Having efficient data pipelines is of paramount importance for any machine learning model. In this blog, we will learn how to use TensorFlow’s Dataset module tf.data to build efficient data pipelines.

Motivation

Most of the introductory articles on TensorFlow would introduce you with the feed_dict method of feeding the data to the model. feed_dict processes the input data in a single thread and while the data is being loaded and processed on CPU, the GPU remains idle and when the GPU is training a batch of data, CPU remains in the idle state. The developers of TensorFlow have advised not to use this method during training or repeated validation of the same datasets.

tf_data improves the performance by prefetching the next batch of data asynchronously so that GPU need not wait for the data. You can also parallelize the process of preprocessing and loading the dataset.

In this blog post we will cover Datasets and Iterators. We will learn how to create Datasets from source data, apply the transformation to Dataset and then consume the data using Iterators.

How to create Datasets?

Tensorflow provides various methods to create Datasets from numpy arrays, text files, CSV files, tensors, etc. Let’s look at few methods below

• from_tensor_slices: It accepts single or multiple numpy arrays or tensors. Dataset created using this method will emit only one data at a time.

# source data - numpy array
data = np.arange(10)
# create a dataset from numpy array
dataset = tf.data.Dataset.from_tensor_slices(data)

The object dataset is a tensorflow Dataset object.

• from_tensors: It also accepts single or multiple numpy arrays or tensors. Dataset created using this method will emit all the data at once.

data = tf.arange(10)
dataset = tf.data.Dataset.from_tensors(data)

3. from_generator: Creates a Dataset whose elements are generated by a function.

def generator():
  for i in range(10):
    yield 2*i
dataset = tf.data.Dataset.from_generator(generator, (tf.int32))

Operations on Datasets

• Batches: Combines consecutive elements of the Dataset into a single batch. Useful when you want to train smaller batches of data to avoid out of memory errors.

data = np.arange(10,40)

create batches of 10
dataset = tf.data.Dataset.from_tensor_slices(data).batch(10)
creates the iterator to consume the data
iterator = dataset.make_one_shot_iterator()

next_ele = iterator.get_next()

with tf.Session() as sess:

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass

You can skip the code where we create an iterator and print the elements of the Dataset. We will learn about Iterators in detail in the later part of this blog. The output is :

[10 11 12 13 14 15 16 17 18 19]

[20 21 22 23 24 25 26 27 28 29]

[30 31 32 33 34 35 36 37 38 39]

• Zip: Creates a Dataset by zipping together datasets. Useful in scenarios where you have features and labels and you need to provide the pair of feature and label for training the model.

datax = np.arange(10,20)

datay = np.arange(11,21)

datasetx = tf.data.Dataset.from_tensor_slices(datax)

datasety = tf.data.Dataset.from_tensor_slices(datay)

dcombined = tf.data.Dataset.zip((datasetx, datasety)).batch(2)

iterator = dcombined.make_one_shot_iterator()

next_ele = iterator.get_next()

with tf.Session() as sess:

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass

The output is

(array([10, 11]), array([11, 12]))

(array([12, 13]), array([13, 14]))

(array([14, 15]), array([15, 16]))

(array([16, 17]), array([17, 18]))

(array([18, 19]), array([19, 20]))

• Repeat: Used to repeat the Dataset.

dataset = tf.data.Dataset.from_tensor_slices(tf.range(10))

dataset = dataset.repeat(count=2)

iterator = dataset.make_one_shot_iterator()

next_ele = iterator.get_next()

with tf.Session() as sess:

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass

The output is

0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9

• Map: Used to transform the elements of the Dataset. Useful in cases where you want to transform your raw data before feeding into the model.

def map_fnc(x):

return x*2;

data = np.arange(10)

dataset = tf.data.Dataset.from_tensor_slices(data)

dataset = dataset.map(map_fnc)

iterator = dataset.make_one_shot_iterator()

next_ele = iterator.get_next()

with tf.Session() as sess:

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass

The output is

0 2 4 6 8 10 12 14 16 18

Creating Iterators

We have learned various ways to create Datasets and apply various transformations to it, but how do we consume the data? Tensorflow provides Iterators to do that.

The iterator is not aware of the number of elements present in the Dataset. It has a get_next function that is used to create an operation in the tensorflow graph and when run over a session, it will return the values from the iterator. Once the Dataset is exhausted, it throws an tf.errors.OutOfRangeError exception.

Let’s look at various Iterators that TensorFlow provides.

• One-shot iterator: This is the most basic form of iterator. It requires no explicit initialization and iterates over the data only one time and once it gets exhausted, it cannot be re-initialized.

data = np.arange(10,15)

#create the dataset

dataset = tf.data.Dataset.from_tensor_slices(data)

#create the iterator

iterator = dataset.make_one_shot_iterator()

next_element = iterator.get_next()

with tf.Session() as sess:

val = sess.run(next_element)

print(val)

• Initializable iterator: This iterator requires you to explicitly initialize the iterator by running iterator.initialize. You can define a tf.placeholder and pass data to it dynamically each time you call the initialize operation.

# define two placeholders to accept min and max value

min_val = tf.placeholder(tf.int32, shape=[])

max_val = tf.placeholder(tf.int32, shape=[])

data = tf.range(min_val, max_val)

dataset = tf.data.Dataset.from_tensor_slices(data)

iterator = dataset.make_initializable_iterator()

next_ele = iterator.get_next()

with tf.Session() as sess:

# initialize an iterator with range of values from 10 to 15

sess.run(iterator.initializer, feed_dict={min_val:10, max_val:15})

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass
# initialize an iterator with range of values from 1 to 10

sess.run(iterator.initializer, feed_dict={min_val:1, max_val:10})

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass

• Reinitializable iterator: This iterator can be initialized from different Dataset objects that have the same structure. Each dataset can pass through its own transformation pipeline.

def map_fnc(ele):

return ele*2

min_val = tf.placeholder(tf.int32, shape=[])

max_val = tf.placeholder(tf.int32, shape=[])

data = tf.range(min_val, max_val)

#Define separate datasets for training and validation

train_dataset =  tf.data.Dataset.from_tensor_slices(data)

val_dataset = tf.data.Dataset.from_tensor_slices(data).map(map_fnc)

#create an iterator

iterator=tf.data.Iterator.from_structure(train_dataset.output_types    ,train_dataset.output_shapes)

train_initializer = iterator.make_initializer(train_dataset)

val_initializer = iterator.make_initializer(val_dataset)

next_ele = iterator.get_next()

with tf.Session() as sess:

initialize an iterator with range of values from 10 to 15
sess.run(train_initializer, feed_dict={min_val:10, max_val:15})

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass
initialize an iterator with range of values from 1 to 10
sess.run(val_initializer, feed_dict={min_val:1, max_val:10})

try:

while True:

val = sess.run(next_ele)

print(val)

except tf.errors.OutOfRangeError:

pass

• Feedable iterator: Can be used to switch between Iterators for different Datasets. Useful when you have different Datasets and you want to have more control over which iterator to use over the Dataset.

def map_fnc(ele):

return ele*2

min_val = tf.placeholder(tf.int32, shape=[])

max_val = tf.placeholder(tf.int32, shape=[])

data = tf.range(min_val, max_val)

train_dataset = tf.data.Dataset.from_tensor_slices(data)

val_dataset = tf.data.Dataset.from_tensor_slices(data).map(map_fnc)

train_val_iterator = tf.data.Iterator.from_structure(train_dataset.output_types , train_dataset.output_shapes)

train_initializer = train_val_iterator.make_initializer(train_dataset)

val_initializer = train_val_iterator.make_initializer(val_dataset)

test_dataset = tf.data.Dataset.from_tensor_slices(tf.range(10,15))

test_iterator = test_dataset.make_one_shot_iterator()

handle = tf.placeholder(tf.string, shape=[])

iterator = tf.data.Iterator.from_string_handle(handle, train_dataset.output_types, train_dataset.output_shapes)

next_ele = iterator.get_next()

with tf.Session() as sess:

train_val_handle = sess.run(train_val_iterator.string_handle())

test_handle = sess.run(test_iterator.string_handle())
training
sess.run(train_initializer, feed_dict={min_val:10, max_val:15})

try:

while True:

val = sess.run(next_ele, feed_dict={handle:train_val_handle})

print(val)

except tf.errors.OutOfRangeError:

pass
#validation

sess.run(val_initializer, feed_dict={min_val:1, max_val:10})

try:

while True:

val = sess.run(next_ele, feed_dict={handle:train_val_handle})

print(val)

except tf.errors.OutOfRangeError:

pass
#testing

try:

while True:

val = sess.run(next_ele, feed_dict={handle:test_handle})

print(val)

except tf.errors.OutOfRangeError:

pass

We have learned about various iterators. Let’s apply the knowledge to a practical dataset. We will train the famous MNIST dataset using the LeNet-5 Model. This tutorial will not dive into the details of implementing the LeNet-5 Model as it is beyond the scope of this article.

LeNet-5 Model

Let’s import the MNIST data from the tensorflow library. The MNIST database contains 60,000 training images and 10,000 testing images. Each image is of size 28281. We need to resize it to 32321 for the LeNet-5 Model.

from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets(“MNIST_data/”, reshape=False, one_hot = True)

X_train, y_train = mnist.train.images, mnist.train.labels

X_val, y_val = mnist.validation.images, mnist.validation.labels

X_test, y_test = mnist.test.images, mnist.test.labels

X_train = np.pad(X_train, ((0,0), (2,2), (2,2), (0,0)), ‘constant’)

X_val =   np.pad(X_val, ((0,0), (2,2), (2,2), (0,0)), ‘constant’)

X_test =  np.pad(X_test, ((0,0), (2,2), (2,2), (0,0)), ‘constant’)

Let’s define the forward propagation of the model.

def forward_pass(X):

W1 = tf.get_variable(“W1”, [5,5,1,6], initializer = tf.contrib.layers.xavier_initializer(seed=0))

# for conv layer2

W2 = tf.get_variable(“W2”, [5,5,6,16], initializer = tf.contrib.layers.xavier_initializer(seed=0))

Z1 = tf.nn.conv2d(X, W1, strides = [1,1,1,1], padding=‘VALID’)

A1 = tf.nn.relu(Z1)

P1 = tf.nn.max_pool(A1, ksize = [1,2,2,1], strides = [1,2,2,1], padding=‘VALID’)

Z2 = tf.nn.conv2d(P1, W2, strides = [1,1,1,1], padding=‘VALID’)

A2= tf.nn.relu(Z2)

P2= tf.nn.max_pool(A2, ksize = [1,2,2,1], strides=[1,2,2,1], padding=‘VALID’)

P2 = tf.contrib.layers.flatten(P2)

Z3 = tf.contrib.layers.fully_connected(P2, 120)
Z4 = tf.contrib.layers.fully_connected(Z3, 84)
Z5 = tf.contrib.layers.fully_connected(Z4,10, activation_fn= None)
return Z5

Let’s define the model operations

def model(X,Y):

logits = forward_pass(X)
cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=Y))
optimizer = tf.train.AdamOptimizer(learning_rate=0.0009)
learner = optimizer.minimize(cost)
correct_predictions = tf.equal(tf.argmax(logits,1),   tf.argmax(Y,1))
accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32))

return (learner, accuracy)

We have now created the model. Before deciding on the Iterator to use for our model, let’s see what are the typical requirements of a machine learning model.

1. Training the data over batches: Dataset can be very huge. To prevent out of memory errors, we would need to train our dataset in small batches.
2. Train the model over n passes of the dataset: Typically you want to run your training model over multiple passes of the dataset.
3. Validate the model at each epoch: You would need to validate your model at each epoch to check your model’s performance.
4. Finally, test your model on unseen data: After the model is trained, you would like to test your model on unseen data.

Let’s see the pros and cons of each iterator.

• One-shot iterator: The Dataset can’t be reinitialized once exhausted. To train for more epochs, you would need to repeat the Dataset before feeding to the iterator. This will require huge memory if the size of the data is large. It also doesn’t provide any option to validate the model.

epochs = 10

batch_size = 64

iterations = len(y_train) * epochs

tf.reset_default_graph()

dataset = tf.data.Dataset.from_tensor_slices((X_train, y_train))

need to repeat the dataset for epoch number of times, as all the data needs
to be fed to the dataset at once
dataset = dataset.repeat(epochs).batch(batch_size)

iterator = dataset.make_one_shot_iterator()

X_batch , Y_batch = iterator.get_next()

(learner, accuracy) = model(X_batch, Y_batch)

with tf.Session() as sess:

sess.run(tf.global_variables_initializer())
total_accuracy = 0

try:

while True:

temp_accuracy, _ = sess.run([accuracy, learner])

total_accuracy += temp_accuracy
except tf.errors.OutOfRangeError:

pass
print(‘Avg training accuracy is {}’.format((total_accuracy * batch_size) / iterations ))

• Initializable iterator: You can dynamically change the Dataset between training and validation Datasets. However, in this case both the Datasets needs to go through the same transformation pipeline.

epochs = 10

batch_size = 64

tf.reset_default_graph()

X_data = tf.placeholder(tf.float32, [None, 32,32,1])

Y_data = tf.placeholder(tf.float32, [None, 10])

dataset = tf.data.Dataset.from_tensor_slices((X_data, Y_data))

dataset = dataset.batch(batch_size)

iterator = dataset.make_initializable_iterator()

X_batch , Y_batch = iterator.get_next()

(learner, accuracy) = model(X_batch, Y_batch)

with tf.Session() as sess:

sess.run(tf.global_variables_initializer())

for epoch in range(epochs):

# train the model
sess.run(iterator.initializer, feed_dict={X_data:X_train, Y_data:y_train})
total_train_accuracy = 0
no_train_examples = len(y_train)
try:
  while True:
    temp_train_accuracy, _ = sess.run([accuracy, learner])
    total_train_accuracy += temp_train_accuracy*batch_size
except tf.errors.OutOfRangeError:
  pass

# validate the model
sess.run(iterator.initializer, feed_dict={X_data:X_val, Y_data:y_val})
total_val_accuracy = 0
no_val_examples = len(y_val)
try:
  while True:
    temp_val_accuracy = sess.run(accuracy)
    total_val_accuracy += temp_val_accuracy*batch_size
except tf.errors.OutOfRangeError:
  pass

print('Epoch {}'.format(str(epoch+1)))
print("---------------------------")
print('Training accuracy is {}'.format(total_train_accuracy/no_train_examples))
print('Validation accuracy is {}'.format(total_val_accuracy/no_val_examples))

• Re-initializable iterator: This iterator overcomes the problem of initializable iterator by using two separate Datasets. Each dataset can go through its own preprocessing pipeline. The iterator can be created using the tf.Iterator.from_structure method.

def map_fnc(X, Y):

return X, Y

epochs = 10

batch_size = 64

tf.reset_default_graph()

X_data = tf.placeholder(tf.float32, [None, 32,32,1])

Y_data = tf.placeholder(tf.float32, [None, 10])

train_dataset = tf.data.Dataset.from_tensor_slices((X_data, Y_data)).batch(batch_size).map(map_fnc)

val_dataset =  tf.data.Dataset.from_tensor_slices((X_data, Y_data)).batch(batch_size)

iterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)

X_batch , Y_batch = iterator.get_next()

(learner, accuracy) = model(X_batch, Y_batch)

train_initializer = iterator.make_initializer(train_dataset)

val_initializer =  iterator.make_initializer(val_dataset)

with tf.Session() as sess:

sess.run(tf.global_variables_initializer())

for epoch in range(epochs):

# train the model
sess.run(train_initializer, feed_dict={X_data:X_train, Y_data:y_train})
total_train_accuracy = 0
no_train_examples = len(y_train)
try:
  while True:
    temp_train_accuracy, _ = sess.run([accuracy, learner])
    total_train_accuracy += temp_train_accuracy*batch_size
except tf.errors.OutOfRangeError:
  pass

# validate the model
sess.run(val_initializer, feed_dict={X_data:X_val, Y_data:y_val})
total_val_accuracy = 0
no_val_examples = len(y_val)
try:
  while True:
    temp_val_accuracy = sess.run(accuracy)
    total_val_accuracy += temp_val_accuracy*batch_size
except tf.errors.OutOfRangeError:
  pass

print('Epoch {}'.format(str(epoch+1)))
print("---------------------------")
print('Training accuracy is {}'.format(total_train_accuracy/no_train_examples))
print('Validation accuracy is {}'.format(total_val_accuracy/no_val_examples))

• Feedable iterator: This iterator provides the option of switching between various iterators. You can create a re-initializable iterator for training and validation purposes. For inference/testing where you require one pass of the dataset, you can use the one shot iterator.

epochs = 10

batch_size = 64

tf.reset_default_graph()

X_data = tf.placeholder(tf.float32, [None, 32,32,1])

Y_data = tf.placeholder(tf.float32, [None, 10])

train_dataset = tf.data.Dataset.from_tensor_slices((X_data, Y_data)).batch(batch_size)

val_dataset =  tf.data.Dataset.from_tensor_slices((X_data, Y_data)).batch(batch_size)

test_dataset =  tf.data.Dataset.from_tensor_slices((X_test, y_test.astype(np.float32)).batch(batch_size)

handle = tf.placeholder(tf.string, shape=[])

iterator = tf.data.Iterator.from_string_handle(handle, train_dataset.output_types, train_dataset.output_shapes)

X_batch , Y_batch = iterator.get_next()

(learner, accuracy) = model(X_batch, Y_batch)

train_val_iterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)

train_iterator = train_val_iterator.make_initializer(train_dataset)

val_iterator = train_val_iterator.make_initializer(val_dataset)

test_iterator = test_dataset.make_one_shot_iterator()

with tf.Session() as sess:

sess.run(tf.global_variables_initializer())

train_val_string_handle = sess.run(train_val_iterator.string_handle())

test_string_handle = sess.run(test_iterator.string_handle())

for epoch in range(epochs):
# train the model
sess.run(train_iterator, feed_dict={X_data:X_train, Y_data:y_train})
total_train_accuracy = 0
no_train_examples = len(y_train)
try:
  while True:
    temp_train_accuracy, _ = sess.run([accuracy, learner], feed_dict={handle:train_val_string_handle})
    total_train_accuracy += temp_train_accuracy*batch_size
except tf.errors.OutOfRangeError:
  pass

# validate the model
sess.run(val_iterator, feed_dict={X_data:X_val, Y_data:y_val})
total_val_accuracy = 0
no_val_examples = len(y_val)
try:
  while True:
    temp_val_accuracy, _ = sess.run([accuracy, learner], feed_dict={handle:train_val_string_handle})
    total_val_accuracy += temp_val_accuracy*batch_size
except tf.errors.OutOfRangeError:
  pass

print('Epoch {}'.format(str(epoch+1)))
print("---------------------------")
print('Training accuracy is {}'.format(total_train_accuracy/no_train_examples))
print('Validation accuracy is {}'.format(total_val_accuracy/no_val_examples))

print(“Testing the model --------”)
total_test_accuracy = 0

no_test_examples = len(y_test)

try:

while True:

temp_test_accuracy, _ = sess.run([accuracy, learner], feed_dict={handle:test_string_handle})

total_test_accuracy += temp_test_accuracy*batch_size

except tf.errors.OutOfRangeError:

pass
print(‘Testing accuracy is {}’.format(total_test_accuracy/no_test_examples))

Thanks for reading the blog. The code examples used in this blog can be found in this jupyter notebook.

Do leave your comments below if you have any questions or if you have any suggestions for improving this blog.

Originally published by Animesh Agarwal at https://towardsdatascience.com/

Learn More

☞ Complete Guide to TensorFlow for Deep Learning with Python

☞ Modern Deep Learning in Python

☞ TensorFlow 101: Introduction to Deep Learning

☞ Tensorflow Bootcamp For Data Science in Python

☞ Machine Learning A-Z™: Hands-On Python & R In Data Science

☞ Data Science A-Z™: Real-Life Data Science Exercises Included

#tensorflow #data-science