Changelog#

This is a record of all past rAI-toolbox releases and what went into them, in reverse chronological order. All previous releases should still be available on pip.

0.3.0 - 2023-05-11#

Improvements#

  • Python 3.10 is officially supported.

  • rai_experiments v0.2.0 is available via pypi.

    • rai_experiments.models.pretrained.load_model was added as a means download/cache model weights for rai-toolbox examples and tutorials.

  • Updated Optimizer protocol for compatibility with torch 1.12.0+.

  • Improved documentation formatting and fixed typos.

  • Added format and spell checking to CI.

Compatibility-Breaking changes#

  • Support for MultiRunMetricsWorkflow.evaluation_task was deprecated in v0.2.0 and is now removed.

  • ParamTransformingOptimizer.project was deprecated in v0.2.0 and is now removed.

  • The minimum supported version of PyTorch is now 1.10.0.

  • The minimum supported version of hydra-zen is now 0.9.0.

Deprecations#

  • rai_toolbox.mushin.zen has been deprecated. Use hydra_zen.zen instead.

0.2.1 - 2022-06-16#

This patch fixes the following bugs:

  • TopQGradient device mismatch with user-specified RNG (see pull request #64)

  • MultiRunMetricsWorkflow.to_xarray raises ValueError when target_job_dirs points to job that performed a multirun over sequence-type values (see pull request #68)

0.2.0 - 2022-06-01#

This release predominantly focuses on improvements to rai_toolbox.mushin.MultiRunMetricsWorkflow, which is still in early-beta and may be subject to substantial future changes

Improvements to MultiRunMetricsWorkflow#

  • A pre_task step can be defined for a workflow; this is useful for seeding random number generators prior to the task’s instantiation phase. See this how-to guide for examples.

  • Loaded workflow overrides now roundtrip appropriately. See pull request #61.

  • metric_load_fn can be overridden to customize how MultiRunMetricsWorkflow loads metric files; the default behavior is to use torch.load. See pull request #63.

  • working_subdir can be included as a data-variable in a workflow’s xarray; this enables users to lookup subdirs by override values. See pull request #52.

  • to_xarray works on lists of array-likes, not just lists of numpy arrays

  • load_metrics can be called directly from MultiRunMetricsWorkflow.

  • load_metrics and to_xarray support loading multiple files; multiple file names can be specified as a sequence and/or a glob patterns can be provided.

Other Improvements#

  • The HydraDDP callback now supports pytorch_lightning.Trainer.predict

  • The project’s CI now has a nightly job that runs our test suite against pre-releases of its dependencies

Compatibility-Breaking changes#

rai_toolbox.mushin.hydra.launch was removed. Use hydra_zen.launch instead.

Deprecations#

  • MultiRunMetricsWorkflow.evaluation_task is deprecated in favor of MultiRunMetricsWorkflow.task. It will be removed in rAI-toolbox v0.3.0. See pull request #62.

  • ParamTransformingOptimizer.project is now deprecated in favor of ParamTransformingOptimizer._post_step_transform_. It will be removed in rAI-toolbox v0.3.0. See pull request #54 and pull request #59 for details.

0.1.1 - 2022-05-10#

This patch fixes two bugs in rai_toolbox.perturbations.init:

  • uniform_like_l1_n_ball_ was not correctly symmeterized; the drawn values only had components in the direction of the positive hemisphere of the \(L^1\) ball.

  • Passing an on-gpu tensor to the in-place init functions would cause a device mismatch error with the default random number generator, which is on CPU.

0.1.0 - 2022-05-04#

This is rAI-toolbox’s first stable release on PyPI! Please check out the rest of our docs to see what the toolbox has to offer.

We plan to have an aggressive release schedule for compatibility-preserving patches of bug-fixes and quality-of-life improvements (e.g. improved type annotations), and to regularly add features. Experimental parts of the toolbox’s API that may undergo significant changes in future releases are documented as such.