']),
1e-4)
assert_rel_error(self, prob['trig.cos_x'], .5 * np.cos(prob['trig.x']),
1e-4)
def test_unequal_training_outputs(self):
mm = MetaModelUnStructuredComp()
mm.add_input('x', 0.)
mm.add_input('y', 0.)
mm.add_output('f', 0.)
mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob = Problem()
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
mm.options['train:x'] = [1.0, 1.0, 1.0, 1.0]
mm.options['train:y'] = [1.0, 2.0, 3.0, 4.0]
mm.options['train:f'] = [1.0, 1.0]
prob['mm.x'] = 1.0
prob['mm.y'] = 1.0
with self.assertRaises(RuntimeError) as cm:
prob.run_model()
expected = (
"MetaModelUnStructuredComp: Each variable must have the same number"
" of training points. Expected 4 but found"
" 2 points for 'f'.")
self.assertEqual(str(cm.exception), expected)
def test_sin_metamodel_rmse(self):
# create MetaModelUnStructuredComp with Kriging, using the rmse option
sin_mm = MetaModelUnStructuredComp()
sin_mm.add_input('x', 0.)
sin_mm.add_output('f_x', 0.)
sin_mm.options['default_surrogate'] = KrigingSurrogate(eval_rmse=True)
# add it to a Problem
prob = Problem()
prob.model.add_subsystem('sin_mm', sin_mm)
prob.setup(check=False)
# train the surrogate and check predicted value
sin_mm.options['train:x'] = np.linspace(0, 10, 20)
sin_mm.options['train:f_x'] = np.sin(sin_mm.options['train:x'])
prob['sin_mm.x'] = 2.1
prob.run_model()
assert_rel_error(self, prob['sin_mm.f_x'], np.sin(2.1), 1e-4) # mean
self.assertTrue(
self,
sin_mm._metadata('f_x')['rmse'] < 1e-5) # std deviation
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_metamodel_vector_errors(self):
# first dimension of all inputs/outputs must be 3
mm = MetaModelUnStructuredComp(vec_size=3)
with self.assertRaises(RuntimeError) as cm:
mm.add_input('x', np.zeros(2))
self.assertEqual(str(cm.exception),
"Metamodel: First dimension of input 'x' must be 3")
with self.assertRaises(RuntimeError) as cm:
mm.add_output('y', np.zeros(4))
self.assertEqual(str(cm.exception),
"Metamodel: First dimension of output 'y' must be 3")
def test_sin_metamodel(self):
# create a MetaModelUnStructuredComp for sine and add it to a Problem
sin_mm = MetaModelUnStructuredComp()
sin_mm.add_input('x', 0.)
sin_mm.add_output('f_x', 0.)
prob = Problem()
prob.model.add_subsystem('sin_mm', sin_mm)
# check that missing surrogate is detected in check_config
testlogger = TestLogger()
prob.setup(check=True, logger=testlogger)
# Conclude setup but don't run model.
prob.final_setup()
msg = ("No default surrogate model is defined and the "
"following outputs do not have a surrogate model:\n"
"['f_x']\n"
"Either specify a default_surrogate, or specify a "
"surrogate model for all outputs.")
self.assertEqual(len(testlogger.get('error')), 1)
self.assertTrue(msg in testlogger.get('error')[0])
# check that output with no specified surrogate gets the default
sin_mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob.setup(check=False)
surrogate = sin_mm._metadata('f_x').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate),
'sin_mm.f_x should get the default surrogate')
# check error message when no training data is provided
with self.assertRaises(RuntimeError) as cm:
prob.run_model()
msg = (
"MetaModelUnStructuredComp: The following training data sets must be "
"provided as options for sin_mm: ['train:x', 'train:f_x']")
self.assertEqual(str(cm.exception), msg)
# train the surrogate and check predicted value
sin_mm.options['train:x'] = np.linspace(0, 10, 20)
sin_mm.options['train:f_x'] = .5 * np.sin(sin_mm.options['train:x'])
prob['sin_mm.x'] = 2.1
prob.run_model()
assert_rel_error(self, prob['sin_mm.f_x'],
.5 * np.sin(prob['sin_mm.x']), 1e-4)
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_derivatives(self):
mm = MetaModelUnStructuredComp()
mm.add_input('x', 0.)
mm.add_output('f', 0.)
mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob = Problem()
prob.model.add_subsystem('p',
IndepVarComp('x', 0.),
promotes_outputs=['x'])
prob.model.add_subsystem('mm', mm, promotes_inputs=['x'])
prob.setup()
mm.options['train:x'] = [0., .25, .5, .75, 1.]
mm.options['train:f'] = [1., .75, .5, .25, 0.]
prob['x'] = 0.125
prob.run_model()
data = prob.check_partials(out_stream=None)
Jf = data['mm'][('f', 'x')]['J_fwd']
Jr = data['mm'][('f', 'x')]['J_rev']
assert_rel_error(self, Jf[0][0], -1., 1.e-3)
assert_rel_error(self, Jr[0][0], -1., 1.e-3)
abs_errors = data['mm'][('f', 'x')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-6)
# Complex step
prob.setup(force_alloc_complex=True)
prob.model.mm.set_check_partial_options(wrt='*', method='cs')
data = prob.check_partials(out_stream=None)
abs_errors = data['mm'][('f', 'x')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-6)
def test_sin_metamodel_preset_data(self):
# preset training data
x = np.linspace(0, 10, 200)
f_x = .5 * np.sin(x)
# create a MetaModelUnStructuredComp for Sin and add it to a Problem
sin_mm = MetaModelUnStructuredComp()
sin_mm.add_input('x', 0., training_data=x)
sin_mm.add_output('f_x', 0., training_data=f_x)
prob = Problem()
prob.model.add_subsystem('sin_mm', sin_mm)
# check that missing surrogate is detected in check_setup
testlogger = TestLogger()
prob.setup(check=True, logger=testlogger)
# Conclude setup but don't run model.
prob.final_setup()
msg = ("No default surrogate model is defined and the "
"following outputs do not have a surrogate model:\n"
"['f_x']\n"
"Either specify a default_surrogate, or specify a "
"surrogate model for all outputs.")
self.assertEqual(len(testlogger.get('error')), 1)
self.assertTrue(msg in testlogger.get('error')[0])
# check that output with no specified surrogate gets the default
sin_mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob.setup(check=False)
surrogate = sin_mm._metadata('f_x').get('surrogate')
self.assertTrue(isinstance(surrogate, FloatKrigingSurrogate),
'sin_mm.f_x should get the default surrogate')
prob['sin_mm.x'] = 2.22
prob.run_model()
assert_rel_error(self, prob['sin_mm.f_x'],
.5 * np.sin(prob['sin_mm.x']), 1e-4)
def test_error_no_surrogate(self):
# Seems like the error message from above should also be present and readable even if the
# user chooses to skip checking the model.
sin_mm = MetaModelUnStructuredComp()
sin_mm.add_input('x', 0.)
sin_mm.add_output('f_x', 0.)
prob = Problem()
prob.model.add_subsystem('sin_mm', sin_mm)
prob.setup(check=False)
sin_mm.options['train:x'] = np.linspace(0, 10, 20)
sin_mm.options['train:f_x'] = .5 * np.sin(sin_mm.options['train:x'])
with self.assertRaises(RuntimeError) as cm:
prob.run_model()
msg = ("Metamodel 'sin_mm': No surrogate specified for output 'f_x'")
self.assertEqual(str(cm.exception), msg)
def test_vector_inputs(self):
mm = MetaModelUnStructuredComp()
mm.add_input('x', np.zeros(4))
mm.add_output('y1', 0.)
mm.add_output('y2', 0.)
mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob = Problem()
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
mm.options['train:x'] = [[1.0, 1.0, 1.0, 1.0], [2.0, 1.0, 1.0, 1.0],
[1.0, 2.0, 1.0, 1.0], [1.0, 1.0, 2.0, 1.0],
[1.0, 1.0, 1.0, 2.0]]
mm.options['train:y1'] = [3.0, 2.0, 1.0, 6.0, -2.0]
mm.options['train:y2'] = [1.0, 4.0, 7.0, -3.0, 3.0]
prob['mm.x'] = [1.0, 2.0, 1.0, 1.0]
prob.run_model()
assert_rel_error(self, prob['mm.y1'], 1.0, .00001)
assert_rel_error(self, prob['mm.y2'], 7.0, .00001)
def test_derivatives_vectorized_multiD(self):
vec_size = 5
mm = MetaModelUnStructuredComp(vec_size=vec_size)
mm.add_input('x', np.zeros((vec_size, 2, 3)))
mm.add_input('xx', np.zeros((vec_size, 1)))
mm.add_output('y', np.zeros((vec_size, 4, 2)))
mm.options['default_surrogate'] = FloatKrigingSurrogate()
prob = Problem()
prob.model.add_subsystem('mm', mm)
prob.setup(check=False)
mm.options['train:x'] = [[[1.0, 2.0, 1.0], [1.0, 2.0, 1.0]],
[[2.0, 1.0, 1.0], [1.0, 1.0, 1.0]],
[[1.0, 1.0, 2.0], [1.0, 2.0, 1.0]],
[[1.0, 1.0, 1.0], [2.0, 1.0, 1.0]],
[[1.0, 2.0, 1.0], [1.0, 2.0, 2.0]]]
mm.options['train:xx'] = [1.0, 2.0, 1.0, 1.0, 2.0]
mm.options['train:y'] = [[[30.0, 10.0], [30.0, 25.0], [50.0, 10.7],
[15.0, 25.7]],
[[20.0, 40.0], [20.0, 40.0], [80.0, 30.3],
[12.0, 20.7]],
[[10.0, 70.0], [10.0, 70.0], [20.0, 10.9],
[13.0, 15.7]],
[[60.0, -30.0], [60.0, -30.0], [50.0, 50.5],
[14.0, 10.7]],
[[-20.0, 30.0], [-20.0, 30.0], [20.2, 10.0],
[15.0, 60.7]]]
prob['mm.x'] = [[[1.3, 1.3, 1.3], [1.5, 1.5, 1.5]],
[[1.4, 1.4, 1.4], [1.5, 1.5, 1.5]],
[[1.5, 1.5, 1.5], [1.5, 1.5, 1.5]],
[[1.5, 1.5, 1.5], [1.4, 1.4, 1.4]],
[[1.5, 1.5, 1.5], [1.3, 1.3, 1.3]]]
prob['mm.xx'] = [[1.4], [1.5], [1.6], [1.5], [1.4]]
prob.run_model()
data = prob.check_partials(out_stream=None)
abs_errors = data['mm'][('y', 'x')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-5)
abs_errors = data['mm'][('y', 'xx')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-5)
# Complex step
prob.setup(force_alloc_complex=True)
prob.model.mm.set_check_partial_options(wrt='*', method='cs')
data = prob.check_partials(out_stream=None)
abs_errors = data['mm'][('y', 'x')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-5)
abs_errors = data['mm'][('y', 'xx')]['abs error']
self.assertTrue(len(abs_errors) > 0)
for match in abs_errors:
abs_error = float(match)
self.assertTrue(abs_error < 1.e-5)