Contribute to the TensorFlow API documentation

Testable docstrings

TensorFlow uses DocTest to test code snippets in Python docstrings. The snippet must be executable Python code. To enable testing, prepend the line with >>> (three left-angle brackets). For example, here's a excerpt from the tf.concat function in the array_ops.py source file:

def concat(values, axis, name="concat"):
  """Concatenates tensors along one dimension.
  ...

  >>> t1 = [[1, 2, 3], [4, 5, 6]]
  >>> t2 = [[7, 8, 9], [10, 11, 12]]
  >>> concat([t1, t2], 0)
  <tf.Tensor: shape=(4, 3), dtype=int32, numpy=
  array([[ 1,  2,  3],
         [ 4,  5,  6],
         [ 7,  8,  9],
         [10, 11, 12]], dtype=int32)>

  <... more description or code snippets ...>

  Args:
    values: A list of `tf.Tensor` objects or a single `tf.Tensor`.
    axis: 0-D `int32` `Tensor`.  Dimension along which to concatenate. Must be
      in the range `[-rank(values), rank(values))`. As in Python, indexing for
      axis is 0-based. Positive axis in the rage of `[0, rank(values))` refers
      to `axis`-th dimension. And negative axis refers to `axis +
      rank(values)`-th dimension.
    name: A name for the operation (optional).

    Returns:
      A `tf.Tensor` resulting from concatenation of the input tensors.
  """

  <code here>

To assess reference documentation quality, see the example section of the TensorFlow 2 API Docs advice. (Be aware that the Task Tracker on this sheet is no longer in use.)

Make the code testable with DocTest

Currently, many docstrings use backticks (```) to identify code. To make the code testable with DocTest:

• Remove the backticks (```) and use the left-brackets (>>>) in front of each line. Use (...) in front of continued lines.
• Add a newline to separate DocTest snippets from Markdown text to render properly on tensorflow.org.

Customizations

TensorFlow uses a few customizations to the builtin doctest logic:

• It does not compare float values as text: Float values are extracted from the text and compared using allclose with liberal atol and rtol tolerences. This allows :
  • Clearer docs - Authors don't need to include all decimal places.
  • More robust tests - Numerical changes in the underlying implementation should never cause a doctest to fail.
• It only checks the output if the author includes output for a line. This allows for clearer docs because authors usually don't need to capture irrelevant intermediate values to prevent them from being printed.

Docstring considerations

• Overall: The goal of doctest is to provide documentation, and confirm that the documentation works. This is different from unit-testing. So:
  • Keep examples simple.
  • Avoid long or complicated outputs.
  • Use round numbers if possible.
• Output format: The output of the snippet needs to be directly beneath the code that’s generating the output. Also, the output in the docstring has to be exactly equal to what the output would be after the code is executed. See the above example. Also, check out this part in the DocTest documentation. If the output exceeds the 80 line limit, you can put the extra output on the new line and DocTest will recognize it. For example, see multi-line blocks below.
• Globals: The `tf`, np and os modules are always available in TensorFlow's DocTest.

• Use symbols: In DocTest you can directly access symbols defined in the same file. To use a symbol that’s not defined in the current file, please use TensorFlow’s public API tf.xxx instead of xxx. As you can see in the example below, `random.normal` is accessed via `tf.random.normal`. This is because `random.normal` is not visible in NewLayer.

  def NewLayer():
    """This layer does cool stuff.
  
    Example usage:
  
    >>> x = tf.random.normal((1, 28, 28, 3))
    >>> new_layer = NewLayer(x)
    >>> new_layer
    <tf.Tensor: shape=(1, 14, 14, 3), dtype=int32, numpy=...>
    """
  

• Floating point values: The TensorFlow doctest extracts float values from the result strings, and compares using np.allclose with reasonable tolerances (atol=1e-6, rtol=1e-6). This way authors do not need to worry about overly precise docstrings causing failures due to numerical issues. Simply paste in the expected value.

• Non-deterministic output: Use ellipsis(...) for the uncertain parts and DocTest will ignore that substring.

    x = tf.random.normal((1,))
    print(x)
        <tf.Tensor: shape=(1,), dtype=float32, numpy=..., dtype=float32)>
        

• Multi-line blocks: DocTest is strict about the difference between a single and a multi-line statement. Note the usage of (...) below:

    if x > 0:
      print("X is positive")
    model.compile(
      loss="mse",
      optimizer="adam")
        

• Exceptions: Exception details are ignored except the Exception that’s raised. See this for more details.

    np_var = np.array([1, 2])
    tf.keras.backend.is_keras_tensor(np_var)
        Traceback (most recent call last):
    
        ValueError: Unexpectedly found an instance of type `<class 'numpy.ndarray'>`.
        

Use a project-local copy of tf-doctest.

Some API's in TensorFlow come from an external project:

• tf.estimator (from tensorflow_estimator)
• tf.summary tensorboard)
• tf.keras.preprocessing (from keras-preprocessing)

If you're working on an external project, or on TensorFlow APIs that are housed in an external project, these instructions won't work unless that project has its own local copy of tf_doctest, and you use that copy instead of TensorFlow's.

For example: tf_estimator_doctest.py.

Test on your local machine

There are two ways to test the code in the docstring locally:

• If you are only changing the docstring of a class/function/method, then you can test it by passing that file's path to tf_doctest.py. For example:

  python tf_doctest.py --file=<file_path>
  

  This will run it using your installed version of TensorFlow. To be sure you're running the same code that you're testing:

  • Use an up to date tf-nightly pip install -U tf-nightly
  • Rebase your pull request onto a recent pull from TensorFlow's master branch.

• If you are changing the code and the docstring of a class/function/method, then you will need to build TensorFlow from source. Once you are setup to build from source, you can run the tests:

  bazel run //tensorflow/tools/docs:tf_doctest
  

  or

  bazel run //tensorflow/tools/docs:tf_doctest -- --module=ops.array_ops
  

  The --module is relative to tensorflow.python.