在 tensorflow 中使用 DNNLinearCombinedEstimator 进行多标签分类

Question

我有一个多标签数据集，我想使用宽 n 深神经网络对样本进行分类。

这是一个非常小的示例，仅供测试：

import numpy as np
import pandas as pd
import tensorflow as tf

tf.enable_eager_execution()

training_df: pd.DataFrame = pd.DataFrame(
    data={
        'feature1': np.random.rand(10),
        'feature2': np.random.rand(10),
        'feature3': np.random.rand(10),
        'feature4': np.random.randint(0, 3, 10),
        'feature5': np.random.randint(0, 3, 10),
        'feature6': np.random.randint(0, 3, 10),
        'target1': np.random.randint(0, 2, 10),
        'target2': np.random.randint(0, 2, 10),
        'target3': np.random.randint(0, 2, 10)
    }
)
features = ['feature1', 'feature2', 'feature3','feature4', 'feature5', 'feature6']
targets = ['target1', 'target2', 'target3']
Categorical_Cols = ['feature4', 'feature5', 'feature6']
Numerical_Cols = ['feature1', 'feature2', 'feature3']


wide_columns = [tf.feature_column.categorical_column_with_vocabulary_list(key=x, vocabulary_list=[0, 1, -1])
                                    for x in Categorical_Cols]


deep_columns = [tf.feature_column.numeric_column(x) for x in Numerical_Cols]


def wrap_dataset(df, features, labels):
  dataset = (
      tf.data.Dataset.from_tensor_slices(
          (
              tf.cast(df[features].values, tf.float32),
              tf.cast(df[labels].values, tf.int32),
          )
      )
  )

  return(dataset)

input_fn_train = wrap_dataset(training_df, features, targets)

m = tf.contrib.estimator.DNNLinearCombinedEstimator(
          head=tf.contrib.estimator.multi_label_head(n_classes=2),
          # wide settings
          linear_feature_columns=wide_columns,
          #     linear_optimizer=tf.train.FtrlOptimizer(...),
          # deep settings
          dnn_feature_columns=deep_columns,
          #     dnn_optimizer=tf.train.ProximalAdagradOptimizer(...),
          dnn_hidden_units=[10, 30, 10])

m.train(input_fn=input_fn_train)

在这个例子中，我们有 6 个特征，包括：

• 3 个数值特征：特征 1、特征 2 和特征 3
• 3 个分类特征：feature4、feature5 和 feature6

其中每个样本有三个标签，每个标签都有一个二进制值：0 或 1。

错误与输入函数有关，我无法弄清楚如何以正确的方式定义输入函数。 任何纠正代码的帮助表示赞赏。

更新：错误是：

TypeError: <TensorSliceDataset shapes: ((6,), (3,)), types: (tf.float32, tf.int32)> is not a callable object

Answer 1

既然它说它不是一个可调用对象，你可以简单地添加 lambda 并且它应该可以工作

input_fn_train = lambda: wrap_dataset(training_df, features, targets)

另外，我认为您需要理清如何将数据传递给 Estimator。由于您使用的是特征列，因此可能需要字典。现在你正在传递张量而不是张量字典。看看这个useful post。

Answer 2

最后，我想出了如何使代码正常工作。我在这里发布它是为了帮助那些想要使用来自 tensorflow 包 1.13 版的内置函数 DNNLinearCombinedEstimator 进行多标签分类的人。

import numpy as np
import pandas as pd
import tensorflow as tf
# from tensorflow import contrib
tf.enable_eager_execution()

training_df: pd.DataFrame = pd.DataFrame(
    data={
        'feature1': np.random.rand(10),
        'feature2': np.random.rand(10),
        'feature3': np.random.rand(10),
        'feature4': np.random.randint(0, 3, 10),
        'feature5': np.random.randint(0, 3, 10),
        'feature6': np.random.randint(0, 3, 10),
        'target1': np.random.randint(0, 2, 10),
        'target2': np.random.randint(0, 2, 10),
        'target3': np.random.randint(0, 2, 10)
    }
)
features = ['feature1', 'feature2', 'feature3','feature4', 'feature5', 'feature6']
targets = ['target1', 'target2', 'target3']
Categorical_Cols = ['feature4', 'feature5', 'feature6']
Numerical_Cols = ['feature1', 'feature2', 'feature3']


wide_columns = [tf.feature_column.categorical_column_with_vocabulary_list(key=x, vocabulary_list=[0, 1, -1])
                                    for x in Categorical_Cols]


deep_columns = [tf.feature_column.numeric_column(x) for x in Numerical_Cols]


def input_fn(df):
  # Creates a dictionary mapping from each continuous feature column name (k) to
  # the values of that column stored in a constant Tensor.
  continuous_cols = {k: tf.constant(df[k].values)
                     for k in Numerical_Cols}
  # Creates a dictionary mapping from each categorical feature column name (k)
  # to the values of that column stored in a tf.SparseTensor.
  categorical_cols = {k: tf.SparseTensor(
      indices=[[i, 0] for i in range(df[k].size)],
      values=df[k].values,
      dense_shape=[df[k].size, 1])
                      for k in Categorical_Cols}
  # Merges the two dictionaries into one.
  feature_cols = continuous_cols.copy()
  feature_cols.update(categorical_cols)

  labels =tf.convert_to_tensor(training_df.as_matrix(training_df[targets].columns.tolist()), dtype=tf.int32)

  return feature_cols, labels



def train_input_fn():
  return input_fn(training_df)

def eval_input_fn():
  return input_fn(training_df)



m = tf.contrib.learn.DNNLinearCombinedEstimator(
          head=tf.contrib.learn.multi_label_head(n_classes=3),
          # wide settings
          linear_feature_columns=wide_columns,
          #     linear_optimizer=tf.train.FtrlOptimizer(...),
          # deep settings
          dnn_feature_columns=deep_columns,
          #     dnn_optimizer=tf.train.ProximalAdagradOptimizer(...),
          dnn_hidden_units=[10, 10])

m.train(input_fn=train_input_fn, steps=20)
results = m.evaluate(input_fn=eval_input_fn, steps=1)
print("#########################################################")
for key in sorted(results):
    print("%s: %s" % (key, results[key]))