def create_calibrated_lattice(feature_columns, config, quantiles_dir):
"""Creates a calibrated Lattice estimator."""
feature_names = [fc.name for fc in feature_columns]
hparams = tfl.CalibratedLatticeHParams(
feature_names=feature_names,
learning_rate=FLAGS.learning_rate,
lattice_l2_laplacian_reg=5.0e-4,
lattice_l2_torsion_reg=1.0e-4,
interpolation_type='hypercube',
num_keypoints=8,
feature__outerIters__monotonicity=1,
feature__innerIters__monotonicity=1,
# feature__bitsPerCycle__monotonicity=-1,
# feature__innerIters__lattice_size=3,
# feature__outerIters__lattice_size=3,
# monotonicity=frozenset({'outerIters': 1}.items),
lattice_size=3,
lattice_rank=len(feature_columns),
num_lattices=FLAGS.num_lattices,
optimizer=tf.train.AdamOptimizer)
# Specific feature parameters.
hparams.parse(FLAGS.hparams)
_pprint_hparams(hparams)
return tfl.calibrated_lattice_regressor(feature_columns=feature_columns,
model_dir=config.model_dir,
config=config,
quantiles_dir=quantiles_dir,
keypoints_initializers_fn=None,
optimizer=tf.train.AdamOptimizer,
hparams=hparams)
def create_calibrated_lattice(feature_columns, config, quantiles_dir):
feature_names = [fc.name for fc in feature_columns]
hparams = tfl.CalibratedLatticeHParams(feature_names=feature_names,
num_keypoints=200,
lattice_l2_laplacian_reg=5e-4,
lattice_l2_torsion_reg=1e-4,
learning_rate=lr,
lattice_size=2)
hparams.set_feature_param("feature1", "monotonicity", 1)
return tfl.calibrated_lattice_classifier(feature_columns=feature_columns,
model_dir=config.model_dir,
config=config,
hparams=hparams,
quantiles_dir=quantiles_dir)
def create_calibrated_lattice(feature_columns, config, quantiles_dir):
"""Creates a calibrated lattice estimator."""
feature_names = [fc.name for fc in feature_columns]
hparams = tfl.CalibratedLatticeHParams(feature_names=feature_names,
num_keypoints=200,
lattice_l2_laplacian_reg=5.0e-3,
lattice_l2_torsion_reg=1.0e-4,
learning_rate=0.1,
lattice_size=2)
hparams.parse(FLAGS.hparams)
_pprint_hparams(hparams)
return tfl.calibrated_lattice_classifier(feature_columns=feature_columns,
model_dir=config.model_dir,
config=config,
hparams=hparams,
quantiles_dir=quantiles_dir)
test_features = {
'distance': np.array([5.0, 10.0]),
'quality': np.array([3.0, 3.0]),
}
# Feature definition.
feature_columns = [
tf.feature_column.numeric_column('distance'),
tf.feature_column.numeric_column('quality'),
]
# Hyperparameters.
num_keypoints = 10
hparams = tfl.CalibratedLatticeHParams(
feature_names=['distance', 'quality'],
num_keypoints=num_keypoints,
learning_rate=0.1,
)
# Set feature monotonicity.
hparams.set_feature_param('distance', 'monotonicity', -1)
hparams.set_feature_param('quality', 'monotonicity', +1)
# Define keypoint init.
keypoints_init_fns = {
'distance':
lambda: tfl.uniform_keypoints_for_signal(num_keypoints,
input_min=0.0,
input_max=10.0,
output_min=0.0,
output_max=1.0),
{"x": x_train}, y_train, batch_size=batch_size, num_epochs=1000, shuffle=False)
# ====================================
# 訓練(トレーニング)
# ====================================
# 機能のリストを宣言する。 1つの数値機能しかありません。 より複雑で有用な他の多くのタイプの列があります。
feature_columns = [
tf.feature_column.numeric_column("x")
]
# Hyperparameters.
num_keypoints = 20
hparams = tfl.CalibratedLatticeHParams(
feature_names=['x'],
num_keypoints=num_keypoints,
learning_rate=0.1,
lattice_rank=2
)
# Set feature monotonicity.
hparams.set_feature_param('x', 'monotonicity', -1)
# Define keypoint init.
keypoints_init_fns = {
'x': lambda: tfl.uniform_keypoints_for_signal(num_keypoints,
input_min=-5.0,
input_max=5.0,
output_min=0.0,
output_max=25.0),
}