Ajb/tde notebook - #3765
Draft
TonyBagnall wants to merge 46 commits into
Draft
Conversation
- SFA._transform_case: build whole word bags in single njit calls (_create_bag_flat/_create_bag_pyramid) instead of per-window typed Dict operations from Python - TDE nearest neighbour: convert bags once to sorted key/value arrays and compute histogram intersections with numba merge kernels instead of per-pair dict lookups - SFA._mcb: vectorise rounding; SFA._binning_dft: replace linspace/np.split with direct slicing - TDE fit 12.8s -> 1.5s on a 120x250 benchmark, outputs verified bit-identical across univariate/multivariate, levels 1-3, IGB/MCB, typed/untyped, loocv train estimates Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- loocv_train_acc: whole LOOCV in one numba call computing each symmetric pair intersection once (upper triangle), preserving the sequential early-abandon semantics; used by _individual_train_acc and _select_dims - replace StandardScaler + KernelRidge parameter selection with the same computation in plain numpy (_kernel_ridge_preds), and keep a float array mirror of the candidate parameters instead of re-converting the list every iteration - trig lookup tables (cos/sin of (n*i) mod w) in _binning_dft_all and the first-window DFT of _mft_all, removing all per-term trig calls - predict: one nn_similarities_all call per member for all test cases instead of a joblib task per case, and lazy tie-break rng construction (seeding a RandomState per test case dominated predict) - _combine_dim_bags: numba k-way merge of the already-sorted per-dimension bags instead of per-case double stable argsorts Evaluated and rejected: flat levels==1 NN kernels (~6% kernel gain only) and an ensemble-level per-window DFT cache (slower at both 40 and 250 parameter samples once the trig tables landed). Benchmark (120x250 univariate 40 members / BasicMotions 10 members): uni fit 703 -> 234 ms, uni predict 686 -> 449 ms, mv fit 147 -> 55 ms, mv predict 36 -> 21 ms. Outputs verified identical on univariate and multivariate train/test probabilities; all TDE, expected-output, ordinal and estimator-check suites pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- fix typos (lcoefficients, TemporalDictinaryEnsemble, 'to use to use', contacting, para space, each dimensions) and call histogram intersection a similarity rather than a distance - update the typed_dict parameter descriptions: it has no effect on newly fitted models and is retained for unpickling older models - deprecate IndividualTDE alphabet_size (fixed to 4 by the new SFA, other values ignored): docstring .. deprecated:: note plus a DeprecationWarning in fit when a non-default value is used, removal targeted for v1.7.0 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rework the earlier deprecations to match docs/developer_guide/deprecation.md: sentinel 'deprecated' defaults, a FutureWarning raised in __init__ only when a value is passed, TODO removal comments for v1.7.0, and plain docstring deprecation notes. The ensemble no longer forwards either parameter to its members, and the unused _alphabet_size attribute is removed. Also correct the parameter selection description: TDE uses kernel ridge regression (the reference paper describes this step as a Gaussian process regressor). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- test_tde.py: cover the loocv train estimate path plus ensemble predict/predict_proba, the deprecated alphabet_size/typed_dict FutureWarnings, numerical equivalence of the numpy kernel ridge helper with sklearn StandardScaler + KernelRidge, and agreement of the per-case LOOCV fallback with the symmetric kernel (via the new _SYMMETRIC_LOOCV_MAX_N constant); pickle test now also checks predictions survive the round trip and no longer passes the deprecated typed_dict; multivariate test made deterministic with dim_threshold=0 - test_tde_sfa.py: direct tests for loocv_train_acc (hand-checked predictions, accuracy count and the early-abandon sentinel), nn_similarities_all (full test x train matrix) and combine_dim_bags (levels > 1 key shifting including the -1 bigram tag, and flat dimension tagging) - remove the now-unused nn_similarities kernel (superseded by nn_similarities_all) Coverage with NUMBA_DISABLE_JIT=1: _tde_sfa.py 99% (0 lines missed), _tde.py 55% with remaining gaps confined to legacy typed-Dict pickle paths and branches exercised by the global estimator checks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- test_tde.py: module and per-test docstrings now state what behaviour each test pins down and why (oob voting, validation/correction on fit, histogram intersection input types, imbalanced subsampling) - _tde_sfa.py: standard aeon header with __maintainer__ and __all__ - TonyBagnall and MatthewMiddlehurst added as maintainers of both TDE modules Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Models pickled with aeon <= 1.5 could not be used with the rewritten TDE anyway (verified: predict fails on an unpickled old model because the old SFA transformer output format is no longer handled), so the code retained for that path was dead: - __getstate__/__setstate__ typed Dict conversions (array bags pickle natively), the _typed_dict attribute and the _bags_cache / _get_bag_arrays conversion cache - the Dict-bag branches of _predict, _train_predict and _individual_train_acc, the _test_nn method and the joblib threading branch they used - the eight module-level legacy numba kernels (_bags_to_arrays_*, _intersection_*, _nn_index_*, _histogram_intersection_to_all_*) - the now-unused _is_tde_sfa_bags helper and os/joblib/NumbaList imports The typed_dict parameter docstrings no longer claim old models can be unpickled. histogram_intersection and its numba helper are kept: they are public API used by OrdinalTDE and the tests. All test lanes pass (jit and no-jit), estimator checks pass, and a multivariate pickle round trip is prediction-identical. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- nogil=True on the predict-path numba kernels: plain njit holds the GIL, so the previous joblib threading could never scale - TemporalDictionaryEnsemble._predict_proba threads members' _predict calls (X validated once by the public wrapper); results gathered in member order so probabilities are identical for any n_jobs - IndividualTDE._predict chunks test cases across threads, stacking chunk similarity matrices in order - tie-breaking moved off the hot path exactly: with an integer seed every case's fresh RandomState yields the same draw sequence, so one precomputed pool + the nn_tie_break kernel resolves all cases in numba; unseeded generators keep the sequential python walk but only for rows nn_first_max flags as having tie events. This removes ~2000 GIL-bound RandomState seedings per predict, cutting single-threaded predict 495 -> 207 ms on the 120x250 benchmark and letting threads scale: 76 ms at n_jobs=2, ~70 ms at n_jobs=8 (7x vs the previous sequential predict) - ensemble members are constructed single threaded (the ensemble owns predict parallelism); n_jobs docstrings state fit is single threaded - new tests: n_jobs equivalence for both classes and unit tests for nn_first_max/nn_tie_break; the framework multithreading checks (including the MULTITHREAD_TESTING-gated results comparison) pass Outputs verified identical to the sequential implementation on the univariate and multivariate snapshots. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
References the TDE paper and describes TDE's role as the dictionary based component of HIVE-COTE 2.0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Repair the mangled intro sentence, defer the merge mechanics to the comparison section instead of pre-explaining them in the motivation, explain why the standalone transform can stand in for the classifier's in the by-hand prediction, rename the LOOCV reproduction variables to avoid clashing with the prediction ones, trim the overlapping channel selection segue and remove a leftover part-number reference. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
# Conflicts: # aeon/classification/dictionary_based/_tde.py
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
aeon-actions-bot
Bot
added
classification
Contributor
Thank you for contributing to
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
holder for TDE notebook