Google is committed to advancing racial equity for Black communities. See how.

TensorFlow code style guide

Python style

Follow the PEP 8 Python style guide, except TensorFlow uses 2 spaces instead of 4. Please conform to the Google Python Style Guide, and use pylint to check your Python changes.

pylint

To install pylint and retrieve TensorFlow's custom style definition:


$ pip install pylint
$ wget -O /tmp/pylintrc https://raw.githubusercontent.com/tensorflow/tensorflow/master/tensorflow/tools/ci_build/pylintrc

To check a file with pylint:

$ pylint --rcfile=/tmp/pylintrc myfile.py

Supported Python versions

TensorFlow supports Python >= 3.5. See the installation guide for details.

See the TensorFlow continuous build status for official and community supported builds.

C++ coding style

Changes to TensorFlow C++ code should conform to the Google C++ Style Guide. Use clang-format to check your C/C++ changes.

To install on Ubuntu 16+, do:

$ apt-get install -y clang-format

You can check the format of a C/C++ file with the following:

$ clang-format <my_cc_file> --style=google > /tmp/my_cc_file.cc
$ diff <my_cc_file> /tmp/my_cc_file.cc

Other languages

TensorFlow conventions and special uses

Python operations

A TensorFlow operation is a function that, given input tensors returns output tensors (or adds an op to a graph when building graphs).

  • The first argument should be tensors, followed by basic Python parameters. The last argument is name with a default value of None.
  • Tensor arguments should be either a single tensor or an iterable of tensors. That is, a "Tensor or list of Tensors" is too broad. See assert_proper_iterable.
  • Operations that take tensors as arguments should call convert_to_tensor to convert non-tensor inputs into tensors if they are using C++ operations. Note that the arguments are still described as a Tensor object of a specific dtype in the documentation.
  • Each Python operation should have a name_scope. As seen below, pass the name of the op as a string.
  • Operations should contain an extensive Python comment with Args and Returns declarations that explain both the type and meaning of each value. Possible shapes, dtypes, or ranks should be specified in the description. See documentation details.
  • For increased usability, include an example of usage with inputs / outputs of the op in Example section.
  • Avoid making explicit use of tf.Tensor.eval or tf.Session.run. For example, to write logic that depends on the Tensor value, use the TensorFlow control flow. Alternatively, restrict the operation to only run when eager execution is enabled (tf.executing_eagerly()).

Example:

def my_op(tensor_in, other_tensor_in, my_param, other_param=0.5,
          output_collections=(), name=None):
  """My operation that adds two tensors with given coefficients.

  Args:
    tensor_in: `Tensor`, input tensor.
    other_tensor_in: `Tensor`, same shape as `tensor_in`, other input tensor.
    my_param: `float`, coefficient for `tensor_in`.
    other_param: `float`, coefficient for `other_tensor_in`.
    output_collections: `tuple` of `string`s, name of the collection to
                        collect result of this op.
    name: `string`, name of the operation.

  Returns:
    `Tensor` of same shape as `tensor_in`, sum of input values with coefficients.

  Example:
    >>> my_op([1., 2.], [3., 4.], my_param=0.5, other_param=0.6,
              output_collections=['MY_OPS'], name='add_t1t2')
    [2.3, 3.4]
  """
  with tf.name_scope(name or "my_op"):
    tensor_in = tf.convert_to_tensor(tensor_in)
    other_tensor_in = tf.convert_to_tensor(other_tensor_in)
    result = my_param * tensor_in + other_param * other_tensor_in
    tf.add_to_collection(output_collections, result)
    return result

Usage:

output = my_op(t1, t2, my_param=0.5, other_param=0.6,
               output_collections=['MY_OPS'], name='add_t1t2')