feat(datasets) Add ContinuousPartitioner (#5235)

Co-authored-by: jafermarq <javier@flower.ai>

Author: Haoran-Jie

datasets/flwr_datasets/partitioner/__init__.py

@@ -15,6 +15,7 @@
 """Flower Datasets Partitioner package."""
 
 
+from .continuous_partitioner import ContinuousPartitioner
 from .dirichlet_partitioner import DirichletPartitioner
 from .distribution_partitioner import DistributionPartitioner
 from .exponential_partitioner import ExponentialPartitioner
@@ -33,6 +34,7 @@
 from .vertical_size_partitioner import VerticalSizePartitioner
 
 __all__ = [
+    "ContinuousPartitioner",
     "DirichletPartitioner",
     "DistributionPartitioner",
     "ExponentialPartitioner",

datasets/flwr_datasets/partitioner/continuous_partitioner.py

@@ -0,0 +1,207 @@
+# Copyright 2025 Flower Labs GmbH. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Continuous partitioner class that works with Hugging Face Datasets."""
+
+
+# pylint: disable=R0913, R0917
+from typing import Optional
+
+import numpy as np
+
+from datasets import Dataset
+from flwr_datasets.partitioner.partitioner import Partitioner
+
+
+class ContinuousPartitioner(
+    Partitioner
+):  # pylint: disable=too-many-instance-attributes
+    r"""Partitioner based on a real-valued dataset property with adjustable strictness.
+
+    This partitioner enables non-IID partitioning by sorting the dataset according to a
+    continuous (i.e., real-valued, not categorical) property and introducing controlled noise
+    to adjust the level of heterogeneity.
+
+    To interpolate between IID and non-IID partitioning, a `strictness` parameter
+    (𝜎 ∈ [0, 1]) blends a standardized property vector (z ∈ ℝⁿ) with Gaussian noise
+    (ε ~ 𝒩(0, I)), producing blended scores:
+
+
+    .. math::
+
+        b = \sigma \cdot z + (1 - \sigma) \cdot ε
+
+
+    Samples are then sorted by `b` to assign them to partitions. When `strictness` is 0,
+    partitioning is purely random (IID), while a value of 1 strictly follows the property ranking
+    (strongly non-IID).
+
+    Parameters
+    ----------
+    num_partitions : int
+        Number of partitions to create.
+    partition_by : str
+        Name of the continuous feature to partition the dataset on.
+    strictness : float
+        Controls how strongly the feature influences partitioning (0 = iid, 1 = non-iid).
+    shuffle : bool
+        Whether to shuffle the indices within each partition (default: True).
+    seed : Optional[int]
+        Random seed for reproducibility.
+
+    Examples
+    --------
+    >>> from datasets import Dataset
+    >>> import numpy as np
+    >>> import pandas as pd
+    >>> from flwr_datasets.partitioner import ContinuousPartitioner
+    >>> import matplotlib.pyplot as plt
+    >>>
+    >>> # Create synthetic data
+    >>> df = pd.DataFrame({
+    >>>     "continuous": np.linspace(0, 10, 10_000),
+    >>>     "category": np.random.choice([0, 1, 2, 3], size=10_000)
+    >>> })
+    >>> hf_dataset = Dataset.from_pandas(df)
+    >>>
+    >>> # Partition dataset
+    >>> partitioner = ContinuousPartitioner(
+    >>>     num_partitions=5,
+    >>>     partition_by="continuous",
+    >>>     strictness=0.7,
+    >>>     shuffle=True
+    >>> )
+    >>> partitioner.dataset = hf_dataset
+    >>>
+    >>> # Plot partitions
+    >>> plt.figure(figsize=(10, 6))
+    >>> for i in range(5):
+    >>>     plt.hist(
+    >>>         partitioner.load_partition(i)["continuous"],
+    >>>         bins=64,
+    >>>         alpha=0.5,
+    >>>         label=f"Partition {i}"
+    >>>     )
+    >>> plt.legend()
+    >>> plt.xlabel("Continuous Value")
+    >>> plt.ylabel("Frequency")
+    >>> plt.title("Partition distributions")
+    >>> plt.grid(True)
+    >>> plt.show()
+    """
+
+    def __init__(
+        self,
+        num_partitions: int,
+        partition_by: str,
+        strictness: float,
+        shuffle: bool = True,
+        seed: Optional[int] = 42,
+    ) -> None:
+        super().__init__()
+        if not 0 <= strictness <= 1:
+            raise ValueError("`strictness` must be between 0 and 1")
+        if num_partitions <= 0:
+            raise ValueError("`num_partitions` must be greater than 0")
+
+        self._num_partitions = num_partitions
+        self._partition_by = partition_by
+        self._strictness = strictness
+        self._shuffle = shuffle
+        self._seed = seed
+        self._rng = np.random.default_rng(seed)
+
+        # Lazy initialization
+        self._partition_id_to_indices: dict[int, list[int]] = {}
+        self._partition_id_to_indices_determined = False
+
+    def load_partition(self, partition_id: int) -> Dataset:
+        """Load a single partition based on the partition index.
+
+        Parameters
+        ----------
+        partition_id : int
+            The index that corresponds to the requested partition.
+
+        Returns
+        -------
+        dataset_partition : Dataset
+            A single dataset partition.
+        """
+        self._check_and_generate_partitions_if_needed()
+        return self.dataset.select(self._partition_id_to_indices[partition_id])
+
+    @property
+    def num_partitions(self) -> int:
+        """Total number of partitions."""
+        self._check_and_generate_partitions_if_needed()
+        return self._num_partitions
+
+    @property
+    def partition_id_to_indices(self) -> dict[int, list[int]]:
+        """Mapping from partition ID to dataset indices."""
+        self._check_and_generate_partitions_if_needed()
+        return self._partition_id_to_indices
+
+    def _check_and_generate_partitions_if_needed(self) -> None:
+        """Lazy evaluation of the partitioning logic."""
+        if self._partition_id_to_indices_determined:
+            return
+
+        if self._num_partitions > self.dataset.num_rows:
+            raise ValueError(
+                "Number of partitions must be less than or equal to number of dataset samples."
+            )
+
+        # Extract property values
+        property_values = np.array(self.dataset[self._partition_by], dtype=np.float32)
+
+        # Check for missing values (None or NaN)
+        if np.any(property_values is None) or np.isnan(property_values).any():
+            raise ValueError(
+                f"The column '{self._partition_by}' contains None or NaN values, "
+                f"which are not supported by {self.__class__.__qualname__}. "
+                "Please clean or filter your dataset before partitioning."
+            )
+
+        # Standardize
+        std = np.std(property_values)
+        if std < 1e-6 and self._strictness > 0:
+            raise ValueError(
+                f"Cannot standardize column '{self._partition_by}' "
+                f"because it has near-zero std (std={std}). "
+                "All values are nearly identical, which prevents meaningful non-IID partitioning. "
+                "To resolve this, choose a different partition property "
+                "or set strictness to 0 to enable IID partitioning."
+            )
+
+        standardized_values = (property_values - np.mean(property_values)) / std
+
+        # Blend noise
+        noise = self._rng.normal(loc=0, scale=1, size=len(standardized_values))
+        blended_values = (
+            self._strictness * standardized_values + (1 - self._strictness) * noise
+        )
+
+        # Sort and partition
+        sorted_indices = np.argsort(blended_values)
+        partition_indices = np.array_split(sorted_indices, self._num_partitions)
+
+        for pid, indices in enumerate(partition_indices):
+            indices_list = indices.tolist()
+            if self._shuffle:
+                self._rng.shuffle(indices_list)
+            self._partition_id_to_indices[pid] = indices_list
+
+        self._partition_id_to_indices_determined = True