tensorflow version 1 or version 2 api. + Minor change to the way root_init_template flag is passed in (now it should be a location path instead of file name).
PiperOrigin-RevId: 286729526
Change-Id: I55ebaa0cfe0fe3db3f4d1e699082b1f7b11df4da
Specifically, it makes autocomplete work for `tf.image.` (with `import
tensorflow as tf`) or `image.` (with
`from tensorflow import image`).
PiperOrigin-RevId: 264901051
- Updates existing DeprecationWrapper with the ability to import modules only when they are referred.
- Updates how TensorFlow generates public API. Wraps all generated TensorFlow __init__.py modules with this enhanced wrapper.
To enable lazy-loading in the future, toggle _LAZY_LOADING flag in create_python_api.py.
Once lazy loading is enabled, the wrapper will have the following behaviors:
- dir() will always return module?s attributes.
- __all__ will always return all public APIs.
- __dict__ will be populated as attributes are being referred.
- After wrapper instance is created, to add more attributes, use setattr(import does not explicitly call setattr) to make sure dir, __all__, __dict__ are updated.
- import * will work as expected.
Built and tested with pip package.
PiperOrigin-RevId: 257240535
We eagerly initialize __all__ variable in the API template and then we make more imports. This fix just adds required names to __all__.
PiperOrigin-RevId: 252469926
setattr in package_hook. This is more verbose now but fixes issue where PyCharm autocomplete won't find modules such as 'keras'. Resolves#26502.
PiperOrigin-RevId: 244473718
When deleting `compiler`, `core` and `python` symbols from `tensorflow`,
we currently don't remove them from `__all__`. This breaks anybody who
tries to do `from tensorflow import *`.
PiperOrigin-RevId: 243897920
This change has the effect of importing scalar(), image() and friends from TensorBoard's summary API into the TF 2.0 tf.summary module. It actually does so by replacing tf.summary entirely, via the component_api_helper() mechanism (in use for tf.estimator), with a replacement package from TensorBoard that re-exports the original tf.summary symbols, thereby creating a fused namespace.
The component package approach mitigates circular dependency issues (a risk because the TensorBoard ops in turn depend on TF) because the TF API __init__.py imports the core TF API symbols before it imports components, so if they in turn `import tensorflow` the module has been mostly initialized. It also means that TensorBoard needn't be importable at API generation time, only when TF is actually imported, which avoids issues like #22395.
Tracking issue: #25356.
PiperOrigin-RevId: 233874463
When moving Keras to a separate repo, we will have to separate the API
generation from the rest of TensorFlow. The generated API files will have
to part of the Keras package, as core TensorFlow will have no way of knowing
which version of the tf.Keras API will be installed.
This CL does a few things...
- Move ALL @tf_export("keras.*") declarations in to keras/ directory.
See python/keras/ops.py where I had to move lots of these decorators
out of TensorFlow code.
- Change ALL @tf_export("keras.*") to @keras_export("keras.*"). Doing
this will allow us to generate the public Keras python __init__.py files
separately from the rest of the TensorFlow API.
- Add a new api_gen targets to keras/api/BUILD to generate this API.
PiperOrigin-RevId: 226050838
Add tensorflow.contrib.lite into traversal of api generator.
+ Needed to change API generator to take list of deps and packages.
+ Exempt tensorflow.conrib.lite from METADATA dependency check (since
we are moving)
PiperOrigin-RevId: 219483886
