tensorflow：如何从特征字典动态选择张量

Question

说我正在使用数据集+ feature_columns生成特征张量， 我想执行以下注释中的操作，以选择另一个indicator标量张量的特征量中的一个张量：

def _parse_function(example_proto):
  features = {"f_0": tf.FixedLenFeature((), tf.int32, default_value=0),
              "f_1": tf.FixedLenFeature((), tf.int32, default_value=0), 
              "f_2": tf.FixedLenFeature((), tf.int32, default_value=0), 
              "indicator": tf.FixedLenFeature((), tf.int32, default_value=0),}
  parsed_features = tf.parse_single_example(example_proto, features)

  # what i want to do:
  # i = parsed_features['indicator']
  # parsed_features["f"] = parsed_features["f_{}".format(i)]

  return parsed_features

dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(_parse_function)

# f is selected from [f_0, f_1, f_2]
feature_column = tf.feature_column.numeric_column(key="f")

我想知道如何在tensorflow中实现它。

Answer 1

有两种方法可以做到这一点。下面是使用查找表和重组输入数据的示例。

---

查询表

从解析的示例中创建一个HashTable，然后使用table.lookup获得所需的功能。

可运行的示例：

import tensorflow as tf

# Create example proto
def _int_feature(value):
  return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

e = tf.train.Example(features=tf.train.Features(feature={
  "f_0": _int_feature([10]),
  "f_1": _int_feature([20]),
  "f_2": _int_feature([30]),
  "indicator": _int_feature([1])}))
e_proto = e.SerializeToString()

def _parse_function(example_proto):
  features = {"f_0": tf.FixedLenFeature((), tf.int64, default_value=0),
              "f_1": tf.FixedLenFeature((), tf.int64, default_value=0),
              "f_2": tf.FixedLenFeature((), tf.int64, default_value=0),
              "indicator": tf.FixedLenFeature((), tf.int64, default_value=0)}
  parsed_features = tf.parse_single_example(example_proto, features)

  # Create HashTable
  keys = ["f_0", "f_1", "f_2"]
  values = [parsed_features[k] for k in keys]
  table = tf.contrib.lookup.HashTable(
      tf.contrib.lookup.KeyValueTensorInitializer(keys, values),
      default_value=-1)

  # Add lookup value to the parsed_features dictionary
  indicator_key = "f_" + tf.as_string(parsed_features["indicator"])
  parsed_features["f"] = table.lookup(indicator_key)

  return parsed_features

features = _parse_function(e_proto)

with tf.Session() as sess:
  sess.run(tf.initialize_all_tables())
  print(sess.run(features))

输出： {'indicator': 1, 'f_0': 10, 'f_1': 20, 'f_2': 30, 'f': 20}

---

重组输入数据

将要素组合到一个矩阵中，并使用指标索引到要素矩阵中。

可运行的示例：

import tensorflow as tf

# Create example proto
def _int_feature(value):
  return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

e = tf.train.Example(features=tf.train.Features(feature={
  "all_features": _int_feature([10, 20, 30]),
  "indicator": _int_feature([1])}))
e_proto = e.SerializeToString()

def _parse_function(example_proto):
  features = {"all_features": tf.FixedLenFeature([3], tf.int64),
              "indicator": tf.FixedLenFeature((), tf.int64, default_value=0)}
  parsed_features = tf.parse_single_example(example_proto, features)

  indicator_key = parsed_features["indicator"]
  parsed_features["f"] = parsed_features["all_features"][indicator_key]

  return parsed_features

features = _parse_function(e_proto)

with tf.Session() as sess:
  print(sess.run(features))

输出： {'indicator': 1, 'all_features': array([10, 20, 30]), 'f': 20}