v0.10.2

v0.10.2 (2022-08-16)

Full Changelog

Fixed

• Fixed a bug in the MMDDrift detector with pytorch backend, where the kernel attribute was not sent to the selected device (#587).

Development

• Code Coverage added (#584).

v0.10.1

v0.10.1 (2022-08-10)

Full Changelog

Fixed

• Corrected a missing optional dependency error when tensorflow was installed without tensorflow-probability (#580).

Development

• An upper version bound has been added for torch (<1.13.0) (#575).

v0.10.0

v0.10.0 (2022-07-26)

Full Changelog

Added

• New feature Drift detectors save/load functionality has been significantly reworked. All offline and online drift detectors (tensorflow backend only) can now be saved and loaded via config.toml files, allowing for more flexibility. Config files are also validated with pydantic. See the documentation for more info (#516).
• New feature Option to use out-of-bag predictions when using a RandomForestClassifier with ClassifierDrift (#426).
• Python 3.10 support. Note that PyTorch at the time of writing doesn't support Python 3.10 on Windows (#485).

Fixed

• Fixed a bug in the TensorFlow trainer which occured when the data was a minibatch of size 2 (#492).

Changed

• TensorFlow is now an optional dependency. Error messages for incorrect use of detectors that are dependent on missing optional dependencies have been improved to include installation instructions and be more informative (#537).
• The optional dependency work has resulted in some imports being reorganised. The original imports will still work as long as the relevant optional dependencies are installed (#538).
  • from alibi_detect.utils.tensorflow.kernels import DeepKernel -> from alibi_detect.utils.tensorflow import DeepKernel
  • from alibi_detect.utils.tensorflow.prediction import predict_batch -> from alibi_detect.utils.tensorflow import predict_batch
  • from alibi_detect.utils.pytorch.data import TorchDataset -> from alibi_detect.utils.pytorch import TorchDataset
  • from alibi_detect.models.pytorch.trainer import trainer -> from alibi_detect.models.pytorch import trainer
  • from alibi_detect.models.tensorflow.resnet import scale_by_instance -> from alibi_detect.models.tensorflow import scale_by_instance
  • from alibi_detect.models.tensorflow.resnet import scale_by_instance -> from alibi_detect.models.tensorflow import scale_by_instance
  • from alibi_detect.utils.pytorch.kernels import DeepKernel -> from alibi_detect.utils.pytorch import DeepKernel
  • from alibi_detect.models.tensorflow.autoencoder import eucl_cosim_features -> from alibi_detect.models.tensorflow import eucl_cosim_features
  • from alibi_detect.utils.tensorflow.prediction import predict_batch -> from alibi_detect.utils.tensorflow import predict_batch
  • from alibi_detect.models.tensorflow.losses import elbo -> from alibi_detect.models.tensorflow import elbo
  • from alibi_detect.models import PixelCNN -> from alibi_detect.models.tensorflow import PixelCNN
  • from alibi_detect.utils.tensorflow.data import TFDataset -> from alibi_detect.utils.tensorflow import TFDataset
  • from alibi_detect.utils.pytorch.data import TorchDataset -> from alibi_detect.utils.pytorch import TorchDataset
• The maximum tensorflow version has been bumped from 2.8 to 2.9 (#508).
• breaking change The detector_type field in the detector.meta dictionary now indicates whether a detector is a 'drift', 'outlier' or 'adversarial' detector. Its previous meaning, whether a detector is online or offline, is now covered by the online field (#564).

Development

• Added MissingDependency class and import_optional for protecting objects that are dependent on optional dependencies (#537).
• Added BackendValidator to factor out similar logic across detectors with backends (#538).
• Added missing CI test for ClassifierDrift with sklearn backend (#523).
• Fixed typing for ContextMMDDrift pytorch backend with numpy>=1.22 (#520).
• Drift detectors with backends refactored to perform distance threshold computation in score instead of predict (#489).
• Factored out PyTorch device setting to utils.pytorch.misc.get_device() (#503). Thanks to @kuutsav!
• Added utils._random submodule and pytest-randomly to manage determinism in CI build tests (#496).
• From this release onwards we exclude the directories doc/ and examples/ from the source distribution (by adding prune directives in MANIFEST.in). This results in considerably smaller file sizes for the source distribution.
• mypy has been updated to ~=0.900 which requires additional development dependencies for type stubs, currently only types-requests and types-toml have been necessary to add to requirements/dev.txt.

v0.9.1

v0.9.1 (2022-04-01)

Full Changelog

Fixed

• Fixed an issue whereby simply importing the library in any capacity caused tensorflow to occupy all available GPU memory. This was due to the instantiation of tf.keras.Model objects within a class definition (GaussianRBF objects within the DeepKernel class).

v0.9.0

v0.9.0 (2022-03-17)

Full Changelog

Added

• Added the ContextMMDDrift detector. The context-aware maximum mean discrepancy drift detector (Cobb and Van Looveren, 2022) is a kernel based method for detecting drift in a manner that can take relevant context into account.

Fixed

• Fixed an issue experienced when the Model uncertainty based drift detection example is run on GPU's (#445).
• Fixed an issue with the Text drift detection on IMDB example to allow PyTorch to be used (#438).

Development

• The maximum tensorflow version has been bumped from 2.7 to 2.8 (#444).

v0.8.1

v0.8.1 (2022-01-18)

Full Changelog

Added

• New feature ClassifierDrift now supports sklearn models (#414). See this example.

Changed

• Python 3.6 has been deprecated from the supported versions as it has reached end-of-life.

Fixed

• The SpectralResidual detector now uses padding to prevent spikes occuring at the beginning and end of scores (#396).
• The handling of url's in the dataset and model fetching methods has been modified to fix behaviour on Windows platforms.

Development

• numpy typing has been updated to be compatible with numpy 1.22 (#403). This is a prerequisite for upgrading to tensorflow 2.7.
• The Alibi Detect CI tests now include Windows and MacOS platforms (#423).
• The maximum tensorflow version has been bumped from 2.6 to 2.7 (#377).

v0.8.0

v0.8.0 (2021-12-09)

Full Changelog

Added

• Offline and online versions of Fisher's Exact Test detector for supervised drift detection on binary data: from alibi_detect.cd import FETDrift, FETDriftOnline.
• Offline and online versions of Cramér-von Mises detector for supervised drift detection on continuous data: from alibi_detect.cd import CVMDrift, CVMDriftOnline.
• Offline supervised drift detection example on the penguin classification dataset.

Changed

• Refactored online detectors to separate updating of state (#371).
• Update tensorflow lower bound to 2.2 due to minimum requirements from transformers.

Fixed

• Fixed incorrect kwarg name in utils.tensorflow.distance.permed_lsdd function (#399).

Development

• Updated sphinx for documentation building to >=4.2.0.
• Added a CITATIONS.cff file for consistent citing of the library.
• CI actions are now not triggered on draft PRs (apart from a readthedoc build).
• Removed dependency on nbsphinx_link and moved examples under doc/source/examples with symlinks from the top-level examples directory.

v0.7.3

v0.7.3 (2021-10-29)

Full Changelog

Added

• DeepKernel is allowed without the kernel_b component, giving a kernel consisting of only a deep kernel component (kernel_a).
• Documentation layout refreshed, and a new "Background to drift detection" added.

Fixed

• Model fetching methods now correctly handle nested filepaths.
• For backward compatibility, fetch and load methods now attept to fetch/load dill files, but fall back to pickle files.
• Prevent dill from extending pickle dispatch table. This prevents undesirable behaviour if using pickle/joblib without dill imported later on (see #326).
• For consistency between save_detector and load_detector, fetch_detector will no longer append detector_name to filepath.

v0.7.2

v0.7.2 (2021-08-17)

Full Changelog

Added

• Learned kernels drift detector with TensorFlow and PyTorch support: from alibi_detect.cd import LearnedKernelDrift
• Spot-the-diff drift detector with TensorFlow and PyTorch support: from alibi_detect.cd import SpotTheDiffDrift
• Online drift detection example on medical imaging data: https://github.com/SeldonIO/alibi-detect/blob/master/examples/cd_online_camelyon.ipynb

v0.7.1

v0.7.1 (2021-07-22)

Full Changelog

Added

• Extend allowed input type for drift detectors to include List[Any] with additional graph and text data examples.
• Allow custom preprocessing steps within alibi_detect.utils.pytorch.prediction.predict_batch and alibi_detect.utils.tensorflow.prediction.predict_batch. This makes it possible to take List[Any] as input and combine instances in the list into batches of data in the right format for the model.

Removed

• PCA preprocessing step for drift detectors.

Fixed

• Improve numerical stability LSDD detectors (offline and online) to avoid overflow/underflow caused by higher dimensionality of the input data.
• Spectral Residual outlier detector test.