def create_calibrated_rtl(window_size, config, lattice_rank=4):
feature_columns = create_feature_columns(window_size)
feature_names = [fc.name for fc in feature_columns]
hparams = tfl.CalibratedRtlHParams(
feature_names=feature_names,
learning_rate=0.02,
calibration_l2_laplacian_reg=1.0e-4,
lattice_l2_laplacian_reg=1.0e-5,
lattice_l2_torsion_reg=1.0e-5,
optimizer=tf.train.AdamOptimizer,
num_lattices=1024,
interpolation_type="hypercube",
lattice_size=4,
lattice_rank=4,
missing_input_value=missing_val
)
hparams.set_feature_param("chr", "num_keypoints", 23)
hparams.set_feature_param("assay", "num_keypoints", 35)
hparams.set_feature_param("line", "num_keypoints", 51)
for offset in window(window_size):
hparams.set_feature_param(f"base_{offset}", "num_keypoints", len(base_symbols))
for i in range(1, num_cell_lines+1):
hparams.set_feature_param(f"C{i}", "num_keypoints", cell_line_keypoints)
for i in range(1, num_assays+1):
hparams.set_feature_param(f"M{i}", "num_keypoints", assay_keypoints)
return tfl.calibrated_rtl_regressor(
feature_columns=feature_columns,
model_dir=model_dir, config=config, hparams=hparams,
quantiles_dir=os.path.join(quantiles_dir, str(window_size))
)
def create_calibrated_rtl(feature_columns, config, quantiles_dir):
feature_names = [fc.name for fc in feature_columns]
hparams = tfl.CalibratedRtlHParams(feature_names=feature_names,
num_keypoints=200,
learning_rate=lr,
lattice_l2_laplacian_reg=5.0e-4,
lattice_l2_torsion_reg=1.0e-4,
lattice_size=2,
lattice_rank=4,
num_lattices=10)
return tfl.calibrated_rtl_classifier(feature_columns=feature_columns,
model_dir=config.model_dir,
config=config,
hparams=hparams,
quantiles_dir=quantiles_dir)
def create_calibrated_rtl(feature_columns, config, quantiles_dir):
feature_names = [fc.name for fc in feature_columns]
hparams = tfl.CalibratedRtlHParams(feature_names=feature_names,
num_keypoints=200,
learning_rate=0.02,
lattice_l2_laplacian_reg=5.0e-4,
lattice_l2_torsion_reg=1.0e-4,
lattice_size=3,
lattice_rank=4,
num_lattices=100)
#对部分列中的超参单独赋值.
hparams.set_feature_param("capital_gain", "lattice_size", 8)
hparams.set_feature_param("native_area", "lattice_size", 8)
hparams.set_feature_param("marital_status", "lattice_size", 4)
hparams.set_feature_param("age", "lattice_size", 8)
#hparams.parse(hparams)
_pprint_hparams(hparams)
return tfl.calibrated_rtl_classifier(feature_columns=feature_columns,
model_dir=config.model_dir,
config=config,
hparams=hparams,
quantiles_dir=quantiles_dir)
"""A quick test script for TensorFlow Lattice's calibrated RTL estimator."""
import numpy as np
import tensorflow as tf
import tensorflow_lattice as tfl
# Feature definition.
feature_columns = [
tf.feature_column.numeric_column('x0'),
tf.feature_column.numeric_column('x1'),
]
# Hyperparameters.
num_keypoints = 10
hparams = tfl.CalibratedRtlHParams(num_keypoints=num_keypoints,
num_lattices=5,
lattice_rank=2,
learning_rate=0.1)
def init_fn():
return tfl.uniform_keypoints_for_signal(num_keypoints,
input_min=-1.0,
input_max=1.0,
output_min=0.0,
output_max=1.0)
# Estimator.
rtl_estimator = tfl.calibrated_rtl_regressor(feature_columns=feature_columns,
hparams=hparams,
keypoints_initializers_fn=init_fn)
def run_image(image_path, lattice_size=35):
"""Reads image and fits a 2D lattice to compress it."""
im = plt.imread(image_path)[:, :, 2]
im_pixels = _pixels(im)
print('compression ratio is ', lattice_size**2 / float(im.size))
# Hyperparameters.
num_keypoints = 2
hparams = tfl.CalibratedRtlHParams(num_keypoints=num_keypoints,
num_lattices=1,
lattice_rank=2,
learning_rate=0.003,
lattice_size=lattice_size)
# Estimator.
# input: coordinate of the pixel
# output: value of the pixel
feature_columns = [
tf.feature_column.numeric_column('pixel_x'),
tf.feature_column.numeric_column('pixel_y'),
]
def keypoints_initializers():
return tfl.uniform_keypoints_for_signal(num_keypoints,
input_min=0.0,
input_max=im_pixels.max(),
output_min=0.0,
output_max=lattice_size - 1)
rtl_estimator = tfl.calibrated_rtl_regressor(
feature_columns=feature_columns,
hparams=hparams,
keypoints_initializers_fn=keypoints_initializers)
# Example input function.
input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(x={
'pixel_x':
im_pixels[:, 0],
'pixel_y':
im_pixels[:, 1]
},
y=im_pixels[:, 2],
batch_size=5000,
num_epochs=15,
shuffle=True)
# Train!
rtl_estimator.train(input_fn=input_fn)
# Evaluate!
eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
x={
'pixel_x': im_pixels[:, 0],
'pixel_y': im_pixels[:, 1]
},
y=im_pixels[:, 2],
batch_size=5000,
num_epochs=1,
shuffle=True)
print(rtl_estimator.evaluate(input_fn=eval_input_fn))
return rtl_estimator