def testMinimalRun(self):
x = basic.TrainableVariable(shape=(),
initializers={'w':
tf.ones_initializer()})()
x2 = x**2.0
min_value = 0.5
constr = optimization_constraints.OptimizationConstraints().add(
x > min_value)
self.assertFalse(constr._is_connected)
loss = moving_average.MovingAverage()(x2 + tf.random.normal(
(), stddev=1.0)) + constr()
self.assertTrue(constr._is_connected)
with self.assertRaisesRegexp(ValueError,
'Cannot add further constraints'):
constr.add(x > min_value)
with self.assertRaisesRegexp(ValueError,
'Cannot add further constraints'):
constr.add_geq(x, min_value)
with self.assertRaisesRegexp(ValueError,
'Cannot add further constraints'):
constr.add_leq(min_value < x)
opt = tf.train.AdamOptimizer(1e-2, beta1=0.0)
update = opt.minimize(loss)
with tf.control_dependencies([update]):
x2 = tf.identity(x2)
with tf.train.MonitoredSession() as sess:
for _ in range(500):
v, _ = sess.run([x2, update])
self.assertAllClose(v, min_value**2)
def testOpIdentification(self, mocked_add_leq, mocked_add_geq):
calls_to_add_leq = [0]
def mock_add_leq(*args, **kwargs):
del args
del kwargs
calls_to_add_leq[0] += 1
mocked_add_leq.side_effect = mock_add_leq
calls_to_add_geq = [0]
def mock_add_geq(*args, **kwargs):
del args
del kwargs
calls_to_add_geq[0] += 1
mocked_add_geq.side_effect = mock_add_geq
cons = optimization_constraints.OptimizationConstraints()
lhs = tf.zeros_like(1.0)
rhs = tf.ones_like(1.0)
self.assertEqual(calls_to_add_leq[0], 0)
self.assertEqual(calls_to_add_geq[0], 0)
cons.add(lhs < rhs)
self.assertEqual(calls_to_add_leq[0], 1)
self.assertEqual(calls_to_add_geq[0], 0)
cons.add(lhs <= rhs)
self.assertEqual(calls_to_add_leq[0], 2)
self.assertEqual(calls_to_add_geq[0], 0)
cons.add(lhs > rhs)
self.assertEqual(calls_to_add_geq[0], 1)
self.assertEqual(calls_to_add_leq[0], 2)
cons.add(lhs >= rhs)
self.assertEqual(calls_to_add_geq[0], 2)
self.assertEqual(calls_to_add_leq[0], 2)
def testValidRangeOverrides(self):
cons = optimization_constraints.OptimizationConstraints()
lhs = tf.zeros_like(1.0)
rhs = tf.ones_like(1.0)
valid_range = (1.0, 2.0)
cons.add(lhs < rhs, initializer=3.0, valid_range=valid_range)()
with tf.train.MonitoredSession() as sess:
lag_mul = sess.run(cons.lagrange_multipliers[0])
self.assertAllClose(lag_mul, valid_range[1])
def testLagrangeMultInit(self, initializer, exp_lag_mul):
cons = optimization_constraints.OptimizationConstraints()
lhs = tf.zeros_like(1.0)
rhs = tf.ones_like(1.0)
cons.add(lhs > rhs, initializer=initializer)()
l = cons.lagrange_multipliers[0]
with tf.train.MonitoredSession() as sess:
lag_mul = sess.run(l)
self.assertAllClose(lag_mul, exp_lag_mul)
def testRateOverrides(self, mocked_parametrized):
mocked_parametrized.side_effect = (
lambda x, rate: scale_gradient.scale_gradient(x, -rate))
rate = 7.3
cons = optimization_constraints.OptimizationConstraints()
lhs = tf.zeros_like(1.0)
rhs = tf.ones_like(1.0)
x = cons.add(lhs < rhs, rate=rate)()
v = tf.all_variables()[0]
dxdl = tf.gradients(x, v)
with tf.train.MonitoredSession() as sess:
grads = sess.run(dxdl)
self.assertAllClose(grads[0], rate)