Skip to content

Releases: keras-team/keras-tuner

Release v1.3.4

02 Apr 22:17
f3840fd
Compare
Choose a tag to compare

Bug fixes

  • If you have a protobuf version > 3.20, it would through an error when import KerasTuner. It is now fixed.

Release v1.3.3

27 Mar 19:08
cbd32f7
Compare
Choose a tag to compare
  • KerasTuner would install protobuf 3.19 with protobuf<=3.20. We want to install 3.20.3, so we changed it to protobuf<=3.20.3. It is now fixed.

Full Changelog: v1.3.2...v1.3.3

Release v1.3.2

27 Mar 17:58
bf2cce4
Compare
Choose a tag to compare

Bug fixes

  • It use to install protobuf 4.22.1 if install with TensorFlow 2.12, which is not compatible with KerasTuner. We limited the protobuf version to <=3.20, which is compatible with all TensorFlow versions so far.

Full Changelog: v1.3.1...v1.3.2

Release v1.3.1

27 Mar 02:59
a7bd433
Compare
Choose a tag to compare

Bug fixes

  • The Tuner.results_summary() did not print error messages for failed trials and did not display Objective information correctly. It is now fixed.
  • The BayesianOptimization would break when not specifying the num_initial_points and overriding .run_trial(). It is now fixed.
  • TensorFlow 2.12 would break because the different protobuf version. It is now fixed.

New Contributors

Full Changelog: v1.3.0...v1.3.1

Release v1.3.0

23 Feb 01:48
091dcb8
Compare
Choose a tag to compare

Breaking changes

  • Removed Logger and CloudLogger and the related arguments in BaseTuner.__init__(logger=...).
  • Removed keras_tuner.oracles.BayesianOptimization, keras_tuner.oracles.Hyperband, keras_tuner.oracles.RandomSearch, which were actually Oracles instead of Tuners. Please usekeras_tuner.oracles.BayesianOptimizationOracle, keras_tuner.oracles.HyperbandOracle, keras_tuner.oracles.RandomSearchOracle instead.
  • Removed keras_tuner.Sklearn. Please use keras_tuner.SklearnTuner instead.

New features

  • keras_tuner.oracles.GridSearchOracle is now available as a standalone Oracle to be used with custom tuners.

Full Changelog: 1.2.1...v1.3.0

Release v1.2.1

10 Feb 01:06
aba01ea
Compare
Choose a tag to compare

Bug fixes

  • The resume feature (overwrite=False) would crash in 1.2.0. This is now fixed.

New Contributors

Full Changelog: 1.2.0...1.2.1

Release v1.2.0

28 Jan 18:25
c48d239
Compare
Choose a tag to compare

Release v1.2.0

Breaking changes

  • If you implemented your own Tuner, the old use case of reporting results with Oracle.update_trial() in Tuner.run_trial() is deprecated. Please return the metrics in Tuner.run_trial() instead.
  • If you implemented your own Oracle and overrided Oracle.end_trial(), you need to change the signature of the function from Oracle.end_trial(trial.trial_id, trial.status) to Oracle.end_trial(trial).
  • The default value of the step argument in keras_tuner.HyperParameters.Int() is changed to None, which was 1 before. No change in default behavior.
  • The default value of the sampling argument in keras_tuner.HyperParameters.Int() is changed to "linear", which was None
    before. No change in default behavior.
  • The default value of the sampling argument in keras_tuner.HyperParameters.Float() is changed to "linear", which was
    None before. No change in default behavior.
  • If you explicitly rely on protobuf values, the new protobuf bug fix may affect you.
  • Changed the mechanism of how a random sample is drawn for a hyperparameter. They now all start from a random value between 0 and 1, and convert the value to a random sample.

New features

  • A new tuner is added, keras_tuner.GridSearch, which can exhaust all the possible hyperparameter combinations.
  • Better fault tolerance during the search. Added two new arguments to Tuner and Oracle initializers, max_retries_per_trial and max_consecutive_failed_trials.
  • You can now mark a Trial as failed by raise keras_tuner.FailedTrialError("error message.") in HyperModel.build(), HyperModel.fit(), or your model build function.
  • Provides better error messages for invalid configs for Int and Float type hyperparameters.
  • A decorator @keras_tuner.synchronized is added to decorate the methods in Oracle and its subclasses to synchronize the concurrent calls to ensure thread safety in parallel tuning.

Bug fixes

  • Protobuf was not converting Boolean type hyperparameter correctly. This is now fixed.
  • Hyperband was not loading the weights correctly for half-trained models. This is now fixed.
  • KeyError may occur if using hp.conditional_scope(), or the parent argument for hyperparameters. This is now fixed.
  • num_initial_points of the BayesianOptimization should defaults to 3 * dimension, but it defaults to 2. This is now fixed.
  • It would through an error when using a concrete Keras optimizer object to override the HyperModel compile arg. This is now fixed.
  • Workers might crash due to Oracle reloading when running in parallel. This is now fixed.

New Contributors

Full Changelog: 1.1.3...1.2.0

Release v1.2.0 RC0

13 Jan 01:03
Compare
Choose a tag to compare
Release v1.2.0 RC0 Pre-release
Pre-release

Breaking changes

  • If you implemented your own Tuner, the old use case of reporting results with Oracle.update_trial() in Tuner.run_trial() is deprecated. Please return the metrics in Tuner.run_trial() instead.
  • If you implemented your own Oracle and overrided Oracle.end_trial(), you need to change the signature of the function from Oracle.end_trial(trial.trial_id, trial.status) to Oracle.end_trial(trial).
  • The default value of the step argument in keras_tuner.HyperParameters.Int() is changed to None, which was 1 before. No change in default behavior.
  • The default value of the sampling argument in keras_tuner.HyperParameters.Int() is changed to "linear", which was None before. No change in default behavior.
  • The default value of the sampling argument in keras_tuner.HyperParameters.Float() is changed to "linear", which was None before. No change in default behavior.
  • If you explicitly rely on protobuf values, the new protobuf bug fix may affect you.
  • Changed the mechanism of how a random sample is drawn for a hyperparameter. They now all start from a random value between 0 and 1, and convert the value to a random sample.

New features

  • A new tuner is added, keras_tuner.GridSearch, which can exhaust all the possible hyperparameter combinations.
  • Better fault tolerance during the search. Added two new arguments to Tuner and Oracle initializers, max_retries_per_trial and max_consecutive_failed_trials.
  • Provides better error messages for invalid configs for Int and Float type hyperparameters.
  • A decorator @keras_tuner.synchronized is added to decorate the methods in Oracle and its subclasses to synchronize the concurrent calls to ensure thread safety in parallel tuning.

Bug fixes

  • Protobuf was not converting Boolean type hyperparameter correctly. This is now fixed.
  • Hyperband was not loading the weights correctly for half-trained models. This is now fixed.
  • KeyError may occur if using hp.conditional_scope(), or the parent argument for hyperparameters. This is now fixed.
  • num_initial_points of the BayesianOptimization should defaults to 3 * dimension, but it defaults to 2. This is now fixed.

New Contributors

Full Changelog: 1.1.3...1.2.0rc0

Release v1.1.3

16 Jul 04:22
Compare
Choose a tag to compare

Summary

Bug fixes to better support AutoKeras.

What's Changed

New Contributors

Full Changelog: 1.1.2...1.1.3

Release v1.1.3RC0

15 Jul 20:15
Compare
Choose a tag to compare
Release v1.1.3RC0 Pre-release
Pre-release

What's Changed

New Contributors

Full Changelog: 1.1.2...1.1.3rc0