TFL 레이어로 Keras 모델 만들기

TFL Keras 레이어를 사용하여 단조 및 기타 형상 제약 조건이 있는 Keras 모델을 구성할 수 있습니다. 이 예제에서는 TFL 레이어를 사용하여 UCI heart 데이터세트에 대해 보정된 격자 모델을 구축하고 훈련합니다.

보정된 격자 모델에서 각 특성은 tfl.layers.PWLCalibration 또는 tfl.layers.CategoricalCalibration 레이어에 의해 변환되고 결과는 tfl.layers.Lattice를 사용하여 비선형적으로 융합됩니다.

설정

TF Lattice 패키지 설치하기

pip install tensorflow-lattice pydot

필수 패키지 가져오기

import tensorflow as tf

import logging
import numpy as np
import pandas as pd
import sys
import tensorflow_lattice as tfl
from tensorflow import feature_column as fc
logging.disable(sys.maxsize)

UCI Statlog(Heart) 데이터세트 다운로드하기

# UCI Statlog (Heart) dataset.
csv_file = tf.keras.utils.get_file(
    'heart.csv', 'http://storage.googleapis.com/download.tensorflow.org/data/heart.csv')
training_data_df = pd.read_csv(csv_file).sample(
    frac=1.0, random_state=41).reset_index(drop=True)
training_data_df.head()

이 가이드에서 훈련에 사용되는 기본값 설정하기

LEARNING_RATE = 0.1
BATCH_SIZE = 128
NUM_EPOCHS = 100

순차형 Keras 모델

이 예제는 순차형 Keras 모델을 생성하고 TFL 레이어만 사용합니다.

격자 레이어는 input[i]이 [0, lattice_sizes[i] - 1.0] 내에 있을 것으로 예상하므로 보정 레이어보다 먼저 격자 크기를 정의해야 보정 레이어의 출력 범위를 올바르게 지정할 수 있습니다.

# Lattice layer expects input[i] to be within [0, lattice_sizes[i] - 1.0], so
lattice_sizes = [3, 2, 2, 2, 2, 2, 2]

tfl.layers.ParallelCombination 레이어를 사용하여 순차형 모델을 생성하기 위해 병렬로 실행해야 하는 보정 레이어를 그룹화합니다.

combined_calibrators = tfl.layers.ParallelCombination()

각 특성에 대한 보정 레이어를 만들고 병렬 조합 레이어에 추가합니다. 숫자 특성에는 tfl.layers.PWLCalibration을 사용하고 범주형 특성에는 tfl.layers.CategoricalCalibration을 사용합니다.

# ############### age ###############
calibrator = tfl.layers.PWLCalibration(
    # Every PWLCalibration layer must have keypoints of piecewise linear
    # function specified. Easiest way to specify them is to uniformly cover
    # entire input range by using numpy.linspace().
    input_keypoints=np.linspace(
        training_data_df['age'].min(), training_data_df['age'].max(), num=5),
    # You need to ensure that input keypoints have same dtype as layer input.
    # You can do it by setting dtype here or by providing keypoints in such
    # format which will be converted to desired tf.dtype by default.
    dtype=tf.float32,
    # Output range must correspond to expected lattice input range.
    output_min=0.0,
    output_max=lattice_sizes[0] - 1.0,
)
combined_calibrators.append(calibrator)

# ############### sex ###############
# For boolean features simply specify CategoricalCalibration layer with 2
# buckets.
calibrator = tfl.layers.CategoricalCalibration(
    num_buckets=2,
    output_min=0.0,
    output_max=lattice_sizes[1] - 1.0,
    # Initializes all outputs to (output_min + output_max) / 2.0.
    kernel_initializer='constant')
combined_calibrators.append(calibrator)

# ############### cp ###############
calibrator = tfl.layers.PWLCalibration(
    # Here instead of specifying dtype of layer we convert keypoints into
    # np.float32.
    input_keypoints=np.linspace(1, 4, num=4, dtype=np.float32),
    output_min=0.0,
    output_max=lattice_sizes[2] - 1.0,
    monotonicity='increasing',
    # You can specify TFL regularizers as a tuple ('regularizer name', l1, l2).
    kernel_regularizer=('hessian', 0.0, 1e-4))
combined_calibrators.append(calibrator)

# ############### trestbps ###############
calibrator = tfl.layers.PWLCalibration(
    # Alternatively, you might want to use quantiles as keypoints instead of
    # uniform keypoints
    input_keypoints=np.quantile(training_data_df['trestbps'],
                                np.linspace(0.0, 1.0, num=5)),
    dtype=tf.float32,
    # Together with quantile keypoints you might want to initialize piecewise
    # linear function to have 'equal_slopes' in order for output of layer
    # after initialization to preserve original distribution.
    kernel_initializer='equal_slopes',
    output_min=0.0,
    output_max=lattice_sizes[3] - 1.0,
    # You might consider clamping extreme inputs of the calibrator to output
    # bounds.
    clamp_min=True,
    clamp_max=True,
    monotonicity='increasing')
combined_calibrators.append(calibrator)

# ############### chol ###############
calibrator = tfl.layers.PWLCalibration(
    # Explicit input keypoint initialization.
    input_keypoints=[126.0, 210.0, 247.0, 286.0, 564.0],
    dtype=tf.float32,
    output_min=0.0,
    output_max=lattice_sizes[4] - 1.0,
    # Monotonicity of calibrator can be decreasing. Note that corresponding
    # lattice dimension must have INCREASING monotonicity regardless of
    # monotonicity direction of calibrator.
    monotonicity='decreasing',
    # Convexity together with decreasing monotonicity result in diminishing
    # return constraint.
    convexity='convex',
    # You can specify list of regularizers. You are not limited to TFL
    # regularizrs. Feel free to use any :)
    kernel_regularizer=[('laplacian', 0.0, 1e-4),
                        tf.keras.regularizers.l1_l2(l1=0.001)])
combined_calibrators.append(calibrator)

# ############### fbs ###############
calibrator = tfl.layers.CategoricalCalibration(
    num_buckets=2,
    output_min=0.0,
    output_max=lattice_sizes[5] - 1.0,
    # For categorical calibration layer monotonicity is specified for pairs
    # of indices of categories. Output for first category in pair will be
    # smaller than output for second category.
    #
    # Don't forget to set monotonicity of corresponding dimension of Lattice
    # layer to '1'.
    monotonicities=[(0, 1)],
    # This initializer is identical to default one('uniform'), but has fixed
    # seed in order to simplify experimentation.
    kernel_initializer=tf.keras.initializers.RandomUniform(
        minval=0.0, maxval=lattice_sizes[5] - 1.0, seed=1))
combined_calibrators.append(calibrator)

# ############### restecg ###############
calibrator = tfl.layers.CategoricalCalibration(
    num_buckets=3,
    output_min=0.0,
    output_max=lattice_sizes[6] - 1.0,
    # Categorical monotonicity can be partial order.
    monotonicities=[(0, 1), (0, 2)],
    # Categorical calibration layer supports standard Keras regularizers.
    kernel_regularizer=tf.keras.regularizers.l1_l2(l1=0.001),
    kernel_initializer='constant')
combined_calibrators.append(calibrator)

그런 다음 calibrator의 출력을 비선형적으로 융합하기 위해 격자 레이어를 만듭니다.

필요한 차원에 대해 증가할 격자의 단조를 지정해야 합니다. 보정에서 단조로운 방향의 구성은 단조의 엔드 투 엔드 방향을 올바르게 만듭니다. 여기에는 CategoricalCalibration 레이어의 부분 단조가 포함됩니다.

lattice = tfl.layers.Lattice(
    lattice_sizes=lattice_sizes,
    monotonicities=[
        'increasing', 'none', 'increasing', 'increasing', 'increasing',
        'increasing', 'increasing'
    ],
    output_min=0.0,
    output_max=1.0)

그런 다음 결합된 calibrator 및 격자 레이어를 사용하여 순차형 모델을 만들 수 있습니다.

model = tf.keras.models.Sequential()
model.add(combined_calibrators)
model.add(lattice)

훈련은 다른 Keras 모델과 동일하게 동작합니다.

features = training_data_df[[
    'age', 'sex', 'cp', 'trestbps', 'chol', 'fbs', 'restecg'
]].values.astype(np.float32)
target = training_data_df[['target']].values.astype(np.float32)

model.compile(
    loss=tf.keras.losses.mean_squared_error,
    optimizer=tf.keras.optimizers.Adagrad(learning_rate=LEARNING_RATE))
model.fit(
    features,
    target,
    batch_size=BATCH_SIZE,
    epochs=NUM_EPOCHS,
    validation_split=0.2,
    shuffle=False,
    verbose=0)

model.evaluate(features, target)

10/10 [==============================] - 0s 2ms/step - loss: 0.1551
0.1550564467906952

함수형 Keras 모델

이 예제에서는 Keras 모델 생성을 위한 함수형 API를 사용합니다.

이전 섹션에서 언급했듯이 격자 레이어는 input[i]가 [0, lattice_sizes[i] - 1.0] 내에 있을 것으로 예상되므로 보정 레이어보다 먼저 격자 크기를 정의해야 보정 레이어의 출력 범위를 적절하게 지정할 수 있습니다.

# We are going to have 2-d embedding as one of lattice inputs.
lattice_sizes = [3, 2, 2, 3, 3, 2, 2]

각 특성에 대해 입력 레이어와 보정 레이어를 만들어야 합니다. 숫자 특성에는 tfl.layers.PWLCalibration을 사용하고 범주형 특성에는 tfl.layers.CategoricalCalibration을 사용합니다.

model_inputs = []
lattice_inputs = []
# ############### age ###############
age_input = tf.keras.layers.Input(shape=[1], name='age')
model_inputs.append(age_input)
age_calibrator = tfl.layers.PWLCalibration(
    # Every PWLCalibration layer must have keypoints of piecewise linear
    # function specified. Easiest way to specify them is to uniformly cover
    # entire input range by using numpy.linspace().
    input_keypoints=np.linspace(
        training_data_df['age'].min(), training_data_df['age'].max(), num=5),
    # You need to ensure that input keypoints have same dtype as layer input.
    # You can do it by setting dtype here or by providing keypoints in such
    # format which will be converted to desired tf.dtype by default.
    dtype=tf.float32,
    # Output range must correspond to expected lattice input range.
    output_min=0.0,
    output_max=lattice_sizes[0] - 1.0,
    monotonicity='increasing',
    name='age_calib',
)(
    age_input)
lattice_inputs.append(age_calibrator)

# ############### sex ###############
# For boolean features simply specify CategoricalCalibration layer with 2
# buckets.
sex_input = tf.keras.layers.Input(shape=[1], name='sex')
model_inputs.append(sex_input)
sex_calibrator = tfl.layers.CategoricalCalibration(
    num_buckets=2,
    output_min=0.0,
    output_max=lattice_sizes[1] - 1.0,
    # Initializes all outputs to (output_min + output_max) / 2.0.
    kernel_initializer='constant',
    name='sex_calib',
)(
    sex_input)
lattice_inputs.append(sex_calibrator)

# ############### cp ###############
cp_input = tf.keras.layers.Input(shape=[1], name='cp')
model_inputs.append(cp_input)
cp_calibrator = tfl.layers.PWLCalibration(
    # Here instead of specifying dtype of layer we convert keypoints into
    # np.float32.
    input_keypoints=np.linspace(1, 4, num=4, dtype=np.float32),
    output_min=0.0,
    output_max=lattice_sizes[2] - 1.0,
    monotonicity='increasing',
    # You can specify TFL regularizers as tuple ('regularizer name', l1, l2).
    kernel_regularizer=('hessian', 0.0, 1e-4),
    name='cp_calib',
)(
    cp_input)
lattice_inputs.append(cp_calibrator)

# ############### trestbps ###############
trestbps_input = tf.keras.layers.Input(shape=[1], name='trestbps')
model_inputs.append(trestbps_input)
trestbps_calibrator = tfl.layers.PWLCalibration(
    # Alternatively, you might want to use quantiles as keypoints instead of
    # uniform keypoints
    input_keypoints=np.quantile(training_data_df['trestbps'],
                                np.linspace(0.0, 1.0, num=5)),
    dtype=tf.float32,
    # Together with quantile keypoints you might want to initialize piecewise
    # linear function to have 'equal_slopes' in order for output of layer
    # after initialization to preserve original distribution.
    kernel_initializer='equal_slopes',
    output_min=0.0,
    output_max=lattice_sizes[3] - 1.0,
    # You might consider clamping extreme inputs of the calibrator to output
    # bounds.
    clamp_min=True,
    clamp_max=True,
    monotonicity='increasing',
    name='trestbps_calib',
)(
    trestbps_input)
lattice_inputs.append(trestbps_calibrator)

# ############### chol ###############
chol_input = tf.keras.layers.Input(shape=[1], name='chol')
model_inputs.append(chol_input)
chol_calibrator = tfl.layers.PWLCalibration(
    # Explicit input keypoint initialization.
    input_keypoints=[126.0, 210.0, 247.0, 286.0, 564.0],
    output_min=0.0,
    output_max=lattice_sizes[4] - 1.0,
    # Monotonicity of calibrator can be decreasing. Note that corresponding
    # lattice dimension must have INCREASING monotonicity regardless of
    # monotonicity direction of calibrator.
    monotonicity='decreasing',
    # Convexity together with decreasing monotonicity result in diminishing
    # return constraint.
    convexity='convex',
    # You can specify list of regularizers. You are not limited to TFL
    # regularizrs. Feel free to use any :)
    kernel_regularizer=[('laplacian', 0.0, 1e-4),
                        tf.keras.regularizers.l1_l2(l1=0.001)],
    name='chol_calib',
)(
    chol_input)
lattice_inputs.append(chol_calibrator)

# ############### fbs ###############
fbs_input = tf.keras.layers.Input(shape=[1], name='fbs')
model_inputs.append(fbs_input)
fbs_calibrator = tfl.layers.CategoricalCalibration(
    num_buckets=2,
    output_min=0.0,
    output_max=lattice_sizes[5] - 1.0,
    # For categorical calibration layer monotonicity is specified for pairs
    # of indices of categories. Output for first category in pair will be
    # smaller than output for second category.
    #
    # Don't forget to set monotonicity of corresponding dimension of Lattice
    # layer to '1'.
    monotonicities=[(0, 1)],
    # This initializer is identical to default one ('uniform'), but has fixed
    # seed in order to simplify experimentation.
    kernel_initializer=tf.keras.initializers.RandomUniform(
        minval=0.0, maxval=lattice_sizes[5] - 1.0, seed=1),
    name='fbs_calib',
)(
    fbs_input)
lattice_inputs.append(fbs_calibrator)

# ############### restecg ###############
restecg_input = tf.keras.layers.Input(shape=[1], name='restecg')
model_inputs.append(restecg_input)
restecg_calibrator = tfl.layers.CategoricalCalibration(
    num_buckets=3,
    output_min=0.0,
    output_max=lattice_sizes[6] - 1.0,
    # Categorical monotonicity can be partial order.
    monotonicities=[(0, 1), (0, 2)],
    # Categorical calibration layer supports standard Keras regularizers.
    kernel_regularizer=tf.keras.regularizers.l1_l2(l1=0.001),
    kernel_initializer='constant',
    name='restecg_calib',
)(
    restecg_input)
lattice_inputs.append(restecg_calibrator)

그런 다음 calibrator의 출력을 비선형적으로 융합하기 위해 격자 레이어를 만듭니다.

필요한 차원에 대해 증가할 격자의 단조를 지정해야 합니다. 보정에서 단조로운 방향의 구성은 단조의 엔드 투 엔드 방향을 올바르게 만듭니다. 여기에는 tfl.layers.CategoricalCalibration 레이어의 부분 단조가 포함됩니다.

lattice = tfl.layers.Lattice(
    lattice_sizes=lattice_sizes,
    monotonicities=[
        'increasing', 'none', 'increasing', 'increasing', 'increasing',
        'increasing', 'increasing'
    ],
    output_min=0.0,
    output_max=1.0,
    name='lattice',
)(
    lattice_inputs)

모델에 더 많은 유연성을 추가하기 위해 출력 보정 레이어를 추가합니다.

model_output = tfl.layers.PWLCalibration(
    input_keypoints=np.linspace(0.0, 1.0, 5),
    name='output_calib',
)(
    lattice)

이제 입력과 출력을 사용하여 모델을 만들 수 있습니다.

model = tf.keras.models.Model(
    inputs=model_inputs,
    outputs=model_output)
tf.keras.utils.plot_model(model, rankdir='LR')

png

훈련은 다른 Keras 모델과 동일하게 동작합니다. 설정에서 입력 특성은 별도의 텐서로 전달됩니다.

feature_names = ['age', 'sex', 'cp', 'trestbps', 'chol', 'fbs', 'restecg']
features = np.split(
    training_data_df[feature_names].values.astype(np.float32),
    indices_or_sections=len(feature_names),
    axis=1)
target = training_data_df[['target']].values.astype(np.float32)

model.compile(
    loss=tf.keras.losses.mean_squared_error,
    optimizer=tf.keras.optimizers.Adagrad(LEARNING_RATE))
model.fit(
    features,
    target,
    batch_size=BATCH_SIZE,
    epochs=NUM_EPOCHS,
    validation_split=0.2,
    shuffle=False,
    verbose=0)

model.evaluate(features, target)

10/10 [==============================] - 0s 3ms/step - loss: 0.1591
0.15907342731952667