edmkit

Why edmkit?

Fast

Built on NumPy, SciPy, and usearch for high-performance nearest-neighbor search and numerical computation.

Simple API

Module-level functions with clear signatures. No unnecessary class hierarchies.

Flexible

Supports batched operations, custom metrics, pluggable prediction functions, and optional tinygrad GPU acceleration.

Complete

Covers the full EDM workflow: embedding, simplex projection, S-Map, CCM, cross-validation, and synthetic data generation.

Quick Example

import numpy as np
from edmkit.generate import lorenz
from edmkit.embedding import lagged_embed, scan, select
from edmkit.simplex_projection import simplex_projection

# Generate Lorenz attractor
t, X = lorenz(sigma=10, rho=28, beta=8/3,
              X0=np.array([1.0, 1.0, 1.0]), dt=0.01, t_max=50)
x = X[:, 0]  # Use x-component

# Find optimal embedding parameters
scores = scan(x, None, E=list(range(1, 11)), tau=[1, 2, 3, 5])
best_E, best_tau, best_score = select(scores, E=list(range(1, 11)), tau=[1, 2, 3, 5])

# Embed and predict
embedded = lagged_embed(x, tau=best_tau, e=best_E)
n = len(embedded)
lib = embedded[:n // 2]
target = x[best_tau * (best_E - 1) + 1 : n // 2 + best_tau * (best_E - 1) + 1]
query = embedded[n // 2:]

predictions = simplex_projection(lib, target, query)