def _CalibratedEtlRegressor(self,
feature_names,
feature_columns,
weight_column=None,
**hparams_args):
def init_fn():
return keypoints_initialization.uniform_keypoints_for_signal(
_NUM_KEYPOINTS, -1., 1., 0., 1.)
hparams = tfl_hparams.CalibratedEtlHParams(
feature_names,
num_keypoints=_NUM_KEYPOINTS,
monotonic_num_lattices=1,
monotonic_lattice_rank=1,
monotonic_lattice_size=2,
non_monotonic_num_lattices=1,
non_monotonic_lattice_rank=1,
non_monotonic_lattice_size=2,
**hparams_args)
# Turn off monotonic calibrator.
hparams.set_param('calibration_monotonic', None)
hparams.set_param('learning_rate', 0.1)
return calibrated_etl.calibrated_etl_regressor(
feature_columns=feature_columns,
weight_column=weight_column,
hparams=hparams,
keypoints_initializers_fn=init_fn)
def testCalibratedEtlWithMissingTraining(self):
# x0 is missing with it's own vertex: so it can take very different values,
# while x1 is missing and calibrated, in this case to the middle of the
# lattice.
x0 = np.array([0., 0., 1., 1., -1., -1., 0., 1.])
x1 = np.array([0., 1., 0., 1., 0., 1., -1., -1.])
training_y = np.array([1., 3., 7., 11., 23., 27., 2., 9.])
x_samples = {'x0': x0, 'x1': x1}
train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
x=x_samples,
y=training_y,
batch_size=x0.shape[0],
num_epochs=2000,
shuffle=False)
test_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
x=x_samples, y=training_y, shuffle=False)
feature_columns = [
tf.feature_column.numeric_column('x0'),
tf.feature_column.numeric_column('x1'),
]
def init_fn():
return keypoints_initialization.uniform_keypoints_for_signal(
2, 0., 1., 0., 1.)
hparams = tfl_hparams.CalibratedEtlHParams(
['x0', 'x1'],
num_keypoints=2,
non_monotonic_num_lattices=5,
non_monotonic_lattice_rank=2,
non_monotonic_lattice_size=2,
learning_rate=0.1,
missing_input_value=-1.)
estimator = calibrated_etl.calibrated_etl_regressor(
feature_columns=feature_columns,
hparams=hparams,
keypoints_initializers_fn=init_fn)
estimator.train(input_fn=train_input_fn)
# Here we only check the successful evaluation.
# Checking the actual number, accuracy, etc, makes the test too flaky.
_ = estimator.evaluate(input_fn=test_input_fn)
def testCalibratedEtlWithMissingTraining(self):
# x0 is missing with it's own vertex: so it can take very different values,
# while x1 is missing and calibrated, in this case to the middle of the
# lattice.
x0 = np.array([0., 0., 1., 1., -1., -1., 0., 1.])
x1 = np.array([0., 1., 0., 1., 0., 1., -1., -1.])
training_y = np.array([1., 3., 7., 11., 23., 27., 2., 9.])
x_samples = {'x0': x0, 'x1': x1}
train_input_fn = numpy_io.numpy_input_fn(x=x_samples,
y=training_y,
batch_size=x0.shape[0],
num_epochs=2000,
shuffle=False)
test_input_fn = numpy_io.numpy_input_fn(x=x_samples,
y=training_y,
shuffle=False)
feature_columns = [
feature_column_lib.numeric_column('x0'),
feature_column_lib.numeric_column('x1'),
]
def init_fn():
return keypoints_initialization.uniform_keypoints_for_signal(
2, 0., 1., 0., 1.)
hparams = tfl_hparams.CalibratedEtlHParams(
['x0', 'x1'],
num_keypoints=2,
non_monotonic_num_lattices=5,
non_monotonic_lattice_rank=2,
non_monotonic_lattice_size=2,
learning_rate=0.1,
missing_input_value=-1.)
estimator = calibrated_etl.calibrated_etl_regressor(
feature_columns=feature_columns,
hparams=hparams,
keypoints_initializers_fn=init_fn)
estimator.train(input_fn=train_input_fn)
results = estimator.evaluate(input_fn=test_input_fn)
self.assertLess(results['average_loss'], 0.1)