def setUp(self):
super(CannedEstimatorsTest, self).setUp()
self.eps = 0.001
# UCI Statlog (Heart) dataset.
heart_csv_file = tf.keras.utils.get_file(
'heart.csv',
'http://storage.googleapis.com/download.tensorflow.org/data/heart.csv')
heart_df = pd.read_csv(heart_csv_file)
heart_target = heart_df.pop('target')
heart_train_size = int(len(heart_df) * 0.8)
self.heart_train_x = heart_df[:heart_train_size]
self.heart_train_y = heart_target[:heart_train_size]
self.heart_test_x = heart_df[heart_train_size:]
self.heart_test_y = heart_target[heart_train_size:]
# Feature columns.
# - age
# - sex
# - cp chest pain type (4 values)
# - trestbps resting blood pressure
# - chol serum cholestoral in mg/dl
# - fbs fasting blood sugar > 120 mg/dl
# - restecg resting electrocardiographic results (values 0,1,2)
# - thalach maximum heart rate achieved
# - exang exercise induced angina
# - oldpeak ST depression induced by exercise relative to rest
# - slope the slope of the peak exercise ST segment
# - ca number of major vessels (0-3) colored by flourosopy
# - thal 3 = normal; 6 = fixed defect; 7 = reversable defect
self.heart_feature_columns = [
fc.numeric_column('age', default_value=-1),
fc.categorical_column_with_vocabulary_list('sex', [0, 1]),
fc.numeric_column('cp'),
fc.numeric_column('trestbps', default_value=-1),
fc.numeric_column('chol'),
fc.categorical_column_with_vocabulary_list('fbs', [0, 1]),
fc.categorical_column_with_vocabulary_list('restecg', [0, 1, 2]),
fc.numeric_column('thalach'),
fc.categorical_column_with_vocabulary_list('exang', [0, 1]),
fc.numeric_column('oldpeak'),
fc.categorical_column_with_vocabulary_list('slope', [0, 1, 2]),
fc.numeric_column('ca'),
fc.categorical_column_with_vocabulary_list(
'thal', ['normal', 'fixed', 'reversible']),
]
# Feature configs. Each model can pick and choose which features to use.
self.heart_feature_configs = [
configs.FeatureConfig(
name='age',
lattice_size=3,
pwl_calibration_num_keypoints=5,
monotonicity=1,
pwl_calibration_clip_max=100,
),
configs.FeatureConfig(
name='cp',
pwl_calibration_num_keypoints=4,
pwl_calibration_input_keypoints='uniform',
monotonicity='increasing',
),
configs.FeatureConfig(
name='chol',
pwl_calibration_input_keypoints=[126.0, 210.0, 247.0, 286.0, 564.0],
monotonicity=1,
pwl_calibration_clip_min=130,
pwl_calibration_clamp_min=True,
pwl_calibration_clamp_max=True,
regularizer_configs=[
configs.RegularizerConfig(name='calib_hessian', l2=1e-4),
],
),
configs.FeatureConfig(
name='fbs',
monotonicity=[(0, 1)],
),
configs.FeatureConfig(
name='trestbps',
pwl_calibration_num_keypoints=5,
monotonicity='decreasing',
),
configs.FeatureConfig(
name='thalach',
pwl_calibration_num_keypoints=5,
monotonicity=-1,
),
configs.FeatureConfig(
name='restecg',
monotonicity=[(0, 1), (0, 2)],
),
configs.FeatureConfig(
name='exang',
monotonicity=[(0, 1)],
),
configs.FeatureConfig(
name='oldpeak',
pwl_calibration_num_keypoints=5,
monotonicity=1,
),
configs.FeatureConfig(
name='slope',
monotonicity=[(0, 1), (1, 2)],
),
configs.FeatureConfig(
name='ca',
pwl_calibration_num_keypoints=4,
monotonicity='increasing',
),
configs.FeatureConfig(
name='thal',
monotonicity=[('normal', 'fixed'), ('normal', 'reversible')],
),
]
# UCI Boston dataset.
boston_dataset = load_boston()
boston_df = pd.DataFrame(
boston_dataset.data, columns=boston_dataset.feature_names)
boston_df['CHAS'] = boston_df['CHAS'].astype(np.int32)
boston_target = pd.Series(boston_dataset.target)
boston_train_size = int(len(boston_df) * 0.8)
self.boston_train_x = boston_df[:boston_train_size]
self.boston_train_y = boston_target[:boston_train_size]
self.boston_test_x = boston_df[boston_train_size:]
self.boston_test_y = boston_target[boston_train_size:]
# Feature columns.
# - CRIM per capita crime rate by town
# - ZN proportion of residential land zoned for lots over 25,000 sq.ft
# - INDUS proportion of non-retail business acres per town
# - CHAS Charles River dummy variable (= 1 if tract bounds river)
# - NOX nitric oxides concentration (parts per 10 million)
# - RM average number of rooms per dwelling
# - AGE proportion of owner-occupied units built prior to 1940
# - DIS weighted distances to five Boston employment centres
# - RAD index of accessibility to radial highways
# - TAX full-value property-tax rate per $10,000
# - PTRATIO pupil-teacher ratio by town
# - B 1000(Bk - 0.63)^2 where Bk is the proportion of blacks by town
# - LSTAT % lower status of the population
# - Target Median value of owner-occupied homes in $1000's
self.boston_feature_columns = [
fc.numeric_column('CRIM'),
fc.numeric_column('ZN'),
fc.numeric_column('INDUS'),
fc.categorical_column_with_vocabulary_list('CHAS', [0, 1]),
fc.numeric_column('NOX'),
fc.numeric_column('RM'),
fc.numeric_column('AGE'),
fc.numeric_column('DIS'),
fc.numeric_column('RAD'),
fc.numeric_column('TAX'),
fc.numeric_column('PTRATIO'),
fc.numeric_column('B'),
fc.numeric_column('LSTAT'),
]
# Feature configs. Each model can pick and choose which features to use.
self.boston_feature_configs = [
configs.FeatureConfig(
name='CRIM',
lattice_size=3,
monotonicity=-1,
pwl_calibration_convexity=1,
),
configs.FeatureConfig(
name='ZN',
pwl_calibration_input_keypoints=[0.0, 25.0, 50.0, 75.0, 100.0],
monotonicity=1,
reflects_trust_in=[
configs.TrustConfig(feature_name='RM', trust_type='trapezoid'),
],
),
configs.FeatureConfig(
name='INDUS',
pwl_calibration_input_keypoints='uniform',
pwl_calibration_always_monotonic=False,
reflects_trust_in=[
configs.TrustConfig(
feature_name='RM',
trust_type='edgeworth',
direction='negative'),
],
regularizer_configs=[
configs.RegularizerConfig(name='calib_wrinkle', l2=1e-4),
],
),
configs.FeatureConfig(name='CHAS',),
configs.FeatureConfig(name='NOX',),
configs.FeatureConfig(
name='RM',
monotonicity='increasing',
pwl_calibration_convexity='concave',
),
configs.FeatureConfig(
name='AGE',
monotonicity=-1,
),
configs.FeatureConfig(
name='DIS',
lattice_size=3,
unimodality=1,
),
configs.FeatureConfig(name='RAD',),
configs.FeatureConfig(name='TAX',),
configs.FeatureConfig(
name='PTRATIO',
monotonicity='decreasing',
),
configs.FeatureConfig(name='B',),
configs.FeatureConfig(
name='LSTAT',
monotonicity=-1,
dominates=[
configs.DominanceConfig(
feature_name='AGE', dominance_type='monotonic'),
],
),
]
def setUp(self):
super(PremadeTest, self).setUp()
# UCI Statlog (Heart) dataset.
heart_csv_file = tf.keras.utils.get_file(
'heart.csv', 'http://storage.googleapis.com/applied-dl/heart.csv')
heart_df = pd.read_csv(heart_csv_file)
heart_train_size = int(len(heart_df) * 0.8)
heart_train_dataframe = heart_df[:heart_train_size]
heart_test_dataframe = heart_df[heart_train_size:]
# Features:
# - age
# - sex
# - cp chest pain type (4 values)
# - trestbps resting blood pressure
# - chol serum cholestoral in mg/dl
# - fbs fasting blood sugar > 120 mg/dl
# - restecg resting electrocardiographic results (values 0,1,2)
# - thalach maximum heart rate achieved
# - exang exercise induced angina
# - oldpeak ST depression induced by exercise relative to rest
# - slope the slope of the peak exercise ST segment
# - ca number of major vessels (0-3) colored by flourosopy
# - thal 3 = normal; 6 = fixed defect; 7 = reversable defect
#
# This ordering of feature names will be the exact same order that we
# construct our model to expect.
self.heart_feature_names = [
'age', 'sex', 'cp', 'chol', 'fbs', 'trestbps', 'thalach', 'restecg',
'exang', 'oldpeak', 'slope', 'ca', 'thal'
]
feature_name_indices = {
name: index for index, name in enumerate(self.heart_feature_names)
}
# This is the vocab list and mapping we will use for the 'thal' categorical
# feature.
thal_vocab_list = ['normal', 'fixed', 'reversible']
thal_map = {category: i for i, category in enumerate(thal_vocab_list)}
# Custom function for converting thal categories to buckets
def convert_thal_features(thal_features):
# Note that two examples in the test set are already converted.
return np.array([
thal_map[feature] if feature in thal_vocab_list else feature
for feature in thal_features
])
# Custom function for extracting each feature.
def extract_features(dataframe, label_name='target'):
features = []
for feature_name in self.heart_feature_names:
if feature_name == 'thal':
features.append(
convert_thal_features(
dataframe[feature_name].values).astype(float))
else:
features.append(dataframe[feature_name].values.astype(float))
labels = dataframe[label_name].values.astype(float)
return features, labels
self.heart_train_x, self.heart_train_y = extract_features(
heart_train_dataframe)
self.heart_test_x, self.heart_test_y = extract_features(
heart_test_dataframe)
# Let's define our label minimum and maximum.
self.heart_min_label = float(np.min(self.heart_train_y))
self.heart_max_label = float(np.max(self.heart_train_y))
# Our lattice models may have predictions above 1.0 due to numerical errors.
# We can subtract this small epsilon value from our output_max to make sure
# we do not predict values outside of our label bound.
self.numerical_error_epsilon = 1e-5
def compute_quantiles(features,
num_keypoints=10,
clip_min=None,
clip_max=None,
missing_value=None):
# Clip min and max if desired.
if clip_min is not None:
features = np.maximum(features, clip_min)
features = np.append(features, clip_min)
if clip_max is not None:
features = np.minimum(features, clip_max)
features = np.append(features, clip_max)
# Make features unique.
unique_features = np.unique(features)
# Remove missing values if specified.
if missing_value is not None:
unique_features = np.delete(unique_features,
np.where(unique_features == missing_value))
# Compute and return quantiles over unique non-missing feature values.
return np.quantile(
unique_features,
np.linspace(0., 1., num=num_keypoints),
interpolation='nearest').astype(float)
self.heart_feature_configs = [
configs.FeatureConfig(
name='age',
lattice_size=3,
monotonicity='increasing',
# We must set the keypoints manually.
pwl_calibration_num_keypoints=5,
pwl_calibration_input_keypoints=compute_quantiles(
self.heart_train_x[feature_name_indices['age']],
num_keypoints=5,
clip_max=100),
# Per feature regularization.
regularizer_configs=[
configs.RegularizerConfig(name='calib_wrinkle', l2=0.1),
],
),
configs.FeatureConfig(
name='sex',
num_buckets=2,
),
configs.FeatureConfig(
name='cp',
monotonicity='increasing',
# Keypoints that are uniformly spaced.
pwl_calibration_num_keypoints=4,
pwl_calibration_input_keypoints=np.linspace(
np.min(self.heart_train_x[feature_name_indices['cp']]),
np.max(self.heart_train_x[feature_name_indices['cp']]),
num=4),
),
configs.FeatureConfig(
name='chol',
monotonicity='increasing',
# Explicit input keypoints initialization.
pwl_calibration_input_keypoints=[126.0, 210.0, 247.0, 286.0, 564.0],
# Calibration can be forced to span the full output range
# by clamping.
pwl_calibration_clamp_min=True,
pwl_calibration_clamp_max=True,
# Per feature regularization.
regularizer_configs=[
configs.RegularizerConfig(name='calib_hessian', l2=1e-4),
],
),
configs.FeatureConfig(
name='fbs',
# Partial monotonicity: output(0) <= output(1)
monotonicity=[(0, 1)],
num_buckets=2,
),
configs.FeatureConfig(
name='trestbps',
monotonicity='decreasing',
pwl_calibration_num_keypoints=5,
pwl_calibration_input_keypoints=compute_quantiles(
self.heart_train_x[feature_name_indices['trestbps']],
num_keypoints=5),
),
configs.FeatureConfig(
name='thalach',
monotonicity='decreasing',
pwl_calibration_num_keypoints=5,
pwl_calibration_input_keypoints=compute_quantiles(
self.heart_train_x[feature_name_indices['thalach']],
num_keypoints=5),
),
configs.FeatureConfig(
name='restecg',
# Partial monotonicity:
# output(0) <= output(1), output(0) <= output(2)
monotonicity=[(0, 1), (0, 2)],
num_buckets=3,
),
configs.FeatureConfig(
name='exang',
# Partial monotonicity: output(0) <= output(1)
monotonicity=[(0, 1)],
num_buckets=2,
),
configs.FeatureConfig(
name='oldpeak',
monotonicity='increasing',
pwl_calibration_num_keypoints=5,
pwl_calibration_input_keypoints=compute_quantiles(
self.heart_train_x[feature_name_indices['oldpeak']],
num_keypoints=5),
),
configs.FeatureConfig(
name='slope',
# Partial monotonicity:
# output(0) <= output(1), output(1) <= output(2)
monotonicity=[(0, 1), (1, 2)],
num_buckets=3,
),
configs.FeatureConfig(
name='ca',
monotonicity='increasing',
pwl_calibration_num_keypoints=4,
pwl_calibration_input_keypoints=compute_quantiles(
self.heart_train_x[feature_name_indices['ca']],
num_keypoints=4),
),
configs.FeatureConfig(
name='thal',
# Partial monotonicity:
# output(normal) <= output(fixed)
# output(normal) <= output(reversible)
monotonicity=[('normal', 'fixed'), ('normal', 'reversible')],
num_buckets=3,
# We must specify the vocabulary list in order to later set the
# monotonicities since we used names and not indices.
vocabulary_list=thal_vocab_list,
),
]
premade_lib.set_categorical_monotonicities(self.heart_feature_configs)
def test_updates(self):
model_config = configs.CalibratedLatticeConfig(
output_min=0,
regularizer_configs=[
configs.RegularizerConfig(name='torsion', l2=2e-3),
],
feature_configs=[
configs.FeatureConfig(
name='feature_a',
pwl_calibration_input_keypoints='quantiles',
pwl_calibration_num_keypoints=8,
monotonicity=1,
pwl_calibration_clip_max=100,
),
configs.FeatureConfig(
name='feature_b',
lattice_size=3,
unimodality='valley',
pwl_calibration_input_keypoints='uniform',
pwl_calibration_num_keypoints=5,
pwl_calibration_clip_min=130,
pwl_calibration_convexity='convex',
regularizer_configs=[
configs.RegularizerConfig(name='calib_hessian',
l2=3e-3),
],
),
configs.FeatureConfig(
name='feature_c',
pwl_calibration_input_keypoints=[0.0, 0.5, 1.0],
reflects_trust_in=[
configs.TrustConfig(feature_name='feature_a'),
configs.TrustConfig(feature_name='feature_b',
direction=-1),
],
),
configs.FeatureConfig(
name='feature_d',
num_buckets=3,
vocabulary_list=['a', 'b', 'c'],
default_value=-1,
),
])
updates = [
# Update values can be passed in as numbers.
('output_max', 1.0), # update
('regularizer__torsion__l2', 0.004), # update
('regularizer__calib_hessian__l1', 0.005), # insert
('feature__feature_a__lattice_size', 3), # update
('feature__feature_e__lattice_size', 4), # insert
# Update values can be strings.
('unrelated_hparams_not_affecting_config', 'unrelated'),
('feature__feature_a__regularizer__calib_wrinkle__l1',
'0.6'), # insert
('feature__feature_b__regularizer__calib_hessian__l1',
'0.7'), # update
('yet__another__unrelated_config', '4'),
]
self.assertEqual(configs.apply_updates(model_config, updates), 7)
model_config.feature_config_by_name('feature_a').monotonicity = 'none'
model_config.feature_config_by_name(
'feature_f').num_buckets = 4 # insert
feature_names = [
feature_config.name
for feature_config in model_config.feature_configs
]
expected_feature_names = [
'feature_a', 'feature_b', 'feature_c', 'feature_d', 'feature_e',
'feature_f'
]
self.assertCountEqual(feature_names, expected_feature_names)
global_regularizer_names = [
regularizer_config.name
for regularizer_config in model_config.regularizer_configs
]
expected_global_regularizer_names = ['torsion', 'calib_hessian']
self.assertCountEqual(global_regularizer_names,
expected_global_regularizer_names)
self.assertEqual(model_config.output_max, 1.0)
self.assertEqual(
model_config.feature_config_by_name('feature_a').lattice_size, 3)
self.assertEqual(
model_config.feature_config_by_name(
'feature_b').pwl_calibration_convexity, 'convex')
self.assertEqual(
model_config.feature_config_by_name('feature_e').lattice_size, 4)
self.assertEqual(
model_config.regularizer_config_by_name('torsion').l2, 0.004)
self.assertEqual(
model_config.regularizer_config_by_name('calib_hessian').l1, 0.005)
self.assertEqual(
model_config.feature_config_by_name(
'feature_a').regularizer_config_by_name('calib_wrinkle').l1,
0.6)
self.assertEqual(
model_config.feature_config_by_name(
'feature_b').regularizer_config_by_name('calib_hessian').l1,
0.7)
import copy
import json
from absl import logging
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow_lattice.python import configs
from tensorflow_lattice.python import premade
from tensorflow_lattice.python import premade_lib
unspecified_feature_configs = [
configs.FeatureConfig(
name='numerical_1',
lattice_size=2,
pwl_calibration_input_keypoints=np.linspace(0.0, 1.0, num=10),
),
configs.FeatureConfig(
name='numerical_2',
lattice_size=2,
pwl_calibration_input_keypoints=np.linspace(0.0, 1.0, num=10),
),
configs.FeatureConfig(
name='categorical',
lattice_size=2,
num_buckets=2,
monotonicity=[('0.0', '1.0')],
vocabulary_list=['0.0', '1.0'],
),
]
def test_from_config(self):
feature_configs = [
configs.FeatureConfig(
name='feature_a',
pwl_calibration_input_keypoints='quantiles',
pwl_calibration_num_keypoints=8,
monotonicity=1,
pwl_calibration_clip_max=100,
),
configs.FeatureConfig(
name='feature_b',
lattice_size=3,
unimodality='valley',
pwl_calibration_input_keypoints='uniform',
pwl_calibration_num_keypoints=5,
pwl_calibration_clip_min=130,
pwl_calibration_convexity='convex',
regularizer_configs=[
configs.RegularizerConfig(name='calib_hesian', l2=3e-3),
],
),
configs.FeatureConfig(
name='feature_c',
pwl_calibration_input_keypoints=[0.0, 0.5, 1.0],
reflects_trust_in=[
configs.TrustConfig(feature_name='feature_a'),
configs.TrustConfig(feature_name='feature_b', direction=-1),
],
dominates=[
configs.DominanceConfig(
feature_name='feature_d', dominance_type='monotonic'),
],
),
configs.FeatureConfig(
name='feature_d',
num_buckets=3,
vocabulary_list=['a', 'b', 'c'],
default_value=-1,
),
]
# First we test CalibratedLatticeEnsembleConfig
model_config = configs.CalibratedLatticeEnsembleConfig(
feature_configs=feature_configs,
lattices=[['feature_a', 'feature_b'], ['feature_c', 'feature_d']],
separate_calibrators=True,
regularizer_configs=[
configs.RegularizerConfig('torsion', l2=1e-4),
],
output_min=0.0,
output_max=1.0,
output_calibration=True,
output_calibration_num_keypoints=5,
output_initialization=[0.0, 1.0])
model_config_copy = configs.CalibratedLatticeEnsembleConfig.from_config(
model_config.get_config(), tfl_custom_objects)
self.assertDictEqual(model_config.get_config(),
model_config_copy.get_config())
# Next we test CalibratedLatticeConfig
model_config = configs.CalibratedLatticeConfig(
feature_configs=feature_configs,
regularizer_configs=[
configs.RegularizerConfig('torsion', l2=1e-4),
],
output_min=0.0,
output_max=1.0,
output_calibration=True,
output_calibration_num_keypoints=8,
output_initialization='quantiles')
model_config_copy = configs.CalibratedLatticeConfig.from_config(
model_config.get_config(), tfl_custom_objects)
self.assertDictEqual(model_config.get_config(),
model_config_copy.get_config())
# Last we test CalibratedLinearConfig
model_config = configs.CalibratedLinearConfig(
feature_configs=feature_configs,
regularizer_configs=[
configs.RegularizerConfig('calib_hessian', l2=1e-4),
],
use_bias=True,
output_min=0.0,
output_max=None,
output_calibration=True,
output_initialization='uniform')
model_config_copy = configs.CalibratedLinearConfig.from_config(
model_config.get_config(), tfl_custom_objects)
self.assertDictEqual(model_config.get_config(),
model_config_copy.get_config())