Skip to content

Latest commit

 

History

History
118 lines (82 loc) · 4.85 KB

CHANGELOG.md

File metadata and controls

118 lines (82 loc) · 4.85 KB
Raw

Changelog

Major KerasTuner changes by version

V0.8: Framework complete

Major features

  • Tuner API refactored and simplified
  • User API refactored and simplified
  • Tunable Model API initial version and initial application with tunable_resnet
  • Experiments recording refactored to remove duplicate, improve bucketing and log additional informations.
  • Display components refactored to offer a more consistent an clearer experience.
  • Tensorflow 2 compatibility added
  • Model exporting added -- allows to export in Keras formats and Tensorflow
  • Model resuming added.
  • Hello world tutorial completed.
  • Add classification metrics reporting.

Extra

  • access to system info and display have been moved to an abstraction layer
  • Major code refactoring to finalize API, cleanup the code and encapsulate all components correctly. This include a rewrite of the following components: abstraction, distribution, collections, callbacks, engine, tuner
  • unit-test refactored and extended with over 100 more tests.

V0.7: Ecosystem integration

The focus of these releases is to make kerastuner works well with other and make it easy to operate

Major features

  • kerastuner-status utility allows to monitor training in commandline
  • kerastuner-summary utility allows to display results overview in commandline
  • Tuner report status in json file every 5 seconds to make it easy to track progress with tool and remotely
  • System information including CPU usage, GPU usage, memory usage, disk space is now reported
  • Model max parameters is now directly infered based from GPU available memory and batch_size

Extras

  • Moved statistics reporting to callback and revamped it to look better and use new display/table system
  • Reduced the amount of boilerplate code needed to write a tuner by shifting burden to the scheduler and callback
  • Output adapts when run in colab/jupyter notebooks to use HTML for better readability
  • New cross-platform display subsystem that produce nice and colorful output
  • Kerastuner warm user if tensorflow is not using GPU and GPU are available

Noteworthy bugfixes

  • setup.py now works in developper mode (issue #41)
  • utilities now use setuptools proprerly to offers cross-platform executable

v0.6: Engine stablization

The focus of this set of release is to have a robust engine that works seamlessly and we can build upon

Major features

  • Upon restarting the tuner -- already existing instances are loaded to prevent retraining them
  • Added avg_accuracy metric computation in callback to allows checkpointing on accuracy for multi-output
  • Replaced model saving at the end of the training with model checkpointing at epoch end: model are now saved when they improve.
  • Added reporting of the search space size for each hyperparameter
  • Added the ability to group hyperparameters in groups to make it easier for post processing
  • Added a summary() function that provide a breakdown of the hyperparameter search space.
  • Added a tool to display results summary in terminal tools/display-results-summary.py

Extras

  • simplified demo
  • Switched from keras to tf.keras
  • Switched xxhash to hashlib due to dependencies issues (rare)
  • Added traceback via verbose for model debug

Noteworthy bugfixes

  • Default metrics not set properly
  • gs:// schema unsupported crashed the tuner
  • Num epochs is properly tracked
  • History is now serializable
  • Callback overflows
  • OOM issues while training on GPU
  • Multi-gpu is now working as intended
  • Model generation stats are now correctly displayed
  • Dryrun mode now works as intended

v0.5

Major features

  • Added the ability to use generator via the search_generator function
  • Added the ability to supply a personalized callback_generator function that is invoked at every execution and receive execution_info as input. Allows to use callbacks which are different from one execution to another which is needed to support complex models such a triplet loss.
  • Added a test(int) function which allows to test the model function before getting started

Extras

  • Add a configurable model over_size detection to avoid OOM errors when TF attempt to train model with too many parameters

  • Moved the cloud configuration to backend() function with notification configuration

  • Added additional statistics reporting:

    • overall: epoch budget, epoch budget remaining, eta, hypertuner used
    • per instance: eta, epoch remaining

  • Various bug fixes including GPU memory release which caused OOM crash

v0.4

  • Cloudstreaming to keraslyzer service implemented.

v0.3

  • Added callback that takes care to periodically records current execution. Mostly meant to make easy to monitor long training.

v0.2

  • Ultraband tuner implemented and becoming default tuner - inspired by ultraband.
  • Breaking API: moved to epoch_budget to define how long the tuner should run. This makes it easier for user to budget optimization.