具有连续和分类列的 HOWTO tf.estimator

Question

I have a tf.estimator which works for continuous variables and I want to expand it to use categorical variables.

Consider a pandas dataframe which looks like this:

label           |  con_col          |  cat_col
(float 0 or 1)  |  (float -1 to 1)  |  (int 0-3)
----------------+-------------------+---------------
0               |   0.123           |  2
0               |   0.456           |  1
1               |  -0.123           |  3
1               |  -0.123           |  3
0               |   0.123           |  2

To build the estimator for just the label and the continuous variable column (con_col) I build the following feature_column variable.

feature_cols = [
                   tf.feature_column.numeric_column('con_col')
               ]

Then I pass it to the DNNClassifer like so.

tf.estimator.DNNClassifier(feature_columns=feature_cols ...)

Later I build a serving_input_fn(). In this function I also specify the columns. This routine is quite small and looks like this:

def serving_input_fn(): 
    feat_placeholders['con_col'] = tf.placeholder(tf.float32, [None])

    return tf.estimator.export.ServingInputReceiver(feat_placeholders.copy(), feat_placeholders)

This works. However, when I try to use the categorical column I have a problem.

So using the categorical column, this part seems to work.

feature_cols = [
    tf.feature_column.sequence_categorical_column_with_identity('cat_col', num_buckets=4))
               ]
tf.estimator.DNNClassifier(feature_columns=feature_cols ...)

For the serving_input_fn() I get suggestions from the stack trace but both suggestions fail.:

def serving_input_fn(): 
    # try #2
    # this fails
    feat_placeholders['cat_col'] = tf.SequenceCategoricalColumn(categorical_column=tf.IdentityCategoricalColumn(key='cat_col', number_buckets=4,default_value=None))

    # try #1
    # this also fails
    # feat_placeholders['cat_col'] = tf.feature_column.indicator_column(tf.feature_column.sequence_categorical_column_with_identity(column, num_buckets=4))

    # try #0
    # this fails. Its using the same form for the con_col
    # the resulting error gave hints for the above code.
    # Note, i'm using this url as a guide.  My cat_col is
    # is similar to that code samples 'dayofweek' except it
    # is not a string.
    # https://github.com/GoogleCloudPlatform/training-data-analyst/blob/master/courses/machine_learning/feateng/taxifare_tft/trainer/model.py
    #feat_placeholders['cat_col'] = tf.placeholder(tf.float32, [None])


    return tf.estimator.export.ServingInputReceiver(feat_placeholders.copy(), feat_placeholders)

This is the error message if try #0 is used.

ValueError: Items of feature_columns must be a <class 'tensorflow.python.feature_column.feature_column_v2.DenseColumn'>. You can wrap a categorical column with an embedding_column or indicator_column. Given: SequenceCategoricalColumn(categorical_column=IdentityCategoricalColumn(key='cat_col', number_buckets=4, default_value=None))

Lak's answer implementation

Using Lak's answer as a guide, this works for both both feature columns.

# This is the list of features we pass as an argument to DNNClassifier
feature_cols = []

# Add the continuous column first
feature_cols.append(tf.feature_column.numeric_column('con_col'))                  

# Add the categorical column which is wrapped?
# This creates new columns from a single column?
category_feature_cols = [tf.feature_column.categorical_column_with_identity('cat_col', num_buckets=4)]
for c in category_feature_cols:
    feat_cols.append(tf.feature_column.indicator_column(c))

# now pass this list to the DNN
tf.estimator.DNNClassifier(feature_columns=feature_cols ...)


def serving_input_fn(): 
    feat_placeholders['con_col'] = tf.placeholder(tf.float32, [None])
    feat_placeholders['cat_col'] = tf.placeholder(tf.int64, [None])

Answer 1

You need to wrap categorical columns before sending to DNN:

cat_feature_cols = [ tf.feature_column.sequence_categorical_column_with_identity('cat_col', num_buckets=4)) ]
feature_cols = [tf.feature_column.indicator_column(c) for c in cat_feature_cols]

Use indicator column to one-hot encode, or embedded column to embed.