def testMonotonicity10dSinOfSum(self):
if self.disable_all:
return
monotonicities = [1] * 10
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 2,
"input_dims": 10,
"num_training_records": 1000,
"num_training_epoch": 100,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": self._SinOfSum,
"monotonicities": [1] * 10,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.179642, delta=self.loss_eps)
monotonicities = [0, 1, 0, 1, 1, 0, 1, 1, 1, 0]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config["monotonicities"] = monotonicities
config["kernel_initializer"] = kernel_initializer
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.181125, delta=self.loss_eps)
def testMonotonicityOneD(self):
if self.disable_all:
return
monotonicities = [1]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 20,
"input_dims": 1,
"num_training_records": 128,
"num_training_epoch": 50,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": self._SinPlusX,
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.123856, delta=self.loss_eps)
self._TestEnsemble(config)
monotonicities = ["increasing"]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 20,
"input_dims": 1,
"num_training_records": 100,
"num_training_epoch": 50,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": lambda x: -self._SinPlusX(x),
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 2.841356, delta=self.loss_eps)
self._TestEnsemble(config)
monotonicities = [1]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 5,
"input_dims": 1,
"num_training_records": 100,
"num_training_epoch": 200,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": self._SinPlusLargeX,
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
# Target function is strictly increasing.
"target_monotonicity_diff": 1e-6,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.000780, delta=self.loss_eps)
def testMonotonicity5d(self):
if self.disable_all:
return
config = {
"lattice_sizes": 2,
"input_dims": 5,
"num_training_records": 100,
"num_training_epoch": 200,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": self._ScaledSum,
"monotonicities": [1, 1, 1, 1, 1],
"kernel_initializer": keras.initializers.Constant(value=0.5),
# Function is strictly increasing everywhere, so request monotonicity
# diff to be strictly positive.
"target_monotonicity_diff": 0.08,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.000524, delta=self.loss_eps)
monotonicities = [1, 1, 1, 1, 1]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 2,
"input_dims": 5,
"num_training_records": 100,
"num_training_epoch": 40,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": lambda x: -self._ScaledSum(x),
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.015019, delta=self.loss_eps)
self._TestEnsemble(config)
monotonicities = [1, "increasing", 1, 1]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 3,
"input_dims": 4,
"num_training_records": 100,
"num_training_epoch": 100,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": self._SinOfSum,
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.377306, delta=self.loss_eps)
self._TestEnsemble(config)
def testInputOutOfBounds(self):
if self.disable_all:
return
config = {
"lattice_sizes": 6,
"input_dims": 1,
"num_training_records": 100,
"num_training_epoch": 20,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformlyExtendedRange,
"y_function": self._Sin,
"kernel_initializer": keras.initializers.Zeros(),
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.018726, delta=self.loss_eps)
self._TestEnsemble(config)
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=None, seed=self.seed)
config = {
"lattice_sizes": 2,
"input_dims": 2,
"num_training_records": 100,
"num_training_epoch": 20,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._TwoDMeshGridExtendedRange,
"y_function": self._SinOfSum,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.130816, delta=self.loss_eps)
self._TestEnsemble(config)
def testInitializerType(self, initializer, expected_loss):
if self.disable_all:
return
if initializer == "random_monotonic_initializer":
initializer = kfll.RandomMonotonicInitializer(monotonicities=None,
seed=self.seed)
config = {
"lattice_sizes": 3,
"input_dims": 2,
"num_training_records": 100,
"num_training_epoch": 0,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._TwoDMeshGrid,
"y_function": self._Max,
"kernel_initializer": initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, expected_loss, delta=self.loss_eps)
self._TestEnsemble(config)
def testHighDimensionsStressTest(self):
if self.disable_all:
return
monotonicities = [0] * 16
monotonicities[3], monotonicities[4], monotonicities[10] = (1, 1, 1)
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 2,
"input_dims": 16,
"num_terms": 128,
"units": 2,
"monotonicities": monotonicities,
"num_training_records": 100,
"num_training_epoch": 3,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._ScatterXUniformly,
"y_function": self._SinOfSum,
"kernel_initializer": kernel_initializer,
"target_monotonicity_diff": -1e-5,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.224262, delta=self.loss_eps)
def testMonotonicityTwoD(self):
if self.disable_all:
return
monotonicities = [1, 1]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 21,
"input_dims": 2,
"num_training_records": 900,
"num_training_epoch": 100,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._TwoDMeshGrid,
"y_function": self._SinPlusXNd,
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.562003, delta=self.loss_eps)
self._TestEnsemble(config)
monotonicities = ["none", "increasing"]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 21,
"input_dims": 2,
"num_training_records": 900,
"num_training_epoch": 100,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._TwoDMeshGrid,
"y_function": self._SinPlusXNd,
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.222727, delta=self.loss_eps)
self._TestEnsemble(config)
monotonicities = [1, 0]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 21,
"input_dims": 2,
"num_training_records": 900,
"num_training_epoch": 100,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.5,
"x_generator": self._TwoDMeshGrid,
"y_function": self._SinPlusXNd,
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.498311, delta=self.loss_eps)
self._TestEnsemble(config)
monotonicities = [1, 1]
kernel_initializer = kfll.RandomMonotonicInitializer(
monotonicities=monotonicities, seed=self.seed)
config = {
"lattice_sizes": 2,
"input_dims": 2,
"num_training_records": 100,
"num_training_epoch": 20,
"optimizer": tf.keras.optimizers.Adagrad,
"learning_rate": 1.0,
"x_generator": self._TwoDMeshGrid,
"y_function": lambda x: -self._ScaledSum(x),
"monotonicities": monotonicities,
"kernel_initializer": kernel_initializer,
} # pyformat: disable
loss = self._TrainModel(config)
self.assertAlmostEqual(loss, 0.050929, delta=self.loss_eps)
self._TestEnsemble(config)