Data partitioners

Dataset-to-worker partitioning strategies for simulations.

A DataPartitioner turns one dataset into n per-worker Dataset instances. Both CentralisedSimulation and DecentralisedSimulation consume this same shape, so partitioning is entirely the caller’s responsibility, IID or not. Wrapping each worker’s dataset into a DataLoader (batch size, shuffling) is the simulation’s job, not the partitioner’s — that separation is what lets partitioners compose (e.g. mixing two partitioners’ outputs) without reaching back into a DataLoader to get at the underlying samples.

Like aggregators and attacks, each strategy is stateless: a @classmethod invoked directly on the class. The dataset is the sole positional argument; n, seed, and any partitioner-specific hyperparameters are keyword-only.

Available partitioners

• IidPartitioner
• DirichletPartitioner
• PerLabelsPartitioner
• MixingPartitioner

Base class

class krum.primitives.data_partitioners.DataPartitioner

Bases: ABC

Abstract base class for stateless dataset-to-worker partitioners.

Subclasses implement partition() as a @classmethod — no instance state is required, and the caller invokes the strategy directly on the class. The dataset is the sole positional argument; n, seed, and any partitioner-specific hyperparameters are keyword-only.

abstract classmethod partition(dataset: Dataset[Any], /, *, n: int, seed: int = 42, **specialized: Any) → Sequence[Dataset[Any]]

Split dataset into n per-worker datasets.

Parameters:

• dataset – Full dataset to partition across workers.

• n – Number of workers to split the dataset across.

• seed – Random seed for reproducibility.

• **specialized – Keyword-only arguments specific to each partitioning strategy (e.g. \(\alpha\) for label skew).

Returns:

Sequence of ``n`` datasets, one per worker.

Raises:

NotImplementedError – If the subclass does not implement this method.