def test_hessian_nodependence(self):
X = T.matrix()
cost = T.sum(X)
with warnings.catch_warnings(record=True) as w:
# The following should emit a warning
hess = tf.hessian(cost, X)
assert len(w) == 1
assert "unused input" in str(w[-1].message)
def test_hessian_nodependence(self):
X = T.matrix()
cost = T.sum(X)
with warnings.catch_warnings(record=True) as w:
# The following should emit a warning
hess = tf.hessian(cost, X)
assert len(w) == 1
assert "unused input" in str(w[-1].message)
def test_hessian_no_Rop(self):
# Break the Rop in T.exp
def new_Rop(x, y):
raise NotImplementedError
T.exp.R_op = new_Rop
# Rebuild graph to force recompile
X = T.vector()
cost = T.exp(T.sum(X**2))
# And check that all is still well
hess = tf.hessian(cost, X)
np_testing.assert_allclose(self.correct_hess, hess(self.Y, self.A))
def test_hessian_no_Rop(self):
# Break the Rop in T.exp
def new_Rop(x, y):
raise NotImplementedError
T.exp.R_op = new_Rop
# Rebuild graph to force recompile
X = T.vector()
cost = T.exp(T.sum(X**2))
# And check that all is still well
hess = tf.hessian(cost, X)
np_testing.assert_allclose(self.correct_hess, hess(self.Y, self.A))
def compile(problem, need_grad, need_hess):
# Conditionally load autodiff backend if needed
if (problem.cost is None or
(need_grad and problem.grad is None and problem.egrad is None) or
(need_hess and problem.hess is None and problem.ehess is None)):
if type(problem.ad_cost).__name__ == 'TensorVariable':
from pymanopt.tools.autodiff import _theano as ad
elif type(problem.ad_cost).__name__ == 'function':
from pymanopt.tools.autodiff import _autograd as ad
else:
warn('Cannot identify autodiff backend from '
'ad_cost variable type.')
if problem.cost is None:
problem.cost = ad.compile(problem.ad_cost, problem.ad_arg)
if need_grad and problem.egrad is None and problem.grad is None:
problem.egrad = ad.gradient(problem.ad_cost, problem.ad_arg)
# Assume if hessian is needed gradient is as well
if need_hess and problem.ehess is None and problem.hess is None:
problem.ehess = ad.hessian(problem.ad_cost, problem.ad_arg)
def test_hessian(self):
hess = tf.hessian(self.cost, self.X)
# Now test hess
np_testing.assert_allclose(self.correct_hess, hess(self.Y, self.A))
def test_hessian(self):
hess = tf.hessian(self.cost, self.X)
# Now test hess
np_testing.assert_allclose(self.correct_hess, hess(self.Y, self.A))