Skip to content

Latest commit

 

History

History
131 lines (110 loc) · 5.3 KB

RELEASE.md

File metadata and controls

131 lines (110 loc) · 5.3 KB

Current version 0.17.0-dev

  • Under development

Release 0.16.0

  • Enforce using Keras 2 by checking for installation of tf_keras
  • Remove deprecated TF 1 API (LatestModuleExporter, register_module_for_export, ...)

Release 0.15.0

  • Require Python 3.9+.
  • Bump rules_license to 0.0.4.
  • Move LICENSE to tensorflow_hub/LICENSE.
  • Use license_files instead of license_file in setup.cfg.
  • Remove empty tools directory.

Release 0.14.0

Release 0.13.0

  • Minor fixes for importing Keras and TF Estimator.
  • Stop using distutils.version.
  • Require protobuf >= 3.19.6.
  • Require Python 3.7+.

Release 0.12.0

  • Improve support for compute_output_shape in hub.KerasLayer:
    • This will now use the output_shape value provided on init, if present, before falling back to the default behavior.
  • Changes to make_image_classifier_tool:
    • Add the option to read input with a tf.data.Dataset and use TF ops for preprocessing. Enabling this feature requires TF 2.5 or higher.
    • Set the default value of shear_range to 0. shear_range is deprecated and incompatible when using TF ops for preprocessing.

Release 0.11.0

  • Use the Keras load context in keras_layer.py.
  • Always use compressed model loading by default.
  • Documentation improvements.

Release 0.10.0

  • Enforce Python 3.5+ and TF1.15+.
  • Add ModelLoadFormat setting to switch between (un)compressed model loading.
  • Support for RaggedTensor inputs/outputs is backported from TF2 SavedModels to the deprecated hub.Module class (Use of tf.compat.v1.ragged.placeholder() in a module_fn is broken for TF2.3 and requires TF2.2 or TF2.4).
  • Bug fixes.

Release 0.9.0

  • Add SavedModel.LoadOptions to hub.KerasLayer API to pass to load_v2.
  • Improved error messaging.
  • Documentation improvements.
  • Bug fixes.

Release 0.8.0

  • Implemented make_nearest_neighbour_index tool.
  • Added text FeatureColumn, hub.text_embedding_column_v2, for TF 2.x.
  • CORD-19 embedding colab.
  • Documentation improvements.
  • Bug fixes.

Release 0.7.0

  • Added support for HubModule v1 in KerasLayer with default tags/signatures.
  • Added support for the KerasLayer to specify tags, signature, as_dict, and output_key arguments.
  • Miscellaneous fixes to hub.KerasLayer
  • Documentation update for TensorFlow 2.
    • Use hub.load() and hub.KerasLayer with TF2 (also works in 1.15).
    • For TF1, hub.Module and it's associated APIs remain available.

Release 0.6.0

  • Added two examples for exporting of Hub/TF2 modules.
  • Switched to dense_features_v2 if a high enough TF version is installed.
  • Added tools/make_image_classifier for use with TF2.

Release 0.5.0

  • Fixes in hub.KerasLayer related with regularizers, config serialization and usage with tf.estimator.
  • Updates to feature columns to be compatible with feature columns V2 apis and add hub.sparse_text_embedding_column for bag of words features.
  • Made hub.Module usable within tf.compat.v1.wrap_function.

Release 0.4.0

  • hub.KerasLayer (for Hub/TF2 modules) can be used in graph mode and can be converted to/from a Keras config.
  • In TF1.x, hub.Module can be used inside a defun (helps with TPU).
  • References to TensorFlow Estimator correctly pick up its v1 API.

Release 0.3.0

  • Initial support for Tensorflow 2.0
  • Tensorflow Hub Library API for Tensorflow 2.0
  • Tensorflow Hub Keras API for Tensorflow 2.0
  • Enable using Tensorflow Hub library against Tensorflow 2.0 run-time.

Release 0.2.0

  • Add support for caching modules on GCS.
  • Add ModuleSpec.export(export_path, checkpoint_path) helper for module creation from a separately trained checkpoint.
  • Modules now have a key-value store of attached protocol messages for domain- or library-specific auxiliary data.
  • Image modules can have input shapes with variable height and width; hub.get_expected_image_size() comes from an attached ImageModuleInfo message.
  • Add with hub.eval_function_for_module("...") as f: out = f(in).
  • New experimental function hub.create_module_spec_from_saved_model().
  • Added property hub.Module.variables.
  • Bazel workspace dependency on protobuf library updated to version 3.6.0.
  • Added progress bar for module downloads when interactive.
  • Numerous minor fixes and updates from the master branch.

Release 0.1.1

  • Removed TensorFlow version checking.

Release 0.1.0

  • Initial TensorFlow Hub release.

Requirements

  • tf-nightly>=1.7.0.dev20180308 || tensorflow>=1.7.0rc0