Creating Keras Models with TFL Layers

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Overview

You can use TFL Keras layers to construct Keras models with monotonicity and other shape constraints. This example builds and trains a calibrated lattice model for the UCI heart dataset using TFL layers.

In a calibrated lattice model, each feature is transformed by a tfl.layers.PWLCalibration or a tfl.layers.CategoricalCalibration layer and the results are nonlinearly fused using a tfl.layers.Lattice.

Setup

Installing TF Lattice package:


!pip install -q tensorflow-lattice pydot

Importing required packages:

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)

Downloading the UCI Statlog (Heart) dataset:

# UCI Statlog (Heart) dataset.
csv_file = tf.keras.utils.get_file(
    'heart.csv', 'http://storage.googleapis.com/applied-dl/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()

Setting the default values used for training in this guide:

LEARNING_RATE = 0.1
BATCH_SIZE = 128
NUM_EPOCHS = 100

Sequential Keras Model

This example creates a Sequential Keras model and only uses TFL layers.

Lattice layers expect input[i] to be within [0, lattice_sizes[i] - 1.0], so we need to define the lattice sizes ahead of the calibration layers so we can properly specify output range of the calibration layers.

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

We use a tfl.layers.ParallelCombination layer to group together calibration layers which have to be executed in paralel in order to be able to create a Sequential model.

combined_calibrators = tfl.layers.ParallelCombination()

We create a calibration layer for each feature and add it to the parallel combination layer. For numeric features we use tfl.layers.PWLCalibration and for categorical features we use tfl.layers.CategoricalCalibration.

# ############### 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 deisred 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)

We then create a lattice layer to nonlinearly fuse the outputs of the calibrators.

Note that we need to specify the monotonicity of the lattice to be increasing for required dimensions. The composition with the direction of the monotonicity in the calibration will result in the correct end-to-end direction of monotonicity. This includes partial monotonicity of CategoricalCalibration layer.

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

We can then create a sequential model using the combined calibrators and lattice layers.

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

Training works the same as any other keras model.

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)
303/303 [==============================] - 0s 50us/sample - loss: 0.1552

0.15515593279509654

Functional Keras Model

This example uses a functional API for Keras model construction.

As mentioned in the previous section, lattice layers expect input[i] to be within [0, lattice_sizes[i] - 1.0], so we need to define the lattice sizes ahead of the calibration layers so we can properly specify output range of the calibration layers.

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

For each feature, we need to create an input layer followed by a calibration layer. For numeric features we use tfl.layers.PWLCalibration and for categorical features we use tfl.layers.CategoricalCalibration.

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 deisred 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)

We then create a lattice layer to nonlinearly fuse the outputs of the calibrators.

Note that we need to specify the monotonicity of the lattice to be increasing for required dimensions. The composition with the direction of the monotonicity in the calibration will result in the correct end-to-end direction of monotonicity. This includes partial monotonicity of tfl.layers.CategoricalCalibration layer.

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)

To add more flexibility to the model, we add an output calibration layer.

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

We can now create a model using the inputs and outputs.

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

png

Training works the same as any other keras model. Note that, with our setup, input features are passed as separate tensors.

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)
303/303 [==============================] - 0s 59us/sample - loss: 0.1590

0.15904678805138017