Skip to content

v0.3.0

Choose a tag to compare

@julien-bloch julien-bloch released this 10 Oct 12:59
· 73 commits to main since this release
f965e1e
  • Prepared initial open-source release.
  • Added reusable ML primitives for multi-output prediction, sample weighting, and sampled-domain validation.
  • Updated public tutorials, runnable examples, API notes, and LLM coding-agent guidance.
  • Added an EEG motor imagery example and tutorial that compares xdflow with plain scikit-learn on PhysioNet data.
  • Added a human activity recognition example, tutorial and Colab notebook (UCI HAR smartphone data) on subject-grouped vs. shuffled-window validation.

Added

  • CrossValidator.oof_predictions_container_: out-of-fold predictions with their trial coordinates and a
    fold coordinate, so CV results can be broken down by subject, session or any other metadata.
  • MNE-Python interop: from_mne_epochs(epochs, subject=...) converts Epochs to a (trial, channel, time)
    container with event labels, metadata columns and extra trial coordinates; CSPTransform (new mne extra)
    wraps mne.decoding.CSP as a stateful step.
  • scikit-learn wrappers align new data to the feature/channel labels seen at fit before predicting, so input
    whose channels arrive in another order gives the same predictions (missing labels raise).
  • SKLearnPredictor.from_estimator(...) wraps an existing scikit-learn estimator or Pipeline instance, so
    existing models can use xdflow validation, caching and tuning unchanged. Predictors accept N-D input
    sample-first (e.g. trial × channel × time for an MNE CSP pipeline).

Fixed

  • Hyperparameters set after construction (set_params, the Tuner, seeds) now reach every CV fold and
    clone. Previously sklearn estimator params were rebuilt from constructor values, so tuning had no effect.
  • Cache keys hash full data and labels instead of a 32-value sample, so use_cache and the CV fold
    cache can no longer return results computed on different data. A cache hit never skips fitting a
    stateful pipeline.
  • Clones get their own copy of a fitted LabelEncoder, so finalizing several pipelines from one
    validator no longer relabels earlier models.
  • GroupedKFoldValidator holdout groups were chosen by trial label as if it were a position.
  • GroupApplyTransform with a Pipeline template ending in a predictor can predict and be
    cross-validated. Previously predict() raised TypeError and validators could not resolve scoring.
  • GroupApplyTransform and domain aligners return samples in input order. GroupApply's unseen-group
    averaging votes on probabilities rather than hard labels.
  • Regression ensembles average member predictions; score-based ensemble weights compare decoded labels.
  • ZCAWhitening whitens (it returned its input); GlobalColoringProjection keeps the top components.
  • PCA/ZCA apply fitted matrices by channel label, not position.
  • Confusion matrices keep rows for classes absent from a split; scoring is re-resolved after changes.
  • sklearn metric functions (e.g. accuracy_score) work as scoring; "accuracy" is a string scorer.
  • Pipelines ending in EnsemblePredictor can be cross-validated. Tuner pruners receive fold scores.
  • HilbertPhaseTransform keeps float output for integer input, and relative_average uses a circular mean.
  • UnflattenTransform keeps coords that vary along unstacked dims.
  • SampleWeightTransform keeps fractional weights; stratified_sample no longer reseeds the global RNG.
  • BalanceClassWeightTransform computes weights per training fold inside CV.
  • LGBMPredictor early stopping works with multiple targets.
  • SklearnCVAdapter recounts splits for each container.
  • Cache keys distinguish different functions/closures, ignore verbose, and include config exposed only via
    get_params (e.g. AdaptWrapperTransform estimator kwargs).
  • GroupApply and domain aligners restore input order without needing unique trial labels. GroupApply's
    unseen-group averaging keeps classes only some group models saw, and falls back to a label vote for
    models without predict_proba.
  • Confusion matrices use the class order from evaluation time, even after finalize_pipeline.
  • Estimator factories (lambdas, functools.partial) clone; Tuner seeds reach GroupApply templates.
  • Cache keys cover ufuncs and other callable instances, bound methods, keyword-only defaults and recursive
    closures, and are stable across processes. Nested GroupApply/aligner steps keep input order.
  • LGBMPredictor early stopping works with target patterns resolving to several targets.
  • RegressOutReferenceTransform(use_fit=True) refuses channels it was not fitted on.

Changed

  • set_params raises for names that are not constructor arguments. Aligners and EnsemblePredictor
    raise for unknown constructor kwargs. Writes to an automatic MultiOutput wrapper's own params raise.
    All of these were previously ignored silently.
  • Domain aligners raise when groups have different class sets. Cross-validation raises on duplicate
    trial labels.
  • Pass-through transforms (e.g. IdentityTransform) return a new container instead of the input.
  • Tuner defaults to use_mlflow=False.
  • Stateless ZScoreTransform/DemeanTransform/RemoveOutliersTransform warn when they pool statistics
    across trials (which leaks validation trials inside CV); use use_fit=True or include trial in per_dim.
  • Cache diagnostics print only when called with verbose=True.
  • Domain aligners build their strategy at fit; get_strategy_and_kwargs was removed. Aligners pickled with
    older versions still load and clone.
  • RegressOutReferenceTransform gains use_fit (default behaviour unchanged) and warns when pooling trials.
  • CropTimeTransform arguments renamed time_window_start_ms/time_window_end_ms -> start/end (they are in
    the time coordinate's units, not milliseconds), and a window that selects no samples raises.
  • BandpassFilterTransform requires fs (the old default of 500 Hz silently mis-filtered other sampling
    rates) and no longer needs the spectral extra; only MultiTaperTransform does.

Refit required for saved models

  • Models containing ZCAWhitening/ZCATransform: the old transform returned its input. Unpickled old
    instances now raise instead of silently producing different features.
  • Models using HilbertPhaseTransform(mode="relative_average"): features change (circular mean). This
    transform is stateless, so old pickles cannot detect it; refit them.