def test_unconstrained(self):
dimension = 5
x = variables.VariableV1(array_ops.zeros(dimension))
optimizer = external_optimizer.ScipyOptimizerInterface(
self._objective(x))
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
optimizer.minimize(sess)
self.assertAllClose(np.ones(dimension), sess.run(x))
def test_optimizer_kwargs(self):
# Checks that the 'method' argument is stil present
# after running optimizer.minimize().
# Bug reference: b/64065260
vector_initial_value = [7., 7.]
vector = variables.VariableV1(vector_initial_value, 'vector')
loss = math_ops.reduce_sum(math_ops.square(vector))
optimizer = external_optimizer.ScipyOptimizerInterface(loss,
method='SLSQP')
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
optimizer.minimize(sess)
method = optimizer.optimizer_kwargs.get('method')
self.assertEqual('SLSQP', method)
def test_vector_bounds(self):
vector_initial_value = [7., 7.]
vector = variables.VariableV1(vector_initial_value, 'vector')
# Make norm as small as possible.
loss = math_ops.reduce_sum(math_ops.square(vector))
var_to_bounds = {vector: ([None, 2.], None)}
optimizer = external_optimizer.ScipyOptimizerInterface(
loss, var_to_bounds=var_to_bounds)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
optimizer.minimize(sess)
self.assertAllClose([0., 2.], sess.run(vector))
def test_scalar_bounds(self):
vector_initial_value = [7., 7.]
vector = variables.VariableV1(vector_initial_value, 'vector')
# Make norm as small as possible.
loss = math_ops.reduce_sum(math_ops.square(vector))
# Make the minimum value of each component be 1.
var_to_bounds = {vector: (1., np.infty)}
optimizer = external_optimizer.ScipyOptimizerInterface(
loss, var_to_bounds=var_to_bounds)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
optimizer.minimize(sess)
self.assertAllClose(np.ones(2), sess.run(vector))
def _test_optimization_method(self,
method,
options,
rtol=1e-5,
atol=1e-5,
dimension=5):
x = variables.VariableV1(array_ops.zeros(dimension))
optimizer = external_optimizer.ScipyOptimizerInterface(
self._objective(x), method=method, options=options)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
optimizer.minimize(sess)
self.assertAllClose(np.ones(dimension),
sess.run(x),
rtol=rtol,
atol=atol)
def test_nonlinear_programming(self):
vector_initial_value = [7., 7.]
vector = variables.VariableV1(vector_initial_value, 'vector')
# Make norm as small as possible.
loss = math_ops.reduce_sum(math_ops.square(vector))
# Ensure y = 1.
equalities = [vector[1] - 1.]
# Ensure x >= 1. Thus optimum should be at (1, 1).
inequalities = [vector[0] - 1.]
optimizer = external_optimizer.ScipyOptimizerInterface(
loss,
equalities=equalities,
inequalities=inequalities,
method='SLSQP')
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
optimizer.minimize(sess)
self.assertAllClose(np.ones(2), sess.run(vector))
def test_callbacks(self):
vector_val = np.array([7., -2.], dtype=np.float32)
vector = variables.VariableV1(vector_val, 'vector')
minimum_location_val = np.arange(2)
minimum_location = constant_op.constant(minimum_location_val,
dtype=dtypes.float32)
loss = math_ops.reduce_sum(
math_ops.square(vector - minimum_location)) / 2.
loss_val_first = ((vector_val - minimum_location_val)**2).sum() / 2.
optimizer = external_optimizer.ScipyOptimizerInterface(loss,
method='SLSQP')
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
initial_vector_val = sess.run(vector)
extra_fetches = [loss]
step_callback = test.mock.Mock()
loss_callback = test.mock.Mock()
optimizer.minimize(sess,
fetches=extra_fetches,
loss_callback=loss_callback,
step_callback=step_callback)
loss_val_last = sess.run(loss)
call_first = test.mock.call(loss_val_first)
call_last = test.mock.call(loss_val_last)
loss_calls = [call_first, call_last]
loss_callback.assert_has_calls(loss_calls, any_order=True)
args, _ = step_callback.call_args
self.assertAllClose(minimum_location_val, args[0])