具有不同数据集的 AttributeError 的 Tensorflow 宽而深的模型

Question

我正在使用 Kaggle Rossmann dataset 来训练一个广泛而深入的模型。该代码与教程中给出的代码非常相似。我只是更改用于建模的数据。

我使用的代码如下：

"""Example code for TensorFlow Wide & Deep Tutorial using TF.Learn API."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import shutil
import sys
import tempfile

import pandas as pd
from six.moves import urllib
import tensorflow as tf


CSV_COLUMNS = [
    'Store', 'DayOfWeek',  'Sales', 'Customers', 'Open', 'Promo',
    'StateHoliday', 'SchoolHoliday', 'StoreType', 'Assortment',
    'CompetitionDistance', 'trend', 'Max_TemperatureC', 'Mean_TemperatureC',
    'Min_TemperatureC', 'Max_Humidity', 'Mean_Humidity', 'Min_Humidity'
]

StateHoliday = tf.feature_column.categorical_column_with_vocabulary_list(
    "StateHoliday", ["True", "False"])
StoreType = tf.feature_column.categorical_column_with_vocabulary_list(
    "StoreType", ['c', 'a', 'd', 'b'])
Assortment = tf.feature_column.categorical_column_with_vocabulary_list(
    "Assortment", ['c', 'a', 'b'])

CompetitionDistance = tf.feature_column.categorical_column_with_hash_bucket(
    "CompetitionDistance", hash_bucket_size=1000)
Customers = tf.feature_column.categorical_column_with_hash_bucket(
    "Customers", hash_bucket_size=1000)
Store = tf.feature_column.categorical_column_with_hash_bucket(
    "Store", hash_bucket_size=1000)

trend = tf.feature_column.numeric_column("trend")
Max_TemperatureC = tf.feature_column.numeric_column("Max_TemperatureC")
Mean_TemperatureC = tf.feature_column.numeric_column("Mean_TemperatureC")
Min_TemperatureC = tf.feature_column.numeric_column("Min_TemperatureC")
Max_Humidity = tf.feature_column.numeric_column("Max_Humidity")
Mean_Humidity = tf.feature_column.numeric_column("Mean_Humidity")
Min_Humidity = tf.feature_column.numeric_column("Min_Humidity")


crossed_columns = [

    tf.feature_column.crossed_column(
        ["Assortment", "StoreType"], hash_bucket_size=1000)
]

deep_columns = [
    tf.feature_column.indicator_column("DayOfWeek"),
    tf.feature_column.indicator_column("Open"),
    tf.feature_column.indicator_column("Promo"),
    tf.feature_column.indicator_column("StateHoliday"),
    tf.feature_column.indicator_column("SchoolHoliday"),
    tf.feature_column.indicator_column("StoreType"),
    tf.feature_column.indicator_column("Assortment"),
    # To show an example of embedding
    tf.feature_column.embedding_column("CompetitionDistance", dimension=8),
    tf.feature_column.embedding_column("Customers", dimension=8),
    tf.feature_column.embedding_column("Store", dimension=8),
    trend,
    Max_TemperatureC,
    Mean_TemperatureC,
    Min_TemperatureC,
    Max_Humidity,
    Mean_Humidity,
    Min_Humidity
]


def build_estimator(model_dir):
    """Build an estimator."""

    m = tf.estimator.DNNLinearCombinedClassifier(
        model_dir=model_dir,
        linear_feature_columns=crossed_columns,
        dnn_feature_columns=deep_columns,
        dnn_hidden_units=[100, 50])
    return m


def input_fn(data_file, num_epochs, shuffle):
    df_data = pd.read_csv(
        "D:/Rossmann/Rossmann_Data/" + data_file + ".csv",
        names=CSV_COLUMNS,
        skipinitialspace=True,
        engine="python",
        skiprows=1)

    # remove NaN elements
    df_data = df_data.dropna(how="any", axis=0)
    print(df_data.dtypes)
    df_data = df_data.sort(['Sales'], ascending=[True])
    labels = df_data["Sales"].apply(lambda x: 1 if x >= 20000 else 0)

    return tf.estimator.inputs.pandas_input_fn(
        x=df_data,
        y=labels,
        batch_size=100,
        num_epochs=num_epochs,
        shuffle=shuffle,
        num_threads=5)


model_dir = "D:/Rossmann/Rossmann_Data"

m = build_estimator(model_dir)

m.train(
    input_fn=input_fn("df1", num_epochs=None, shuffle=True),
    steps=2000)

但不幸的是，我收到以下错误。

Traceback (most recent call last):
  File "timeSeriesPredictionUsingEmbedding2.py", line 121, in <module>
    steps=2000)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\estimator\estimator.py", line 241, in train
    loss = self._train_model(input_fn=input_fn, hooks=hooks)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\estimator\estimator.py", line 630, in _train_model
    model_fn_lib.ModeKeys.TRAIN)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\estimator\estimator.py", line 615, in _call_model_fn
    model_fn_results = self._model_fn(features=features, **kwargs)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\estimator\canned\dnn_linear_combined.py", line 395, in _model_fn
    config=config)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\estimator\canned\dnn_linear_combined.py", line 156, in _dnn_linear_combined_model_fn
    feature_columns=dnn_feature_columns)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\feature_column\feature_column.py", line 207, in input_layer
    _check_feature_columns(feature_columns)
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\feature_column\feature_column.py", line 1662, in _check_feature_columns
    if column.name in name_to_column:
  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\feature_column\feature_column.py", line 2453, in name
    return '{}_indicator'.format(self.categorical_column.name)
AttributeError: 'str' object has no attribute 'name'

你能指导我在哪里得到这个错误吗？当我运行您的代码时，它运行良好。

谢谢！

Answer 1

原因是tf.feature_column.indicator_column 接受分类列实例，而不是列名（“DayOfWeek”、“Open”等）。可能这个 API 在以前的 tf 版本中看起来有所不同，我不确定，但现在你必须创建一个 categorical_column_* 然后用指示器包装。

顺便说一句，我看到您正在使用 DataFrame.sort - 此方法已被弃用，并且不再适用于最新的 pandas。使用sort_values。

更新

我没有注意到代码是对tutorial example 的改编，这就是为什么它故意使用所有可能的特征类型、散列、嵌入、跨列。通常，不必一次使用所有这些，特别是对于 Rossmann 数据来说不是必需的。您可以进一步添加，例如，交叉列，如果您注意到数据中的特征相关性，但只是开始，数据主要是numeric_columns，很少有categorical_column_with_vocabulary_lists。

这是我的完整版代码：

"""Example code for TensorFlow Wide & Deep Tutorial using TF.Learn API."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import pandas as pd
import tensorflow as tf

CSV_COLUMNS = [
  'Store', 'DayOfWeek', 'Sales', 'Customers', 'Open', 'Promo',
  'StateHoliday', 'SchoolHoliday', 'StoreType', 'Assortment',
  'CompetitionDistance', 'trend', 'Max_TemperatureC', 'Mean_TemperatureC',
  'Min_TemperatureC', 'Max_Humidity', 'Mean_Humidity', 'Min_Humidity'
]

Store = tf.feature_column.numeric_column("Store")
DayOfWeek = tf.feature_column.numeric_column("DayOfWeek")
Customers = tf.feature_column.numeric_column("Customers")
Open = tf.feature_column.numeric_column("Open")
Promo = tf.feature_column.numeric_column("Promo")
StateHoliday = tf.feature_column.categorical_column_with_vocabulary_list("StateHoliday", ["True", "False"])
SchoolHoliday = tf.feature_column.numeric_column("SchoolHoliday")
StoreType = tf.feature_column.categorical_column_with_vocabulary_list("StoreType", ['a', 'b', 'c', 'd'])
Assortment = tf.feature_column.categorical_column_with_vocabulary_list("Assortment", ['a', 'b', 'c'])
CompetitionDistance = tf.feature_column.numeric_column("CompetitionDistance")
trend = tf.feature_column.numeric_column("trend")
Max_TemperatureC = tf.feature_column.numeric_column("Max_TemperatureC")
Mean_TemperatureC = tf.feature_column.numeric_column("Mean_TemperatureC")
Min_TemperatureC = tf.feature_column.numeric_column("Min_TemperatureC")
Max_Humidity = tf.feature_column.numeric_column("Max_Humidity")
Mean_Humidity = tf.feature_column.numeric_column("Mean_Humidity")
Min_Humidity = tf.feature_column.numeric_column("Min_Humidity")

deep_columns = [
  Store,
  DayOfWeek,
  Customers,
  Open,
  Promo,
  tf.feature_column.indicator_column(StateHoliday),
  SchoolHoliday,
  tf.feature_column.indicator_column(StoreType),
  tf.feature_column.indicator_column(Assortment),
  CompetitionDistance,
  trend,
  Max_TemperatureC,
  Mean_TemperatureC,
  Min_TemperatureC,
  Max_Humidity,
  Mean_Humidity,
  Min_Humidity
]


def build_estimator(model_dir):
  """Build an estimator."""
  return tf.estimator.DNNLinearCombinedClassifier(
    model_dir=model_dir,
    dnn_feature_columns=deep_columns,
    dnn_hidden_units=[100, 50])


def input_fn(data_file, num_epochs, shuffle):
  df_data = pd.read_csv(data_file + ".csv",
                        names=CSV_COLUMNS,
                        dtype={"StateHoliday": str},
                        skipinitialspace=True,
                        engine="python",
                        skiprows=1)

  # remove NaN elements
  df_data = df_data.dropna(how="any", axis=0)
  df_data = df_data.sort_values(['Sales'], ascending=[True])
  labels = df_data["Sales"].apply(lambda x: 1 if x >= 20000 else 0)

  return tf.estimator.inputs.pandas_input_fn(
    x=df_data,
    y=labels,
    batch_size=100,
    num_epochs=num_epochs,
    shuffle=shuffle,
    num_threads=5)


m = build_estimator(model_dir="./model")
m.train(input_fn=input_fn("df1", num_epochs=None, shuffle=True),
        steps=2000)