Build TensorFlow input pipelines

View on Run in Google Colab View source on GitHub Download notebook

The API enables you to build complex input pipelines from simple, reusable pieces. For example, the pipeline for an image model might aggregate data from files in a distributed file system, apply random perturbations to each image, and merge randomly selected images into a batch for training. The pipeline for a text model might involve extracting symbols from raw text data, converting them to embedding identifiers with a lookup table, and batching together sequences of different lengths. The API makes it possible to handle large amounts of data, read from different data formats, and perform complex transformations.

The API introduces a abstraction that represents a sequence of elements, in which each element consists of one or more components. For example, in an image pipeline, an element might be a single training example, with a pair of tensor components representing the image and its label.

There are two distinct ways to create a dataset:

  • A data source constructs a Dataset from data stored in memory or in one or more files.

  • A data transformation constructs a dataset from one or more objects.

import tensorflow as tf
import pathlib
import os
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np


Basic mechanics

To create an input pipeline, you must start with a data source. For example, to construct a Dataset from data in memory, you can use or Alternatively, if your input data is stored in a file in the recommended TFRecord format, you can use

Once you have a Dataset object, you can transform it into a new Dataset by chaining method calls on the object. For example, you can apply per-element transformations such as, and multi-element transformations such as Dataset.batch. Refer to the documentation for for a complete list of transformations.

The Dataset object is a Python iterable. This makes it possible to consume its elements using a for loop:

dataset =[8, 3, 0, 8, 2, 1])
for elem in dataset:

Or by explicitly creating a Python iterator using iter and consuming its elements using next:

it = iter(dataset)


Alternatively, dataset elements can be consumed using the reduce transformation, which reduces all elements to produce a single result. The following example illustrates how to use the reduce transformation to compute the sum of a dataset of integers.

print(dataset.reduce(0, lambda state, value: state + value).numpy())

Dataset structure

A dataset produces a sequence of elements, where each element is the same (nested) structure of components. Individual components of the structure can be of any type representable by tf.TypeSpec, including tf.Tensor, tf.sparse.SparseTensor, tf.RaggedTensor, tf.TensorArray, or

The Python constructs that can be used to express the (nested) structure of elements include tuple, dict, NamedTuple, and OrderedDict. In particular, list is not a valid construct for expressing the structure of dataset elements. This is because early users felt strongly about list inputs (for example, when passed to being automatically packed as tensors and list outputs (for example, return values of user-defined functions) being coerced into a tuple. As a consequence, if you would like a list input to be treated as a structure, you need to convert it into tuple and if you would like a list output to be a single component, then you need to explicitly pack it using tf.stack.

The Dataset.element_spec property allows you to inspect the type of each element component. The property returns a nested structure of tf.TypeSpec objects, matching the structure of the element, which may be a single component, a tuple of components, or a nested tuple of components. For example:

dataset1 =[4, 10]))

dataset2 =
    tf.random.uniform([4, 100], maxval=100, dtype=tf.int32)))

dataset3 =, dataset2))

# Dataset containing a sparse tensor.
dataset4 =[[0, 0], [1, 2]], values=[1, 2], dense_shape=[3, 4]))

# Use value_type to see the type of value represented by the element spec

The Dataset transformations support datasets of any structure. When using the, and Dataset.filter transformations, which apply a function to each element, the element structure determines the arguments of the function:

dataset1 =
    tf.random.uniform([4, 10], minval=1, maxval=10, dtype=tf.int32))

for z in dataset1:
dataset2 =
    tf.random.uniform([4, 100], maxval=100, dtype=tf.int32)))

dataset3 =, dataset2))

for a, (b,c) in dataset3:
  print('shapes: {a.shape}, {b.shape}, {c.shape}'.format(a=a, b=b, c=c))

Reading input data

Consuming NumPy arrays

Refer to the Loading NumPy arrays tutorial for more examples.

If all of your input data fits in memory, the simplest way to create a Dataset from them is to convert them to tf.Tensor objects and use Dataset.from_tensor_slices.

train, test = tf.keras.datasets.fashion_mnist.load_data()
images, labels = train
images = images/255

dataset =, labels))

Consuming Python generators

Another common data source that can easily be ingested as a is the python generator.

def count(stop):
  i = 0
  while i<stop:
    yield i
    i += 1
for n in count(5):

The Dataset.from_generator constructor converts the python generator to a fully functional

The constructor takes a callable as input, not an iterator. This allows it to restart the generator when it reaches the end. It takes an optional args argument, which is passed as the callable's arguments.

The output_types argument is required because builds a tf.Graph internally, and graph edges require a tf.dtype.

ds_counter =, args=[25], output_types=tf.int32, output_shapes = (), )
for count_batch in ds_counter.repeat().batch(10).take(10):

The output_shapes argument is not required but is highly recommended as many TensorFlow operations do not support tensors with an unknown rank. If the length of a particular axis is unknown or variable, set it as None in the output_shapes.

It's also important to note that the output_shapes and output_types follow the same nesting rules as other dataset methods.

Here is an example generator that demonstrates both aspects: it returns tuples of arrays, where the second array is a vector with unknown length.

def gen_series():
  i = 0
  while True:
    size = np.random.randint(0, 10)
    yield i, np.random.normal(size=(size,))
    i += 1
for i, series in gen_series():
  print(i, ":", str(series))
  if i > 5:

The first output is an int32 the second is a float32.

The first item is a scalar, shape (), and the second is a vector of unknown length, shape (None,)

ds_series =
    output_types=(tf.int32, tf.float32), 
    output_shapes=((), (None,)))


Now it can be used like a regular Note that when batching a dataset with a variable shape, you need to use Dataset.padded_batch.

ds_series_batch = ds_series.shuffle(20).padded_batch(10)

ids, sequence_batch = next(iter(ds_series_batch))

For a more realistic example, try wrapping preprocessing.image.ImageDataGenerator as a

First download the data:

flowers = tf.keras.utils.get_file(

Create the image.ImageDataGenerator

img_gen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1./255, rotation_range=20)
images, labels = next(img_gen.flow_from_directory(flowers))
print(images.dtype, images.shape)
print(labels.dtype, labels.shape)
ds =
    lambda: img_gen.flow_from_directory(flowers), 
    output_types=(tf.float32, tf.float32), 
    output_shapes=([32,256,256,3], [32,5])

for images, labels in ds.take(1):
  print('images.shape: ', images.shape)
  print('labels.shape: ', labels.shape)

Consuming TFRecord data

Refer to the Loading TFRecords tutorial for an end-to-end example.

The API supports a variety of file formats so that you can process large datasets that do not fit in memory. For example, the TFRecord file format is a simple record-oriented binary format that many TensorFlow applications use for training data. The class enables you to stream over the contents of one or more TFRecord files as part of an input pipeline.

Here is an example using the test file from the French Street Name Signs (FSNS).

# Creates a dataset that reads all of the examples from two files.
fsns_test_file = tf.keras.utils.get_file("fsns.tfrec", "")

The filenames argument to the TFRecordDataset initializer can either be a string, a list of strings, or a tf.Tensor of strings. Therefore if you have two sets of files for training and validation purposes, you can create a factory method that produces the dataset, taking filenames as an input argument:

dataset = = [fsns_test_file])

Many TensorFlow projects use serialized tf.train.Example records in their TFRecord files. These need to be decoded before they can be inspected:

raw_example = next(iter(dataset))
parsed = tf.train.Example.FromString(raw_example.numpy())


Consuming text data

Refer to the Load text tutorial for an end-to-end example.

Many datasets are distributed as one or more text files. The provides an easy way to extract lines from one or more text files. Given one or more filenames, a TextLineDataset will produce one string-valued element per line of those files.

directory_url = ''
file_names = ['cowper.txt', 'derby.txt', 'butler.txt']

file_paths = [
    tf.keras.utils.get_file(file_name, directory_url + file_name)
    for file_name in file_names
dataset =

Here are the first few lines of the first file:

for line in dataset.take(5):

To alternate lines between files use Dataset.interleave. This makes it easier to shuffle files together. Here are the first, second and third lines from each translation:

files_ds =
lines_ds = files_ds.interleave(, cycle_length=3)

for i, line in enumerate(lines_ds.take(9)):
  if i % 3 == 0:

By default, a TextLineDataset yields every line of each file, which may not be desirable, for example, if the file starts with a header line, or contains comments. These lines can be removed using the Dataset.skip() or Dataset.filter transformations. Here, you skip the first line, then filter to find only survivors.

titanic_file = tf.keras.utils.get_file("train.csv", "")
titanic_lines =
for line in titanic_lines.take(10):
def survived(line):
  return tf.not_equal(tf.strings.substr(line, 0, 1), "0")

survivors = titanic_lines.skip(1).filter(survived)
for line in survivors.take(10):

Consuming CSV data

Refer to the Loading CSV Files and Loading Pandas DataFrames tutorials for more examples.

The CSV file format is a popular format for storing tabular data in plain text.

For example:

titanic_file = tf.keras.utils.get_file("train.csv", "")
df = pd.read_csv(titanic_file)

If your data fits in memory the same Dataset.from_tensor_slices method works on dictionaries, allowing this data to be easily imported:

titanic_slices =

for feature_batch in titanic_slices.take(1):
  for key, value in feature_batch.items():
    print("  {!r:20s}: {}".format(key, value))

A more scalable approach is to load from disk as necessary.

The module provides methods to extract records from one or more CSV files that comply with RFC 4180.

The function is the high-level interface for reading sets of CSV files. It supports column type inference and many other features, like batching and shuffling, to make usage simple.

titanic_batches =
    titanic_file, batch_size=4,
for feature_batch, label_batch in titanic_batches.take(1):
  print("'survived': {}".format(label_batch))
  for key, value in feature_batch.items():
    print("  {!r:20s}: {}".format(key, value))

You can use the select_columns argument if you only need a subset of columns.

titanic_batches =
    titanic_file, batch_size=4,
    label_name="survived", select_columns=['class', 'fare', 'survived'])
for feature_batch, label_batch in titanic_batches.take(1):
  print("'survived': {}".format(label_batch))
  for key, value in feature_batch.items():
    print("  {!r:20s}: {}".format(key, value))

There is also a lower-level experimental.CsvDataset class which provides finer grained control. It does not support column type inference. Instead you must specify the type of each column.

titanic_types  = [tf.int32, tf.string, tf.float32, tf.int32, tf.int32, tf.float32, tf.string, tf.string, tf.string, tf.string]
dataset =, titanic_types , header=True)

for line in dataset.take(10):
  print([item.numpy() for item in line])

If some columns are empty, this low-level interface allows you to provide default values instead of column types.

%%writefile missing.csv
# Creates a dataset that reads all of the records from two CSV files, each with
# four float columns which may have missing values.

record_defaults = [999,999,999,999]
dataset ="missing.csv", record_defaults)
dataset = *items: tf.stack(items))
for line in dataset:

By default, a CsvDataset yields every column of every line of the file, which may not be desirable, for example if the file starts with a header line that should be ignored, or if some columns are not required in the input. These lines and fields can be removed with the header and select_cols arguments respectively.

# Creates a dataset that reads all of the records from two CSV files with
# headers, extracting float data from columns 2 and 4.
record_defaults = [999, 999] # Only provide defaults for the selected columns
dataset ="missing.csv", record_defaults, select_cols=[1, 3])
dataset = *items: tf.stack(items))
for line in dataset:

Consuming sets of files

There are many datasets distributed as a set of files, where each file is an example.

flowers_root = tf.keras.utils.get_file(
flowers_root = pathlib.Path(flowers_root)

The root directory contains a directory for each class:

for item in flowers_root.glob("*"):

The files in each class directory are examples:

list_ds ='*/*'))

for f in list_ds.take(5):

Read the data using the function and extract the label from the path, returning (image, label) pairs:

def process_path(file_path):
  label = tf.strings.split(file_path, os.sep)[-2]
  return, label

labeled_ds =
for image_raw, label_text in labeled_ds.take(1):

Batching dataset elements

Simple batching

The simplest form of batching stacks n consecutive elements of a dataset into a single element. The Dataset.batch() transformation does exactly this, with the same constraints as the tf.stack() operator, applied to each component of the elements: i.e. for each component i, all elements must have a tensor of the exact same shape.

inc_dataset =
dec_dataset =, -100, -1)
dataset =, dec_dataset))
batched_dataset = dataset.batch(4)

for batch in batched_dataset.take(4):
  print([arr.numpy() for arr in batch])

While tries to propagate shape information, the default settings of Dataset.batch result in an unknown batch size because the last batch may not be full. Note the Nones in the shape:


Use the drop_remainder argument to ignore that last batch, and get full shape propagation:

batched_dataset = dataset.batch(7, drop_remainder=True)

Batching tensors with padding

The above recipe works for tensors that all have the same size. However, many models (including sequence models) work with input data that can have varying size (for example, sequences of different lengths). To handle this case, the Dataset.padded_batch transformation enables you to batch tensors of different shapes by specifying one or more dimensions in which they may be padded.

dataset =
dataset = x: tf.fill([tf.cast(x, tf.int32)], x))
dataset = dataset.padded_batch(4, padded_shapes=(None,))

for batch in dataset.take(2):

The Dataset.padded_batch transformation allows you to set different padding for each dimension of each component, and it may be variable-length (signified by None in the example above) or constant-length. It is also possible to override the padding value, which defaults to 0.

Training workflows

Processing multiple epochs

The API offers two main ways to process multiple epochs of the same data.

The simplest way to iterate over a dataset in multiple epochs is to use the Dataset.repeat() transformation. First, create a dataset of titanic data:

titanic_file = tf.keras.utils.get_file("train.csv", "")
titanic_lines =
def plot_batch_sizes(ds):
  batch_sizes = [batch.shape[0] for batch in ds], batch_sizes)
  plt.xlabel('Batch number')
  plt.ylabel('Batch size')

Applying the Dataset.repeat() transformation with no arguments will repeat the input indefinitely.

The Dataset.repeat transformation concatenates its arguments without signaling the end of one epoch and the beginning of the next epoch. Because of this a Dataset.batch applied after Dataset.repeat will yield batches that straddle epoch boundaries:

titanic_batches = titanic_lines.repeat(3).batch(128)

If you need clear epoch separation, put Dataset.batch before the repeat:

titanic_batches = titanic_lines.batch(128).repeat(3)


If you would like to perform a custom computation (for example, to collect statistics) at the end of each epoch then it's simplest to restart the dataset iteration on each epoch:

epochs = 3
dataset = titanic_lines.batch(128)

for epoch in range(epochs):
  for batch in dataset:
  print("End of epoch: ", epoch)

Randomly shuffling input data

The Dataset.shuffle() transformation maintains a fixed-size buffer and chooses the next element uniformly at random from that buffer.

Add an index to the dataset so you can see the effect:

lines =
counter =

dataset =, lines))
dataset = dataset.shuffle(buffer_size=100)
dataset = dataset.batch(20)

Since the buffer_size is 100, and the batch size is 20, the first batch contains no elements with an index over 120.

n,line_batch = next(iter(dataset))

As with Dataset.batch the order relative to Dataset.repeat matters.

Dataset.shuffle doesn't signal the end of an epoch until the shuffle buffer is empty. So a shuffle placed before a repeat will show every element of one epoch before moving to the next:

dataset =, lines))
shuffled = dataset.shuffle(buffer_size=100).batch(10).repeat(2)

print("Here are the item ID's near the epoch boundary:\n")
for n, line_batch in shuffled.skip(60).take(5):
shuffle_repeat = [n.numpy().mean() for n, line_batch in shuffled]
plt.plot(shuffle_repeat, label="shuffle().repeat()")
plt.ylabel("Mean item ID")

But a repeat before a shuffle mixes the epoch boundaries together:

dataset =, lines))
shuffled = dataset.repeat(2).shuffle(buffer_size=100).batch(10)

print("Here are the item ID's near the epoch boundary:\n")
for n, line_batch in shuffled.skip(55).take(15):
repeat_shuffle = [n.numpy().mean() for n, line_batch in shuffled]

plt.plot(shuffle_repeat, label="shuffle().repeat()")
plt.plot(repeat_shuffle, label="repeat().shuffle()")
plt.ylabel("Mean item ID")

Preprocessing data

The transformation produces a new dataset by applying a given function f to each element of the input dataset. It is based on the map() function that is commonly applied to lists (and other structures) in functional programming languages. The function f takes the tf.Tensor objects that represent a single element in the input, and returns the tf.Tensor objects that will represent a single element in the new dataset. Its implementation uses standard TensorFlow operations to transform one element into another.

This section covers common examples of how to use

Decoding image data and resizing it

When training a neural network on real-world image data, it is often necessary to convert images of different sizes to a common size, so that they may be batched into a fixed size.

Rebuild the flower filenames dataset:

list_ds ='*/*'))

Write a function that manipulates the dataset elements.

# Reads an image from a file, decodes it into a dense tensor, and resizes it
# to a fixed shape.
def parse_image(filename):
  parts = tf.strings.split(filename, os.sep)
  label = parts[-2]

  image =
  image =
  image = tf.image.convert_image_dtype(image, tf.float32)
  image = tf.image.resize(image, [128, 128])
  return image, label

Test that it works.

file_path = next(iter(list_ds))
image, label = parse_image(file_path)

def show(image, label):

show(image, label)

Map it over the dataset.

images_ds =

for image, label in images_ds.take(2):
  show(image, label)

Applying arbitrary Python logic

For performance reasons, use TensorFlow operations for preprocessing your data whenever possible. However, it is sometimes useful to call external Python libraries when parsing your input data. You can use the tf.py_function operation in a transformation.

For example, if you want to apply a random rotation, the tf.image module only has tf.image.rot90, which is not very useful for image augmentation.

To demonstrate tf.py_function, try using the scipy.ndimage.rotate function instead:

import scipy.ndimage as ndimage

def random_rotate_image(image):
  image = ndimage.rotate(image, np.random.uniform(-30, 30), reshape=False)
  return image
image, label = next(iter(images_ds))
image = random_rotate_image(image)
show(image, label)

To use this function with the same caveats apply as with Dataset.from_generator, you need to describe the return shapes and types when you apply the function:

def tf_random_rotate_image(image, label):
  im_shape = image.shape
  [image,] = tf.py_function(random_rotate_image, [image], [tf.float32])
  return image, label
rot_ds =

for image, label in rot_ds.take(2):
  show(image, label)

Parsing tf.Example protocol buffer messages

Many input pipelines extract tf.train.Example protocol buffer messages from a TFRecord format. Each tf.train.Example record contains one or more "features", and the input pipeline typically converts these features into tensors.

fsns_test_file = tf.keras.utils.get_file("fsns.tfrec", "")
dataset = = [fsns_test_file])

You can work with tf.train.Example protos outside of a to understand the data:

raw_example = next(iter(dataset))
parsed = tf.train.Example.FromString(raw_example.numpy())

feature = parsed.features.feature
raw_img = feature['image/encoded'].bytes_list.value[0]
img = tf.image.decode_png(raw_img)
_ = plt.title(feature["image/text"].bytes_list.value[0])
raw_example = next(iter(dataset))
def tf_parse(eg):
  example =
      eg[tf.newaxis], {
          'image/encoded':, dtype=tf.string),
          'image/text':, dtype=tf.string)
  return example['image/encoded'][0], example['image/text'][0]
img, txt = tf_parse(raw_example)
print(repr(img.numpy()[:20]), "...")
decoded =
image_batch, text_batch = next(iter(decoded.batch(10)))

Time series windowing

For an end-to-end time series example see: Time series forecasting.

Time series data is often organized with the time axis intact.

Use a simple Dataset.range to demonstrate:

range_ds =

Typically, models based on this sort of data will want a contiguous time slice.

The simplest approach would be to batch the data:

Using batch

batches = range_ds.batch(10, drop_remainder=True)

for batch in batches.take(5):

Or to make dense predictions one step into the future, you might shift the features and labels by one step relative to each other:

def dense_1_step(batch):
  # Shift features and labels one step relative to each other.
  return batch[:-1], batch[1:]

predict_dense_1_step =

for features, label in predict_dense_1_step.take(3):
  print(features.numpy(), " => ", label.numpy())

To predict a whole window instead of a fixed offset you can split the batches into two parts:

batches = range_ds.batch(15, drop_remainder=True)

def label_next_5_steps(batch):
  return (batch[:-5],   # Inputs: All except the last 5 steps
          batch[-5:])   # Labels: The last 5 steps

predict_5_steps =

for features, label in predict_5_steps.take(3):
  print(features.numpy(), " => ", label.numpy())

To allow some overlap between the features of one batch and the labels of another, use

feature_length = 10
label_length = 3

features = range_ds.batch(feature_length, drop_remainder=True)
labels = range_ds.batch(feature_length).skip(1).map(lambda labels: labels[:label_length])

predicted_steps =, labels))

for features, label in predicted_steps.take(5):
  print(features.numpy(), " => ", label.numpy())

Using window

While using Dataset.batch works, there are situations where you may need finer control. The Dataset.window method gives you complete control, but requires some care: it returns a Dataset of Datasets. Go to the Dataset structure section for details.

window_size = 5

windows = range_ds.window(window_size, shift=1)
for sub_ds in windows.take(5):

The Dataset.flat_map method can take a dataset of datasets and flatten it into a single dataset:

for x in windows.flat_map(lambda x: x).take(30):
   print(x.numpy(), end=' ')

In nearly all cases, you will want to Dataset.batch the dataset first:

def sub_to_batch(sub):
  return sub.batch(window_size, drop_remainder=True)

for example in windows.flat_map(sub_to_batch).take(5):

Now, you can see that the shift argument controls how much each window moves over.

Putting this together you might write this function:

def make_window_dataset(ds, window_size=5, shift=1, stride=1):
  windows = ds.window(window_size, shift=shift, stride=stride)

  def sub_to_batch(sub):
    return sub.batch(window_size, drop_remainder=True)

  windows = windows.flat_map(sub_to_batch)
  return windows
ds = make_window_dataset(range_ds, window_size=10, shift = 5, stride=3)

for example in ds.take(10):

Then it's easy to extract labels, as before:

dense_labels_ds =

for inputs,labels in dense_labels_ds.take(3):
  print(inputs.numpy(), "=>", labels.numpy())


When working with a dataset that is very class-imbalanced, you may want to resample the dataset. provides two methods to do this. The credit card fraud dataset is a good example of this sort of problem.

zip_path = tf.keras.utils.get_file(

csv_path = zip_path.replace('.zip', '.csv')
creditcard_ds =
    csv_path, batch_size=1024, label_name="Class",
    # Set the column types: 30 floats and an int.

Now, check the distribution of classes, it is highly skewed:

def count(counts, batch):
  features, labels = batch
  class_1 = labels == 1
  class_1 = tf.cast(class_1, tf.int32)

  class_0 = labels == 0
  class_0 = tf.cast(class_0, tf.int32)

  counts['class_0'] += tf.reduce_sum(class_0)
  counts['class_1'] += tf.reduce_sum(class_1)

  return counts
counts = creditcard_ds.take(10).reduce(
    initial_state={'class_0': 0, 'class_1': 0},
    reduce_func = count)

counts = np.array([counts['class_0'].numpy(),

fractions = counts/counts.sum()

A common approach to training with an imbalanced dataset is to balance it. includes a few methods which enable this workflow:

Datasets sampling

One approach to resampling a dataset is to use sample_from_datasets. This is more applicable when you have a separate for each class.

Here, just use filter to generate them from the credit card fraud data:

negative_ds = (
    .filter(lambda features, label: label==0)
positive_ds = (
    .filter(lambda features, label: label==1)
for features, label in positive_ds.batch(10).take(1):

To use pass the datasets, and the weight for each:

balanced_ds =
    [negative_ds, positive_ds], [0.5, 0.5]).batch(10)

Now the dataset produces examples of each class with a 50/50 probability:

for features, labels in balanced_ds.take(10):

Rejection resampling

One problem with the above Dataset.sample_from_datasets approach is that it needs a separate per class. You could use Dataset.filter to create those two datasets, but that results in all the data being loaded twice.

The method can be applied to a dataset to rebalance it, while only loading it once. Elements will be dropped or repeated to achieve balance.

The rejection_resample method takes a class_func argument. This class_func is applied to each dataset element, and is used to determine which class an example belongs to for the purposes of balancing.

The goal here is to balance the label distribution, and the elements of creditcard_ds are already (features, label) pairs. So the class_func just needs to return those labels:

def class_func(features, label):
  return label

The resampling method deals with individual examples, so in this case you must unbatch the dataset before applying that method.

The method needs a target distribution, and optionally an initial distribution estimate as inputs.

resample_ds = (
    .rejection_resample(class_func, target_dist=[0.5,0.5],

The rejection_resample method returns (class, example) pairs where the class is the output of the class_func. In this case, the example was already a (feature, label) pair, so use map to drop the extra copy of the labels:

balanced_ds = extra_label, features_and_label: features_and_label)

Now the dataset produces examples of each class with a 50/50 probability:

for features, labels in balanced_ds.take(10):

Iterator Checkpointing

Tensorflow supports taking checkpoints so that when your training process restarts it can restore the latest checkpoint to recover most of its progress. In addition to checkpointing the model variables, you can also checkpoint the progress of the dataset iterator. This could be useful if you have a large dataset and don't want to start the dataset from the beginning on each restart. Note however that iterator checkpoints may be large, since transformations such as Dataset.shuffle and Dataset.prefetch require buffering elements within the iterator.

To include your iterator in a checkpoint, pass the iterator to the tf.train.Checkpoint constructor.

range_ds =

iterator = iter(range_ds)
ckpt = tf.train.Checkpoint(step=tf.Variable(0), iterator=iterator)
manager = tf.train.CheckpointManager(ckpt, '/tmp/my_ckpt', max_to_keep=3)

print([next(iterator).numpy() for _ in range(5)])

save_path =

print([next(iterator).numpy() for _ in range(5)])


print([next(iterator).numpy() for _ in range(5)])

Using with tf.keras

The tf.keras API simplifies many aspects of creating and executing machine learning models. Its and Model.evaluate and Model.predict APIs support datasets as inputs. Here is a quick dataset and model setup:

train, test = tf.keras.datasets.fashion_mnist.load_data()

images, labels = train
images = images/255.0
labels = labels.astype(np.int32)
fmnist_train_ds =, labels))
fmnist_train_ds = fmnist_train_ds.shuffle(5000).batch(32)

model = tf.keras.Sequential([


Passing a dataset of (feature, label) pairs is all that's needed for and Model.evaluate:, epochs=2)

If you pass an infinite dataset, for example by calling Dataset.repeat, you just need to also pass the steps_per_epoch argument:, epochs=2, steps_per_epoch=20)

For evaluation you can pass the number of evaluation steps:

loss, accuracy = model.evaluate(fmnist_train_ds)
print("Loss :", loss)
print("Accuracy :", accuracy)

For long datasets, set the number of steps to evaluate:

loss, accuracy = model.evaluate(fmnist_train_ds.repeat(), steps=10)
print("Loss :", loss)
print("Accuracy :", accuracy)

The labels are not required when calling Model.predict.

predict_ds =
result = model.predict(predict_ds, steps = 10)

But the labels are ignored if you do pass a dataset containing them:

result = model.predict(fmnist_train_ds, steps = 10)