Migrating¶
Migrating from tsxtract (pre-0.7.0 names)¶
0.7.0 renamed the project to Kymora and deleted every old name outright — there are no shims. Numerics, feature names/order, and the NaN/error contracts are unchanged — only names moved:
| Before (≤ 0.6.0) | After (≥ 0.7.0) |
|---|---|
pip install tsxtract-rs |
pip install kymora |
import tsxtract as tsx |
import kymora as km |
tsxtract._core |
kymora._core |
TsxSelector |
KymoraSelector (TsxSelector still works as an alias) |
TSXTRACT_WISDOM / TSXTRACT_POOL |
KYMORA_WISDOM / KYMORA_POOL |
~/.cache/tsxtract/wisdom.json |
~/.cache/kymora/wisdom.json (re-run tune()) |
import kymora is the only spelling — tsxtract and tsxtractor were
deleted with no re-export. Old benchmark
rows and reports still carry the tsxtract library id — that is data, not
a bug; new runs record kymora.
Migrating from tsfresh¶
The shape of the change¶
tsfresh takes a long-format DataFrame and returns a wide one. This library
takes a numpy matrix (or a list of arrays) and returns a matrix.
```python skip
tsfresh (requires the tsfresh package and your own long-format data —¶
not executed by tools/check_snippets.py)¶
from tsfresh import extract_features from tsfresh.feature_extraction import EfficientFCParameters
long_df = pd.DataFrame({"id": ids, "time": times, "value": values}) features = extract_features( long_df, column_id="id", column_sort="time", default_fc_parameters=EfficientFCParameters(), )
kymora¶
import kymora X = values.reshape(n_series, n_steps) # already sorted per series features = kymora.extract_features_df(X) # DataFrame, 33 columns
If your data is already long-format and you want to keep it that way:
```python
import numpy as np
import pandas as pd
import kymora
rng = np.random.default_rng(0)
n_series, n_steps = 4, 50
long_df = pd.DataFrame({
"id": np.repeat(np.arange(n_series), n_steps),
"time": np.tile(np.arange(n_steps), n_series),
"value": rng.standard_normal(n_series * n_steps),
})
groups = long_df.sort_values("time").groupby("id")["value"]
series = [g.to_numpy(dtype="float64") for _, g in groups]
features = kymora.extract_features_df(series) # ragged input is fine
features.index = list(groups.groups) # keep the original ids
Sorting is your responsibility here: there is no column_sort equivalent,
because the input is an array whose order is already meaningful.
Feature-name mapping¶
Many features have a direct tsfresh counterpart under a different name:
| kymora | tsfresh |
|---|---|
mean, std, var, median, min, max |
mean, standard_deviation, variance, median, minimum, maximum |
quantile_10 ... quantile_90 |
quantile__q_0.1 ... quantile__q_0.9 |
skewness, kurtosis |
skewness, kurtosis |
abs_energy |
abs_energy |
root_mean_square |
root_mean_square |
mean_abs_change |
mean_abs_change |
mean_change |
mean_change |
cid_ce |
cid_ce__normalize_True |
mean_second_derivative_central |
mean_second_derivative_central |
zero_crossings |
number_crossing_m__m_0 |
mean_crossings |
no exact equivalent (crossings of the series mean) |
number_of_peaks |
number_peaks__n_3 |
longest_strike_above_mean / _below_mean |
same names |
autocorr_lag_{1,2,5,10} |
autocorrelation__lag_{1,2,5,10} |
trend_slope, trend_r2 |
linear_trend__attr_"slope", linear_trend__attr_"rvalue" (squared) |
permutation_entropy |
permutation_entropy__dimension_3__tau_1 (normalisation differs) |
spectral_entropy |
closest: fourier_entropy__bins_* (definition differs) |
dominant_frequency, spectral_centroid |
no direct equivalent |
Values will not match tsfresh bit for bit where conventions differ — notably
permutation entropy (normalised to [0, 1] here) and the spectral features (DC
bin excluded, sample spacing 1). Treat them as the same concept, not as
drop-in replacements inside a trained model.
Retrain, do not transplant
A model trained on tsfresh columns cannot be fed these columns even where
the names line up. Extract features fresh and retrain.
What has no equivalent here¶
tsfresh's parameterised families — agg_linear_trend, change_quantiles,
fft_coefficient at arbitrary indices, cwt_coefficients, and the rest of the
long tail — are not implemented and will not be. The feature set is closed by
design; see when not to use this. If your model
depends on those, tsfresh is the right tool and this library is not.
Behaviour differences worth knowing¶
| tsfresh | kymora | |
|---|---|---|
| NaN in a series | varies per feature; some impute, some return NaN | every feature for that series is NaN, always |
| Empty series | dropped or NaN-filled | ValueError |
| Input dtype | coerced | float64 required, TypeError otherwise |
| Parallelism | multiprocessing over series, optional | threads over series, always on, GIL released |
| Dependencies | pandas, scipy, statsmodels, tqdm, ... | numpy only |