← All 31 animations

Notebook example metric: root mean square of input - (trend + season + residual) over all samples and channels. This checks additive bookkeeping; it is not a forecast error or a component-recovery score. The separate teaching input and intermediate arrays for the animation are linked below.

How it works

Method maturity: experimental. Implementation: native-backed.

DeTime's experimental operator builds windowed Fourier atoms, assigns their features to a supplied codebook, and reconstructs cluster-specific contributions with overlap-add.

Deterministic synthetic teaching signal; not a benchmark.

Open GIF · Static PNG · Source arrays (NPZ) · Trace metadata

Follow the four steps

1. Window and transform

Slide the normalized Gaussian window and compute real FFT coefficients.

2. Build atom features

Represent each time-frequency bin by normalized time, normalized frequency, and log amplitude.

3. Assign the codebook

Standardize features with the stored mean/scale and select the nearest supplied centroid.

4. Mask and reconstruct

Retain each cluster's complex coefficients, apply inverse FFT and normalized overlap-add, then group contributions heuristically.

Parameters that change the result

What this animation computes

Uses the current repository's window, STFT, feature extraction and inverse-overlap-add functions. Squared Euclidean assignment matches the supplied-centroid operation; native cluster reconstructions are independently compared when available. Model normalization, fixed teaching centroids, and numeric window settings are saved with the trace. The public call requires a GaborClusterModel or a model archive created with its save() method; the generic config is incomplete until that model is supplied.

Interpretation

• This is an experimental, repository-specific operator, not a claimed reproduction of a canonical forecasting or decomposition architecture.
• The fixed four-centroid teaching codebook is neither trained nor evidence of a useful scientific clustering.
• The current frame grid can leave an uncovered tail when the record length is not aligned with the hop; the residual retains that difference.
• Cluster labels are codebook labels; assigning low-frequency clusters to trend is a heuristic.

Try the method

The public method requires a GaborClusterModel object or a saved model_path. The animation generator constructs an explicitly fixed teaching codebook; the arrays are included in its source archive. Supply a model before calling decompose with the configuration below.

{
  "method": "GABOR_CLUSTER",
  "params": {
    "trend_freq_thr": 0.08
  }
}

See the configuration reference for the full interface and optional dependencies.

Sources and reproduction

• DeTime Gabor operator source
• DeTime native Gabor kernels
• DeTime implementation: src/detime/methods/gabor_cluster.py
• DeTime implementation: native/src/gabor.cpp

Regenerate this animation with python scripts/generate_method_animations.py --methods GABOR_CLUSTER. The reproduction guide explains the environment, numerical checks, coverage manifest, and how to inspect the stored arrays.

Code

case = CASES["GABOR_CLUSTER"]
print(case["skip"])

Output

requires a trained GaborClusterModel or model_path plus the experimental clustering backend