GPU Clustering & Balanced Sampling

Pipeline Steps

STEP 0 CPU

Generate & Normalize Data

Create non-uniform point clouds (blobs) and apply StandardScaler to zero-mean, unit-variance. This removes scale bias from distance calculations.

scaler = StandardScaler() X_scaled = scaler.fit_transform(X)

STEP 1 GPU

Farthest Point Sampling

Iteratively pick the point that is farthest from any existing center. Result: evenly spaced centers that cover the entire feature space.

sq_dist = Σ(chunk - center)² next = argmax(min_sq_dists)

STEP 2 GPU

Cluster Assignment

Assign every point to its nearest center using torch.cdist. This creates Voronoi-like partitions. Done in chunks to keep VRAM stable.

dists = torch.cdist(chunk, centers) labels = argmin(dists, dim=1)

STEP 3 CPU

Balanced Sampling

Sample up to N points per cluster. Dense clusters are downsampled, sparse clusters keep all points. This creates a balanced representation of the feature space.

for c in range(n_centers): chosen = rng.choice(idx, N)

Points: 0

Centers: 0

Clusters: 0

Sampled: 0