Example Spec

Create an Example spec used for parsing serialized Examples.

We'll cover the following

Chapter Goals:

  • Create the Example spec that’s used to parse serialized Example objects

A. Creating the Example spec

Since the data is now stored as serialized Example objects in TFRecords files, we need to create an Example spec, which allows us to parse the serialized Examples in the input pipeline. The Example spec gives specifications on each of the dataset’s features, specifically the shape and type of the feature’s values.

The Example spec is just a Python dictionary, mapping feature names to FixedLenFeature objects. For our dataset, each of the FixedLenFeature objects has shape (). This is because each feature contains a single value per data observation, i.e. each feature of our dataset’s Example objects contains one value (as opposed to an array of values).

The type of each FixedLenFeature depends on the type of values in the feature. Integer valued features will have tf.int64 type, float valued features will have tf.float32 type, and string valued features (in our case, only the 'Type' feature) will have tf.string type.

Press + to interact
import tensorflow as tf



example_spec = {}

example_spec['Store'] = tf.io.FixedLenFeature((), tf.int64)

example_spec['CPI'] = tf.io.FixedLenFeature((), tf.float32)

example_spec['Type'] = tf.io.FixedLenFeature((), tf.string)

Time to Code!

In this chapter you’ll be completing the create_example_spec function, which creates an Example spec based on the dataset features. The function is already filled with code that provides the feature names for the dataset.

For each of the integer features, we’ll create a FixedLenFeature object with shape () and type tf.int64.

Create a for loop that iterates through int_vals using a variable named feature_name.
Inside the for loop, set feature_name as a key in example_spec. The value it maps to will be tf.io.FixedLenFeature initialized with the specified shape and type.

We’ll now repeat the previous step for each of the float features. The shape of the FixedLenFeature objects is still (), but now the type is tf.float32.

Create another for loop (outside the previous one) that iterates through float_vals using a variable named feature_name.
Inside the for loop, set feature_name as a key in example_spec. The value it maps to will be tf.io.FixedLenFeature initialized with the specified shape and type.

Finally, we’ll create a FixedLenFeature for the only string feature ('Type'), which will have shape () and type tf.string.

Outside the for loops, set example_spec['Type'] to a tf.io.FixedLenFeature object initialized with the specified shape and type. Then return example_spec.

Press + to interact
import tensorflow as tf
# Create the spec used when parsing the Example object
def create_example_spec(has_labels):
    example_spec = {}
    int_vals = ['Store', 'Dept', 'IsHoliday', 'Size']
    float_vals = ['Temperature', 'Fuel_Price', 'CPI', 'Unemployment']
    if has_labels:
        float_vals.append('Weekly_Sales')
    # CODE HERE

Get hands-on with 1300+ tech skills courses.