What do I pass as the x and y parameters in tf.contrib.learn.LinearClassifier.fit

Question

Question

I've set up a toy example of the TensorFlow linear classifier tutorial . In this example, the fit method is called with a parameter input_fn in which I pass train_input_fn . This is how TensorFlow likes to pass the data. However, I really want to run mini batches. Fortunately, fit has a batch_size parameter, but I need to forgo the use of input_fn and pass x and y instead. I've tried passing ndarray s and DataFrames as well as the output from the train_input_fn function. Nothing works. I need a working example of using the batch_size parameter.

Setup

Here is the setup code split into stuff I have no problem with followed by the problem portion.

No problem with (feel free to copy/paste/run)

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

np.random.seed([3,1415])
df = pd.DataFrame(dict(cat1=np.random.choice(('Yes', 'No'), (100,),),
                       val1=np.random.rand(100),
                       val2=np.random.rand(100),
                       val3=np.random.rand(100),
                       label=np.random.choice((0, 1), (100,))))

LABEL_COLUMN = "label"

trainBegin, trainEnd = 0, 80
testBegin, testEnd = 80, 100
df_train = df.iloc[trainBegin:trainEnd, :]
df_test = df.iloc[testBegin:testEnd, :]

CONTINUOUS_COLUMNS = ['val1', 'val2', 'val3']
CATEGORICAL_COLUMNS = ['cat1']

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 CONTINUOUS_COLUMNS}
    # 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,
            shape=[df[k].size, 1])
                        for k in CATEGORICAL_COLUMNS}
    # Merges the two dictionaries into one.
    feature_cols = dict(continuous_cols.items() + categorical_cols.items())
    # Converts the label column into a constant Tensor.
    label = tf.constant(df[LABEL_COLUMN].values)
    # Returns the feature columns and the label.
    return feature_cols, label

def train_input_fn():
    return input_fn(df_train)

def eval_input_fn():
    return input_fn(df_test)

val1 = tf.contrib.layers.real_valued_column("val1")
val2 = tf.contrib.layers.real_valued_column("val2")
val3 = tf.contrib.layers.real_valued_column("val3")

wide_columns = [val1, val2, val3]

Problem portion working version

model_dir = tempfile.mkdtemp()
m = tf.contrib.learn.LinearClassifier(feature_columns=wide_columns, model_dir=model_dir)
m.fit(input_fn=train_input_fn, steps=200)

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

accuracy: 0.45
eval_auc: 0.459596
loss: 0.771354

Problem portion non working version

model_dir = tempfile.mkdtemp()
m = tf.contrib.learn.LinearClassifier(feature_columns=wide_columns, model_dir=model_dir)
m.fit(input_fn=train_input_fn, steps=200)
# 2 lines that are different ##########################
x, y = train_input_fn()
results = m.evaluate(x=x, y=y, batch_size=100, steps=1)
#######################################################
for key in sorted(results):
    print("%s: %s" % (key, results[key]))

Below is the error I get but I get a different error depending on what I try. The documentation says a matrix. I tried that too.

Dump of entire traceback

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-135-5b53add19aac> in <module>()
     12 # p.fit(input_fn=train_input_fn, steps=10)
     13 x, y = train_input_fn()
---> 14 p.fit(x=df_train, y=df_train, steps=10, batch_size=100)
     15 results = p.evaluate(input_fn=eval_input_fn, steps=1)
     16 for key in sorted(results):

/Users/sean/anaconda/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc in fit(self, x, y, input_fn, steps, batch_size, monitors)
    171       if x is None:
    172         raise ValueError('Either x or input_fn must be provided.')
--> 173       input_fn, feed_fn = _get_input_fn(x, y, batch_size)
    174     elif (x is not None) or (y is not None):
    175       raise ValueError('Can not provide both input_fn and either of x and y.')

/Users/sean/anaconda/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc in _get_input_fn(x, y, batch_size)
     65 def _get_input_fn(x, y, batch_size):
     66   df = data_feeder.setup_train_data_feeder(
---> 67       x, y, n_classes=None, batch_size=batch_size)
     68   return df.input_builder, df.get_feed_dict_fn()
     69 

/Users/sean/anaconda/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/io/data_feeder.pyc in setup_train_data_feeder(X, y, n_classes, batch_size, shuffle, epochs)
     97     ValueError: if one of `X` and `y` is iterable and the other is not.
     98   """
---> 99   X, y = _data_type_filter(X, y)
    100   if HAS_DASK:
    101     # pylint: disable=g-import-not-at-top

/Users/sean/anaconda/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/io/data_feeder.pyc in _data_type_filter(X, y)
     65       y = extract_dask_labels(y)
     66   if HAS_PANDAS:
---> 67     X = extract_pandas_data(X)
     68     if y is not None:
     69       y = extract_pandas_labels(y)

/Users/sean/anaconda/lib/python2.7/site-packages/tensorflow/contrib/learn/python/learn/io/pandas_io.pyc in extract_pandas_data(data)
     51     return data.values.astype('float')
     52   else:
---> 53     raise ValueError('Data types for data must be int, float, or bool.')
     54 
     55 

ValueError: Data types for data must be int, float, or bool.

Answer 1

It seems that the format if you pass x and y is not the same as input_fn . Quote from fit 's docstring :

x: matrix or tensor of shape [n_samples, n_features...]. Can be iterator that returns arrays of features. The training input samples for fitting the model. If set, input_fn must be None.

The example below works. Note that

• I had to substitute 'Yes' / 'No' by booleans (which is probably not equivalent, but illustrates the point) since seems one cannot enter sparse data this way.

• I used infer_real_valued_columns_from_input to get the columns.

Revised version:

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

np.random.seed([3,1415])

_x_df = pd.DataFrame(dict(
    cat1=np.random.choice((True, False), (100,),),
    val1=np.random.rand(100),
    val2=np.random.rand(100),
    val3=np.random.rand(100)))

_y_df = pd.DataFrame(dict(label=np.random.choice((0, 1), (100,))))

trainBegin, trainEnd = 0, 80
testBegin, testEnd = 80, 100
x_df_train = _x_df.iloc[trainBegin:trainEnd, :]
x_df_test = _x_df.iloc[testBegin:testEnd, :]
y_df_train = _y_df.iloc[trainBegin:trainEnd, :]
y_df_test = _y_df.iloc[testBegin:testEnd, :]

wide_columns = tf.contrib.learn.infer_real_valued_columns_from_input(x_df_train)

model_dir = tempfile.mkdtemp()
m = tf.contrib.learn.LinearClassifier(feature_columns=wide_columns, model_dir=model_dir)
m.fit(x_df_train, y_df_train, batch_size=5, steps=200)

results = m.evaluate(x_df_test, y_df_test, batch_size=5, steps=1)
for key in sorted(results):
    print("%s: %s" % (key, results[key]))