Testable docstrings
TensorFlow uses DocTest to
test code snippets in Python docstrings. The snippet must be executable Python
code. To enable testing, preprend 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>
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.
- (```) can still be used for non-Python code or code that cannot be tested for some reason.
- Add a newline as a separation between each DocTest and Markdown text.
Docstring considerations
- 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
andos
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 ofxxx
. As you can see in the example below,random.normal
is accessed viatf.random.normal
. This is becauserandom.normal
is not visible inNewLayer
.
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=...> “””
- 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'>`.
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.
- Use an up to date tf-nightly
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 totensorflow.python
.