def test_pearl_mutter_hvp_2x2(self, a_val, b_val, x_val, y_val, vector):
"""Test Hessian-vector product for a function with two variables."""
a_val = [a_val]
b_val = [b_val]
vector = np.array([vector], dtype=np.float32)
policy = HelperPolicy(n_vars=2)
params = policy.get_params()
x, y = params[0], params[1]
a = tf.constant(a_val)
b = tf.constant(b_val)
f = a * (x**2) + b * (y**2)
expected_hessian = compute_hessian(f, [x, y])
expected_hvp = tf.matmul(vector, expected_hessian)
reg_coeff = 1e-5
hvp = PearlmutterHvp()
self.sess.run(tf.compat.v1.global_variables_initializer())
self.sess.run(x.assign([x_val]))
self.sess.run(y.assign([y_val]))
hvp.update_hvp(f, policy, (a, b), reg_coeff)
hx = hvp.build_eval((np.array(a_val), np.array(b_val)))
hvp = hx(vector[0])
expected_hvp = expected_hvp.eval()
assert np.allclose(hvp, expected_hvp, atol=1e-6)
def __init__(self,
cg_iters=10,
reg_coeff=1e-5,
subsample_factor=1.,
backtrack_ratio=0.8,
max_backtracks=15,
debug_nan=False,
accept_violation=False,
hvp_approach=None,
num_slices=1):
self._cg_iters = cg_iters
self._reg_coeff = reg_coeff
self._subsample_factor = subsample_factor
self._backtrack_ratio = backtrack_ratio
self._max_backtracks = max_backtracks
self._num_slices = num_slices
self._opt_fun = None
self._target = None
self._max_constraint_val = None
self._constraint_name = None
self._debug_nan = debug_nan
self._accept_violation = accept_violation
if hvp_approach is None:
hvp_approach = PearlmutterHvp(num_slices)
self._hvp_approach = hvp_approach
def test_pearl_mutter_hvp_1x1(self):
"""Test Hessian-vector product for a function with one variable."""
policy = HelperPolicy(n_vars=1)
x = policy.get_params()[0]
a_val = np.array([5.0])
a = tf.constant([0.0])
f = a * (x**2)
expected_hessian = 2 * a_val
vector = np.array([10.0])
expected_hvp = expected_hessian * vector
reg_coeff = 1e-5
hvp = PearlmutterHvp()
self.sess.run(tf.compat.v1.global_variables_initializer())
hvp.update_hvp(f, policy, (a, ), reg_coeff)
hx = hvp.build_eval(np.array([a_val]))
computed_hvp = hx(vector)
assert np.allclose(computed_hvp, expected_hvp)
def test_pickleable(self):
policy = HelperPolicy(n_vars=1)
x = policy.get_params()[0]
a_val = np.array([5.0])
a = tf.constant([0.0])
f = a * (x**2)
vector = np.array([10.0])
reg_coeff = 1e-5
hvp = PearlmutterHvp()
self.sess.run(tf.compat.v1.global_variables_initializer())
hvp.update_hvp(f, policy, (a, ), reg_coeff)
hx = hvp.build_eval(np.array([a_val]))
before_pickle = hx(vector)
hvp = pickle.loads(pickle.dumps(hvp))
hvp.update_hvp(f, policy, (a, ), reg_coeff)
after_pickle = hx(vector)
assert np.equal(before_pickle, after_pickle)
