This document provides answers to some of the frequently asked questions about TensorFlow. If you have a question that is not covered here, you might find an answer on one of the TensorFlow community resources.
Features and Compatibility
Can I run distributed training on multiple GPUs?
Yes! TensorFlow gained support for distributed computation back in version 0.8, see the distributed computation guide. TensorFlow supports multiple devices (CPUs and GPUs) on one or more computers.
Does TensorFlow work with Python 3?
TensorFlow graphs and eager execution
Do TensorFlow operations return immediately?
If you enable eager execution, operations like
c = tf.matmul(a, b) are executed
immediately. See the eager execution guide for using eager execution
to create more readable, intuitive TensorFlow code.
However, without eager execution enabled, an operation like
tf.matmul above does
not execute immediately, but, instead, builds a fragment of a TensorFlow graph.
Why graphs? TensorFlow graphs can help with distribution, optimization, and putting
models into production. In the suggested expression,
tf.Tensor object is a symbolic handle to the result of an
operation, but does not actually hold the values of the operation's
output. Instead, you can build up complicated expressions (such as
entire neural networks and their gradients) as a dataflow graph. You
then offload the computation of the entire dataflow graph (or a
subgraph of it) to a TensorFlow
tf.Session, which is able to execute
the whole computation much more efficiently than executing the
How are devices named?
The supported device names are
"/cpu:0") for the CPU
"/gpu:i") for the ith GPU device.
How do I place operations on a particular device?
You can also look at CIFAR-10 tutorial for an example model that uses multiple GPUs.
As of r1.12, we recommend trying
tf.contrib.distribute.DistributionStrategy as an easy way to
distribute computation with Keras and Estimator models. It is under
Running a TensorFlow computation
See the API documentation on running graphs.
What is feeding and placeholders?
However, in some older models you may find feeds and placeholders.
Feeding is a mechanism in the
tf.Session API that allows you
to substitute different values for one or more tensors at run
feed_dict argument to
tf.Session.run is a dictionary
tf.Tensor objects to numpy arrays (and some other types),
which will be used as the values of those tensors in the execution of
What is the difference between
# Using `Session.run()`. sess = tf.Session() c = tf.constant(5.0) print(sess.run(c)) # Using `Tensor.eval()`. c = tf.constant(5.0) with tf.Session(): print(c.eval())
In the second example, the session acts as a
which has the effect of installing it as the default session for the lifetime of
with block. The context manager approach can lead to more concise code for
simple use cases (like unit tests); if your code deals with multiple graphs and
sessions, it may be more straightforward to make explicit calls to
Do Sessions have a lifetime? What about intermediate tensors?
Sessions can own resources, such as
tf.ReaderBase. These resources can sometimes use
a significant amount of memory, and can be released when the session is closed by calling
The intermediate tensors that are created as part of a call to
Session.run() will be freed at or before the
end of the call.
Does the runtime parallelize parts of graph execution?
When you use graph execution, the TensorFlow runtime parallelizes execution across many different dimensions:
- The individual ops have parallel implementations, using multiple cores in a CPU, or multiple threads in a GPU.
- Independent nodes in a TensorFlow graph can run in parallel on multiple devices, which makes it possible to speed up CIFAR-10 training using multiple GPUs.
- The Session API allows multiple concurrent steps (i.e. calls to
tf.Session.runin parallel). This enables the runtime to get higher throughput, if a single step does not use all of the resources in your computer.
Which client languages are supported in TensorFlow?
TensorFlow is designed to support multiple client languages. Currently, the best-supported client language is Python. Experimental interfaces for executing and constructing graphs are also available for C++, Java and Go.
TensorFlow also has a C-based client API to help build support for more client languages. We invite contributions of new language bindings.
Separately, there is the Swift for TensorFlow project, which integrates TensorFlow directly into the Swift programming language.
Session.run() hang when using a reader or a queue?
tf.QueueBase classes provide special operations that
can block until input (or free space in a bounded queue) becomes
available. These operations allow you to build sophisticated
input pipelines, at the cost of making the
TensorFlow computation somewhat more complicated. See the how-to documentation
QueueRunner objects to drive queues and readers
for more information on how to use them.
Should I turn on
use_resource=True when constructing variables?
Yes. This uses safer memory behavior, and will be the default in TensorFlow 2.0.
What is the lifetime of a variable?
In eager execution, variables are freed when their associated Python objects are cleaned up.
How do variables behave when they are concurrently accessed?
Variables allow concurrent read and write operations. The value read from a
variable may change if it is concurrently updated. By default, concurrent
assignment operations to a variable are allowed to run with no mutual exclusion.
To acquire a lock when assigning to a variable, pass
See also the
How can I determine the shape of a tensor in Python?
In TensorFlow, a tensor has both a static (inferred) shape and a dynamic (true)
shape. The static shape can be read using the
method: this shape is inferred from the operations that were used to create the
tensor, and may be partially complete (the static-shape may contain
the static shape is not fully defined, the dynamic shape of a
can be determined using
What is the difference between
x = tf.reshape(x)?
tf.Tensor.set_shape method updates
the static shape of a
Tensor object, and it is typically used to provide
additional shape information when this cannot be inferred directly. It does not
change the dynamic shape of the tensor.
tf.reshape operation creates
a new tensor with a different dynamic shape.
How do I build a graph that works with variable batch sizes?
It is often useful to build a graph that works with variable batch sizes
so that the same code can be used for (mini-)batch training, and
single-instance inference. The resulting graph can be
When building a variable-size graph, the most important thing to remember is not
to encode the batch size as a Python constant, but instead to use a symbolic
Tensor to represent it. The following tips may be useful:
batch_size = tf.shape(input)to extract the batch dimension from a
input, and store it in a
tf.reduce_sum(...) / batch_size.
How can I visualize a TensorFlow graph?
See the graph visualization tutorial.
What is the simplest way to send data to TensorBoard?
Add summary ops to your TensorFlow graph, and write these summaries to a log directory. Then, start TensorBoard using
python tensorflow/tensorboard/tensorboard.py --logdir=path/to/log-directory
For more details, see the Summaries and TensorBoard tutorial.
Every time I launch TensorBoard, I get a network security popup!
You can change TensorBoard to serve on localhost rather than '0.0.0.0' by the flag --host=localhost. This should quiet any security warnings.
See the how-to documentation for adding a new operation to TensorFlow.
My data is in a custom format. How do I read it using TensorFlow?
There are three main options for dealing with data in a custom format.
The easiest option is to write parsing code in Python that transforms the data
into a numpy array. Then, use
create an input pipeline from the in-memory data.
If your data doesn't fit in memory, try doing the parsing in the Dataset
pipeline. Start with an appropriate file reader, like
tf.data.TextLineDataset. Then convert the dataset by mapping
tf.data.Dataset.map appropriate operations over it.
Prefer predefined TensorFlow operations such as
If your data is not easily parsable with the built-in TensorFlow operations,
consider converting it, offline, to a format that is easily parsable, such
The most efficient method to customize the parsing behavior is to add a new op written in C++ that parses your data format. The guide to handling new data formats has more information about the steps for doing this.
What is TensorFlow's coding style convention?
The TensorFlow Python API adheres to the
PEP8 conventions.* In
particular, we use
CamelCase names for classes, and
snake_case names for
functions, methods, and properties. We also adhere to the
Google Python style guide.
The TensorFlow C++ code base adheres to the Google C++ style guide.
(* With one exception: we use 2-space indentation instead of 4-space indentation.)