Configuration

H-Cut parameters

All parameters can be set via the constructor:

from pgcuts import HyCut

model = HyCut(
    n_clusters=10,          # number of clusters
    objective="hyp_ncut",   # "hyp_ncut", "hyp_rcut", "prcut"
    n_neighbors=50,         # KNN graph neighbors
    steps=3000,             # optimization steps
    batch_size=8192,        # edges per step
    lr=1e-3,                # learning rate
    weight_decay=0.1,       # AdamW weight decay
    m=512,                  # polynomial degree for 2F1
    num_bins=16,            # degree bins (hyp_ncut only)
    tau_start=10.0,         # initial softmax temperature
    tau_end=1.0,            # final softmax temperature
    distance="ce",          # "ce" (cross-entropy) or "xor"
    ema=0.9,                # EMA decay for cluster proportions
    device="cuda",          # "cuda" or "cpu"
    seed=42,                # random seed
)

Parameter guide

Objective (objective)

• "hyp_ncut": Default. Hypergeometric NCut with Holder binning.

• "hyp_rcut": Simpler variant without degree binning.

• "prcut": Original PRCut baseline.

Graph density (n_neighbors)

Controls the KNN graph sparsity. Typical: 20–100. Denser graphs are more robust but slower.

Training steps (steps)

Default 3000 is sufficient for most datasets. Increase for very large K or low graph quality.

Temperature schedule (tau_start, tau_end)

Annealed linearly from tau_start to tau_end. Higher start = softer assignments early; lower end = sharper final clusters.

Polynomial degree (m)

Controls tightness of the hypergeometric bound. Higher = tighter but more computation. Default 512 works well.

Benchmark scripts

For reproducible experiments, use the Hydra-based benchmark:

python scripts/benchmark.py \
    dataset=cifar10 model=dinov2 objective=hyp_ncut

Multi-run sweep:

python scripts/benchmark.py --multirun \
    dataset=cifar10,cifar100,stl10 \
    model=dinov2 \
    objective=prcut,hyp_rcut,hyp_ncut

Config files are in scripts/configs/.