Skip to content

Latest commit

 

History

History
192 lines (149 loc) · 8.58 KB

CHANGELOG.md

File metadata and controls

192 lines (149 loc) · 8.58 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[unreleased] -

[0.0.7] - September 5, 2021

Added

  • CI/CD stuff:

    • Added automatic execution of all examples notebooks as part of CI/CD flow.

  • New acquisition functions.

    • Corresponds to implementations of the following acquisition functions from BoTorch, all working out-of-the-box in greattunes. For more details, please consult the BoTorch documentation
      • ExpectedImprovement
      • NoisyExpectedImprovement
      • PosteriorMean
      • ProbabilityOfImprovement
      • qExpectedImprovement
      • qKnowledgeGradient
      • qMaxValueEntropy
      • qMultiFidelityMaxValueEntropy
      • qNoisyExpectedImprovement
      • qProbabilityOfImprovement
      • qSimpleRegret
      • qUpperConfidenceBound
      • UpperConfidenceBound
    • All available acquisition functions added as attribute ACQ_FUNC_LIST to TuneSession.
    • Acqusition function settings parameters and samplers (for Monte Carlo-based methods) can be provided to TuneSession during class initialization as kwargs. Will default to pre-configured settings if none provided. Specifically, the following parameters can be configured via kwargs:

      • beta - tradeoff parameter for acquisition functions UpperConfidenceBound, qUpperConfidenceBound

      • num_fantasies - number of realizations for generating estimates for acquisition functions qKnowledgeGradient, NoisyExpectedImprovement

      • sampler - sampler for Monte Carlo methods, should be an initialized sampler from BoTorch (details in BoTorch documentation).

  • Added best_predicted-method to class, allowing user to see the best predicted result similar to the best observed result. This can be called after each iteration, and will be printed to the prompt. Using the Nelder-Mead algorithm, the method will return the maximum mean value of the surrogate model as well as the maximum of the lower confidence region.

  • Extended CONTRIBUTING.md with details of how to contribute

Changed

  • Acquisition functions
    • Acquisition function Expected Improvement has been renamed from EI to ExpectedImprovement. This the latter going forward as argument acq_func to TuneSession to invoke Expected Improvement acquisition function.

Deprecated

Removed

Fixed

  • Added repo logo as a static link to GitHub so it shows also on PyPI

[0.0.5] - July 15, 2021

Version for this release: 0.0.5

Added

Changed

  • Updated .tell: kwargs for providing covariate and response observations have been renamed. Should now be provided via the less confusion covar_obs and response_obs

Deprecated

Removed

Fixed

[0.0.4] - July 14, 2021

Added

  • Functionality to use integer and categorical covariates as input to the function under optimization, using the method of Garrido-Merchán and Hernandéz-Lobato (journal link, ArXiv preprint). This significantly extends the applicability of the framework.

  • Named covariates

  • Pretty data format for covariates (x_data) and response (y_data) which keeps track of observations in their natural data types (float for doubles, int for integers and str for categorical variables). These are in pandas format

  • Two new end-to-end examples to illustrate a simple use case of integer covariates (Example 6) and a more elaborate combining continuous, integer and categorical (Example 7).

Changed

  • Extended how the package determines covariates enabled via a wider range of options for the parameter covars provided during class initialization. There are now two methods, see README.md:
    1. A simple in which requires a list of tuples, with each tuple giving the guess, the minimum and the maximum of the covariate. Data types are inferred and covariate names are assigned.
    2. An elaborate that allows more control over data type and covariate naming.
  • In _best_response.current_best: switched to storing in pretty user-facing format (pandas df), updated output slightly
  • Extended creative_project.transformed_kernel_models.transformation.GP_kernel_transformation to support high-rank tensors. This allows using botorch's optimize_acqf method to determine best next covariate datapoint from acquisition function.

Deprecated

None

Removed

In ask-tell-approach: reporting observations via covars and response entries to tell-method cannot be done via the backend data format (torch tensor of same format as train_X and train_Y). Instead, use the same user-facing format (in pandas df) to report all entries, including integer and categorical variables in their natural data types (int and str).

Fixed

None

[0.0.3] - February 25, 2021

Version number for this release: 0.0.3

Added

  • Added new random sampling functionality with two purposes. Firstly, during initialization it is known to be good to start with random sampling if no data is available. Secondly, and also to ensure speedier optimization convergence, a single randomly sampled point every now and then in between Bayesian points is known to increase convergence. Random sampling is now available for both auto approach and ask-tell approach with the following features
    • During class initialization, using random initialization or not is controlled by random_start (default: True)
    • Additional parameters during initialization
      • num_initial_random: number of initial random; no default, if not specified will be set to $\sqrt{# dimensions}$
      • random_sampling_method: sampling method with options random (completely random) and latin_hcs (latin hypercube sampling); defaults to latin_hcs
      • random_step_cadence: the cadence of random sampling after initialization (default: 10)

Changed

In TuneSession class initialization:

  • If historical data is added via train_X, train_Y
    • proposed_X has been changed to be a zero tensor of the same size as train_X. This replaces an empty tensor for proposed_X, which confusingly could take any values.
    • optimization cycle counters (iteration counters) model["covars_proposed_iter"], model["covars_sampled_iter"] and model["response_sampled_iter"] are set so the first iterations are taken as those from the historical data. That is, if train_X, train_Y is provided with two observations during initialization, then the counters are set as model["covars_proposed_iter"]=2, model["covars_sampled_iter"]=2 and model["response_sampled_iter"]=2.

Deprecated

None

Removed

Removed the attribute start_from_random as part of adding more elaborate random sampling functionality.

Fixed

None

[0.0.2] - January 17, 2021

Version number for this release: 0.0.2

Added

  • In .tell-method:
    • Optional functionality to provide observations of covariates and response programmatically (provide as input parameters covars and response)
  • In .auto-method:
    • Optional functionality to stop based on relative improvement of best response detected by the algorithm. Users can stop the algorithm as soon as the relative improvement of the best response drops below a user-specified limit in order to improve the speed of reaching an answer. See Example 5 in examples for illustration.
  • Examples of end-to-end workflows of using the library as Jupyter notebooks are added in examples folder with descriptions.

Changed

  • Extended the README.md of the repo to describe usage, design decisions, repo content and how to contribute

Deprecated

None

Removed

None

Fixed

None

[0.0.1] - December 30, 2020

Version number for first release: 0.0.1

Added

  • setup.py to wrap repo as a package.
  • CHANGELOG.md to keep track of changes going forward.
  • __str__ and __repr__ methods to core user-facing method creative_project.TuneSession for improved developer and user experience.

Changed

  • Added conditional to build pipeline so sample_problems only need to pass when merging pull requests in order for code to be considered passing.

Deprecated

None

Removed

None

Fixed

None