TF Lattice 自定义 Estimator

概述

您可以使用自定义 Estimator 通过 TFL 层创建任意单调模型。本指南概述了创建此类 Estimator 所需的步骤。

设置

安装 TF Lattice 软件包：

pip install tensorflow-lattice

导入所需的软件包：

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

from tensorflow_estimator.python.estimator.canned import optimizers
from tensorflow_estimator.python.estimator.head import binary_class_head
logging.disable(sys.maxsize)

下载 UCI Statlog (Heart) 数据集：

csv_file = tf.keras.utils.get_file(
    'heart.csv', 'http://storage.googleapis.com/download.tensorflow.org/data/heart.csv')
df = pd.read_csv(csv_file)
target = df.pop('target')
train_size = int(len(df) * 0.8)
train_x = df[:train_size]
train_y = target[:train_size]
test_x = df[train_size:]
test_y = target[train_size:]
df.head()

设置用于在本指南中进行训练的默认值：

LEARNING_RATE = 0.1
BATCH_SIZE = 128
NUM_EPOCHS = 1000

特征列

与任何其他 TF Estimator 一样，数据通常需要通过 input_fn 传递给 Estimator，并使用 FeatureColumns 进行解析。

# Feature columns.
# - age
# - sex
# - ca        number of major vessels (0-3) colored by flourosopy
# - thal      3 = normal; 6 = fixed defect; 7 = reversable defect
feature_columns = [
    fc.numeric_column('age', default_value=-1),
    fc.categorical_column_with_vocabulary_list('sex', [0, 1]),
    fc.numeric_column('ca'),
    fc.categorical_column_with_vocabulary_list(
        'thal', ['normal', 'fixed', 'reversible']),
]

请注意，分类特征不需要用密集特征列包装，因为 tfl.laysers.CategoricalCalibration 层可以直接使用分类索引。

创建 input_fn

与任何其他 Estimator 一样，您可以使用 input_fn 将数据馈送给模型进行训练和评估。

train_input_fn = tf.compat.v1.estimator.inputs.pandas_input_fn(
    x=train_x,
    y=train_y,
    shuffle=True,
    batch_size=BATCH_SIZE,
    num_epochs=NUM_EPOCHS,
    num_threads=1)

test_input_fn = tf.compat.v1.estimator.inputs.pandas_input_fn(
    x=test_x,
    y=test_y,
    shuffle=False,
    batch_size=BATCH_SIZE,
    num_epochs=1,
    num_threads=1)

创建 model_fn

您可以通过多种方式创建自定义 Estimator。在这里，我们将构造一个在已解析的输入张量上调用 Keras 模型的 model_fn。要解析输入特征，您可以使用 tf.feature_column.input_layer、tf.keras.layers.DenseFeatures 或 tfl.estimators.transform_features。如果使用后者，则不需要使用密集特征列包装分类特征，并且生成的张量不会串联，这样可以更轻松地在校准层中使用特征。

要构造模型，您可以搭配使用 TFL 层或任何其他 Keras 层。在这里，我们从 TFL 层创建一个校准点阵 Keras 模型，并施加一些单调性约束。随后，我们使用 Keras 模型创建自定义 Estimator。

def model_fn(features, labels, mode, config):
  """model_fn for the custom estimator."""
  del config
  input_tensors = tfl.estimators.transform_features(features, feature_columns)
  inputs = {
      key: tf.keras.layers.Input(shape=(1,), name=key) for key in input_tensors
  }

  lattice_sizes = [3, 2, 2, 2]
  lattice_monotonicities = ['increasing', 'none', 'increasing', 'increasing']
  lattice_input = tf.keras.layers.Concatenate(axis=1)([
      tfl.layers.PWLCalibration(
          input_keypoints=np.linspace(10, 100, num=8, dtype=np.float32),
          # The output range of the calibrator should be the input range of
          # the following lattice dimension.
          output_min=0.0,
          output_max=lattice_sizes[0] - 1.0,
          monotonicity='increasing',
      )(inputs['age']),
      tfl.layers.CategoricalCalibration(
          # Number of categories including any missing/default category.
          num_buckets=2,
          output_min=0.0,
          output_max=lattice_sizes[1] - 1.0,
      )(inputs['sex']),
      tfl.layers.PWLCalibration(
          input_keypoints=[0.0, 1.0, 2.0, 3.0],
          output_min=0.0,
          output_max=lattice_sizes[0] - 1.0,
          # You can specify TFL regularizers as tuple
          # ('regularizer name', l1, l2).
          kernel_regularizer=('hessian', 0.0, 1e-4),
          monotonicity='increasing',
      )(inputs['ca']),
      tfl.layers.CategoricalCalibration(
          num_buckets=3,
          output_min=0.0,
          output_max=lattice_sizes[1] - 1.0,
          # Categorical monotonicity can be partial order.
          # (i, j) indicates that we must have output(i) <= output(j).
          # Make sure to set the lattice monotonicity to 'increasing' for this
          # dimension.
          monotonicities=[(0, 1), (0, 2)],
      )(inputs['thal']),
  ])
  output = tfl.layers.Lattice(
      lattice_sizes=lattice_sizes, monotonicities=lattice_monotonicities)(
          lattice_input)

  training = (mode == tf.estimator.ModeKeys.TRAIN)
  model = tf.keras.Model(inputs=inputs, outputs=output)
  logits = model(input_tensors, training=training)

  if training:
    optimizer = optimizers.get_optimizer_instance_v2('Adagrad', LEARNING_RATE)
  else:
    optimizer = None

  head = binary_class_head.BinaryClassHead()
  return head.create_estimator_spec(
      features=features,
      mode=mode,
      labels=labels,
      optimizer=optimizer,
      logits=logits,
      trainable_variables=model.trainable_variables,
      update_ops=model.updates)

训练和 Estimator

使用 model_fn，我们可以创建和训练 Estimator。

estimator = tf.estimator.Estimator(model_fn=model_fn)
estimator.train(input_fn=train_input_fn)
results = estimator.evaluate(input_fn=test_input_fn)
print('AUC: {}'.format(results['auc']))

AUC: 0.5701754689216614