def _CalibratedLinearRegressorWithQuantiles(self, feature_names,
feature_columns, **hparams_args):
"""Model that saves/retrieves quantiles."""
# Quantiles to be used for x2
quantiles_dir = self.get_temp_dir()
keypoints_initialization.save_quantiles_for_keypoints(
input_fn=self._test_data.threed_input_fn(True),
save_dir=quantiles_dir,
feature_columns=feature_columns,
num_steps=1)
# Keypoint initialization function for x0 and x1
def init_fn():
return keypoints_initialization.uniform_keypoints_for_signal(
_NUM_KEYPOINTS, -1., 1., -1., 1.)
hparams = tfl_hparams.CalibratedLinearHParams(
feature_names, num_keypoints=_NUM_KEYPOINTS, **hparams_args)
return calibrated_linear.calibrated_linear_regressor(
feature_columns=feature_columns,
hparams=hparams,
keypoints_initializers_fn={
'x0': init_fn,
'x1': init_fn
},
quantiles_dir=quantiles_dir # Used for 'x2'
)
def _CalibratedLinearClassifier(self, feature_names, feature_columns,
**hparams_args):
def init_fn():
return keypoints_initialization.uniform_keypoints_for_signal(
_NUM_KEYPOINTS, -1., 1., -1., 1.)
hparams = tfl_hparams.CalibratedLinearHParams(
feature_names, num_keypoints=_NUM_KEYPOINTS, **hparams_args)
return calibrated_linear.calibrated_linear_classifier(
feature_columns=feature_columns,
hparams=hparams,
keypoints_initializers_fn=init_fn)
def testConstructorsAllTypes(self):
_ = hparams.CalibratedHParams(['x0', 'x1'])
_ = hparams.CalibratedLinearHParams(['x0', 'x1'], learning_rate=0.1)
_ = hparams.CalibratedLatticeHParams(['x0', 'x1'], learning_rate=0.1)
_ = hparams.CalibratedRtlHParams(['x0', 'x1'], learning_rate=0.1)
etl = hparams.CalibratedEtlHParams(['x0', 'x1'], learning_rate=0.1)
etl.parse('calibration_bound=yes')
self.assertTrue(etl.calibration_bound)
etl.parse('calibration_bound=off')
self.assertFalse(etl.calibration_bound)
with self.assertRaises(ValueError):
etl.parse('calibration_bound=foobar')
def __init__(self,
n_classes,
feature_columns=None,
model_dir=None,
quantiles_dir=None,
keypoints_initializers_fn=None,
optimizer=None,
config=None,
hparams=None):
"""Construct CalibrateLinearClassifier/Regressor."""
if not hparams:
hparams = tfl_hparams.CalibratedLinearHParams([])
self.check_hparams(hparams)
super(_CalibratedLinear,
self).__init__(n_classes, feature_columns, model_dir, quantiles_dir,
keypoints_initializers_fn, optimizer, config, hparams,
'linear')
def testAddNonExistingPerFeatureParam(self):
hp = hparams.CalibratedLinearHParams(['x0', 'x1'])
hp.set_feature_param('x0', 'calibration_l2_laplacian_reg', 0.1)
self.assertAlmostEqual(
hp.get_feature_param('x0', 'calibration_l2_laplacian_reg'), 0.1)
