Repository navigation
Feat/tsfresh parity extractors - #21
Merged
Merged
Conversation
tsfresh_compare.py replaces minimal_bench.py, efficient_bench.py, 1brc/benchmark.py and the stale efficient.json. tsfresh now runs exactly the calculators, parameters and attributes tsfx runs for the tier (derived from .tsfx-config.toml, checked by feature count), both get the same pre-sorted in-memory data with CSV parsing outside the timer, tsfresh's n_jobs matches the Polars thread pool, and both get a warm-up and repeated runs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The noise estimate selected the 10th percentile by sorting a len/20 window per ridge line, and every local maximum was compared with every ridge line. Select the two order statistics instead, match by binary search over the lines' sorted end columns (keeping np.argmin's first-line tie-break), and end lines with one partition instead of repeated Vec::remove. Results are bit-identical (snapshot and 11,857 number_cwt_peaks values on 1BRC and edge-case series). On 50M 1BRC rows: number_cwt_peaks 10.9 s -> 2.6 s, Efficient tier 21.6 s -> 13.7 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
statsmodels 0.15 fits an AR(k) model with exactly as many observations as parameters (0.14 raised ZeroDivisionError, which tsfresh turned into NaN, or 0 for coeff = k). CI's fuzz job did not pin statsmodels, so it picked up 0.15 and failed on ar_coefficient([1.0], k=0): tsfresh 1.0, tsfx 0.0. Fit from 2k + 1 values on, and pin statsmodels==0.15.0 in the fuzz job. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bit-identical (snapshot plus ~12,900 series of 1BRC and edge-case data per feature). Efficient tier on 100M 1BRC rows: 35.3 s -> 25.5 s. - augmented_dickey_fuller: the exact-fit check computed the fit's scale (O(nobs * p) per candidate lag) for every fit. Compare against the cheap bound max|y| + sum_j |b_j| max|X_j| first; the exact scale is only needed when the residual is within twice the tolerance of that bound, so the decision is unchanged. The Jacobi SVD also sums its three dot products in one pass (same order). 16.6 s -> ~10 s. - permutation_entropy: count patterns as base-`dimension` integers (same order as the rank sequences) instead of allocating a Vec per window; dimensions too large for u64 keep the Vec path. 3.7 s -> 2.1 s. - change_quantiles: compute each (ql, qh, isabs) corridor's changes once and aggregate a copy (the median reorders it). 3.0 s -> 2.5 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With pyo3's abi3 feature, --find-interpreter builds only the abi3 wheel, so 0.5.0 shipped no wheels for free-threaded CPython or PyPy (0.4.0 had cp314t and pp311 on every Linux target) and those installs fell back to the sdist. Name the interpreters instead: python3.11 (abi3), python3.14t and pypy3.11. maturin has their configuration built in, so they need not be installed in the build images (checked locally for both). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- _ar_designs left out the exactly determined designs (n = 2k + 1) that statsmodels 0.15 now fits, so their conditioning went unchecked: the thorough profile found [6519, 1, 0, 0, 1] (k = 2, condition number 6e7), where statsmodels is off by 2e-9 of the largest coefficient (tsfx gives the exact 6518 to within 3e-12). - NaN or +-inf inside an ADF regression design can make LAPACK's SVD loop forever in statsmodels' adfuller (autolag=None), which hung the thorough fuzz run. Only a non-finite last value, which enters no design, is still compared; the parity tests check tsfx's NaN for the others. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Householder QR in Ols::fit_prefixes sweeps the whole design once per column; for ADF on long series (42 columns x 11k rows, 3.7 MB) that was memory-bandwidth bound and most of ADF's time. Build X^T X and X^T y in one pass over cache-sized row blocks, with column scaling, and use the Cholesky factor, whose leading blocks serve every lag prefix as the QR's did. Not bit-identical: the QR path is still used, unchanged, when the scaled condition number bound reaches 1e3, when the residual (a difference, y'y - |c|^2) has an estimated relative error above 1e-12, or when a fit could be exact. On 1BRC and edge-case series, ADF results agree with the QR path's within 1e-12 relative with the same chosen lags, AR coefficients within 2e-11 of the largest coefficient. ADF alone on 100M 1BRC rows: 12.5 s -> 4.3 s. tsfresh comparison on 1BRC: 39x at 10M rows (1.3 s vs 52.3 s), 27x at 50M (7.1 s vs 192.7 s). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.
No description provided.