def testTriangular2CyclicalLearningRate(self, serialize):
self.skipTest("Failing. See https://github.com/tensorflow/addons/issues/1203")
initial_learning_rate = 0.1
maximal_learning_rate = 1
step_size = 4000
step = tf.resource_variable_ops.ResourceVariable(0)
triangular2_cyclical_lr = cyclical_learning_rate.Triangular2CyclicalLearningRate(
initial_learning_rate=initial_learning_rate,
maximal_learning_rate=maximal_learning_rate,
step_size=step_size,
)
triangular2_cyclical_lr = _maybe_serialized(triangular2_cyclical_lr, serialize)
self.evaluate(tf.compat.v1.global_variables_initializer())
middle_learning_rate = (maximal_learning_rate + initial_learning_rate) / 2
expected = np.concatenate(
[
np.linspace(initial_learning_rate, maximal_learning_rate, num=2001)[1:],
np.linspace(maximal_learning_rate, initial_learning_rate, num=2001)[1:],
np.linspace(initial_learning_rate, middle_learning_rate, num=2001)[1:],
np.linspace(middle_learning_rate, initial_learning_rate, num=2001)[1:],
]
)
for expected_value in expected:
self.assertAllClose(
self.evaluate(triangular2_cyclical_lr(step)), expected_value, 1e-6
)
self.evaluate(step.assign_add(1))
def test_triangular2_cyclical_learning_rate(serialize):
initial_lr = 0.1
maximal_lr = 1
step_size = 30
triangular2_lr = cyclical_learning_rate.Triangular2CyclicalLearningRate(
initial_learning_rate=initial_lr,
maximal_learning_rate=maximal_lr,
step_size=step_size,
)
triangular2_lr = _maybe_serialized(triangular2_lr, serialize)
middle_lr = (maximal_lr + initial_lr) / 2
expected = np.concatenate([
np.linspace(initial_lr, maximal_lr, num=step_size + 1),
np.linspace(maximal_lr, initial_lr, num=step_size + 1)[1:],
np.linspace(initial_lr, middle_lr, num=step_size + 1)[1:],
np.linspace(middle_lr, initial_lr, num=step_size + 1)[1:],
])
for step, expected_value in enumerate(expected):
np.testing.assert_allclose(
triangular2_lr(step).numpy(), expected_value, 1e-6)