Composition

Squeeze includes wrappers for sequential reductions and experimental combinations. Their fit-time convenience does not give every underlying method a native transform.

Sequential reductions

import numpy as np
from sklearn.datasets import load_digits
from squeeze import DRPipeline, PCA, NeighborMap

X = np.asarray(load_digits().data[:100], dtype=np.float64)
pipeline = DRPipeline([
    ("linear", PCA(n_components=16)),
    ("layout", NeighborMap(n_epochs=40, random_state=42)),
])
embedding = pipeline.fit_transform(X)
assert embedding.shape == (100, 2)

This changes the data presented to NeighborMap and is a different experiment from the raw-pixel benchmarks. For inference, each step must have a suitable transform; do not use a fit-only final reducer as though it were a fitted projection.

Ensemble and progressive wrappers

EnsembleDR(methods=[(name, reducer, weight), ...]) fits each method. Its default weighted_average blend requires weights summing to one. blend_mode="procrustes" uses an alignment-based blend. Arbitrary rotations and scales make direct coordinate averaging difficult to interpret; validate the output rather than assuming a gain. The alignment constructor field alone does not change the default blend mode.

ProgressiveDR(coarse, fine, blend_steps=10, blend_function="linear") fits both reducers and blends their coordinates. It does not use the coarse coordinates to initialize the fine optimizer. For new data, these wrappers may call fit_transform again when a component lacks transform, producing a newly fitted coordinate system. Use OutOfSampleDR explicitly when interpolation is what you intend.

Adaptive selection

AdaptiveDR(method_map, strategy="size") selects size:small below 1,000 rows, size:medium below 100,000, and size:large otherwise. The dimensionality strategy uses dim:low below 50 features, dim:medium below 1,000, and dim:high otherwise. These are routing thresholds, not measured scalability guarantees. Selected estimators must accept fit(X, y) and provide transform; raw Rust fit-only reducers do not satisfy that contract.