Documentation for Nebuladataset Module

NebulaDataset

Source code in nebula/core/datasets/nebuladataset.py

class NebulaDataset:
    def __init__(
        self,
        num_classes=10,
        partitions_number=1,
        batch_size=32,
        num_workers=4,
        iid=True,
        partition="dirichlet",
        partition_parameter=0.5,
        seed=42,
        config_dir=None,
    ):
        self.num_classes = num_classes
        self.partitions_number = partitions_number
        self.batch_size = batch_size
        self.num_workers = num_workers
        self.iid = iid
        self.partition = partition
        self.partition_parameter = partition_parameter
        self.seed = seed
        self.config_dir = config_dir

        logging.info(
            f"Dataset {self.__class__.__name__} initialized | Partitions: {self.partitions_number} | IID: {self.iid} | Partition: {self.partition} | Partition parameter: {self.partition_parameter}"
        )

        # Dataset
        self.train_set = None
        self.train_indices_map = None
        self.test_set = None
        self.test_indices_map = None
        self.local_test_indices_map = None

        enable_deterministic(self.seed)

    @abstractmethod
    def initialize_dataset(self):
        """
        Initialize the dataset. This should load or create the dataset.
        """
        raise NotImplementedError("Subclasses must implement this method.")

    def clear(self):
        """
        Clear the dataset. This should remove or reset the dataset.
        """
        if self.train_set is not None and hasattr(self.train_set, "close"):
            self.train_set.close()
        if self.test_set is not None and hasattr(self.test_set, "close"):
            self.test_set.close()

        self.train_set = None
        self.train_indices_map = None
        self.test_set = None
        self.test_indices_map = None
        self.local_test_indices_map = None

    def data_partitioning(self, plot=False):
        """
        Perform the data partitioning.
        """

        logging.info(
            f"Partitioning data for {self.__class__.__name__} | Partitions: {self.partitions_number} | IID: {self.iid} | Partition: {self.partition} | Partition parameter: {self.partition_parameter}"
        )

        self.train_indices_map = (
            self.generate_iid_map(self.train_set)
            if self.iid
            else self.generate_non_iid_map(self.train_set, self.partition, self.partition_parameter)
        )
        self.test_indices_map = self.get_test_indices_map()
        self.local_test_indices_map = self.get_local_test_indices_map()

        if plot:
            self.plot_data_distribution("train", self.train_set, self.train_indices_map)
            self.plot_all_data_distribution("train", self.train_set, self.train_indices_map)
            self.plot_data_distribution("local_test", self.test_set, self.local_test_indices_map)
            self.plot_all_data_distribution("local_test", self.test_set, self.local_test_indices_map)

        self.save_partitions()

    def get_test_indices_map(self):
        """
        Get the indices of the test set for each participant.

        Returns:
            A dictionary mapping participant_id to a list of indices.
        """
        try:
            test_indices_map = {}
            for participant_id in range(self.partitions_number):
                test_indices_map[participant_id] = list(range(len(self.test_set)))
            return test_indices_map
        except Exception as e:
            logging.exception(f"Error in get_test_indices_map: {e}")

    def get_local_test_indices_map(self):
        """
        Get the indices of the local test set for each participant.
        Indices whose labels are the same as the training set are selected.

        Returns:
            A dictionary mapping participant_id to a list of indices.
        """
        try:
            local_test_indices_map = {}
            test_targets = np.array(self.test_set.targets)
            for participant_id in range(self.partitions_number):
                train_labels = np.array([self.train_set.targets[idx] for idx in self.train_indices_map[participant_id]])
                indices = np.where(np.isin(test_targets, train_labels))[0].tolist()
                local_test_indices_map[participant_id] = indices
            return local_test_indices_map
        except Exception as e:
            logging.exception(f"Error in get_local_test_indices_map: {e}")
            raise

    def save_partition(self, obj, file, name):
        try:
            logging.info(f"Saving pickled object of type {type(obj)}")
            pickled = pickle.dumps(obj)
            ds = file.create_dataset(name, data=np.void(pickled))
            ds.attrs["__type__"] = "pickle"
            logging.info(f"Saved pickled object of type {type(obj)} to {name}")
        except Exception as e:
            logging.exception(f"Error saving object to HDF5: {e}")
            raise

    def save_partitions(self):
        """
        Save each partition data (train, test, and local test) to separate pickle files.
        The controller saves one file per partition for each data split.
        """
        try:
            logging.info(f"Saving partitions data for ALL participants ({self.partitions_number}) in {self.config_dir}")
            path = self.config_dir
            if not os.path.exists(path):
                raise FileNotFoundError(f"Path {path} does not exist")
            # Check that the partition maps have the expected number of partitions
            if not (
                len(self.train_indices_map)
                == len(self.test_indices_map)
                == len(self.local_test_indices_map)
                == self.partitions_number
            ):
                raise ValueError("One of the partition maps has an unexpected length.")

            # Save global test data
            file_name = os.path.join(path, "global_test.h5")
            with h5py.File(file_name, "w") as f:
                indices = list(range(len(self.test_set)))
                test_data = [self.test_set[i] for i in indices]
                self.save_partition(test_data, f, "test_data")
                f["test_data"].attrs["num_classes"] = self.num_classes
                test_targets = np.array(self.test_set.targets)
                f.create_dataset("test_targets", data=test_targets, compression="gzip")

            for participant in range(self.partitions_number):
                file_name = os.path.join(path, f"participant_{participant}_train.h5")
                with h5py.File(file_name, "w") as f:
                    logging.info(f"Saving training data for participant {participant} in {file_name}")
                    indices = self.train_indices_map[participant]
                    train_data = [self.train_set[i] for i in indices]
                    self.save_partition(train_data, f, "train_data")
                    f["train_data"].attrs["num_classes"] = self.num_classes
                    train_targets = np.array([self.train_set.targets[i] for i in indices])
                    f.create_dataset("train_targets", data=train_targets, compression="gzip")
                    logging.info(f"Partition saved for participant {participant}.")

            logging.info("Successfully saved all partition files.")

        except Exception as e:
            logging.exception(f"Error in save_partitions: {e}")

        finally:
            self.clear()
            logging.info("Cleared dataset after saving partitions.")

    @abstractmethod
    def generate_non_iid_map(self, dataset, partition="dirichlet", plot=False):
        """
        Create a non-iid map of the dataset.
        """
        pass

    @abstractmethod
    def generate_iid_map(self, dataset, plot=False):
        """
        Create an iid map of the dataset.
        """
        pass

    def plot_data_distribution(self, phase, dataset, partitions_map):
        """
        Plot the data distribution of the dataset.

        Plot the data distribution of the dataset according to the partitions map provided.

        Args:
            phase: The phase of the dataset (train, test, local_test).
            dataset: The dataset to plot (torch.utils.data.Dataset).
            partitions_map: The map of the dataset partitions.
        """
        logging_training.info(f"Plotting data distribution for {phase} dataset")
        # Plot the data distribution of the dataset, one graph per partition
        sns.set()
        sns.set_style("whitegrid", {"axes.grid": False})
        sns.set_context("paper", font_scale=1.5)
        sns.set_palette("Set2")

        # Plot bar charts for each partition
        partition_index = 0
        for partition_index in range(self.partitions_number):
            indices = partitions_map[partition_index]
            class_counts = [0] * self.num_classes
            for idx in indices:
                label = dataset.targets[idx]
                class_counts[label] += 1

            logging_training.info(f"[{phase}] Participant {partition_index} total samples: {len(indices)}")
            logging_training.info(f"[{phase}] Participant {partition_index} data distribution: {class_counts}")

            plt.figure(figsize=(12, 8))
            plt.bar(range(self.num_classes), class_counts)
            for j, count in enumerate(class_counts):
                plt.text(j, count, str(count), ha="center", va="bottom", fontsize=12)
            plt.xlabel("Class")
            plt.ylabel("Number of samples")
            plt.xticks(range(self.num_classes))
            plt.title(
                f"[{phase}] Participant {partition_index} data distribution ({'IID' if self.iid else f'Non-IID - {self.partition}'} - {self.partition_parameter})"
            )
            plt.tight_layout()
            path_to_save = f"{self.config_dir}/participant_{partition_index}_data_distribution_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
            logging_training.info(f"Saving data distribution for participant {partition_index} to {path_to_save}")
            plt.savefig(path_to_save, dpi=300, bbox_inches="tight")
            plt.close()

        if hasattr(self, "tsne") and self.tsne:
            self.visualize_tsne(dataset)

    def visualize_tsne(self, dataset):
        X = []  # List for storing the characteristics of the samples
        y = []  # Ready to store the labels of the samples
        for idx in range(len(dataset)):  # Assuming that 'dataset' is a list or array of your samples
            sample, label = dataset[idx]
            X.append(sample.flatten())
            y.append(label)

        X = np.array(X)
        y = np.array(y)

        tsne = TSNE(n_components=2, verbose=1, perplexity=40, n_iter=300)
        tsne_results = tsne.fit_transform(X)

        plt.figure(figsize=(16, 10))
        sns.scatterplot(
            x=tsne_results[:, 0],
            y=tsne_results[:, 1],
            hue=y,
            palette=sns.color_palette("hsv", self.num_classes),
            legend="full",
            alpha=0.7,
        )

        plt.title("t-SNE visualization of the dataset")
        plt.xlabel("t-SNE axis 1")
        plt.ylabel("t-SNE axis 2")
        plt.legend(title="Class")
        plt.tight_layout()

        path_to_save_tsne = f"{self.config_dir}/tsne_visualization.png"
        plt.savefig(path_to_save_tsne, dpi=300, bbox_inches="tight")
        plt.close()

    def dirichlet_partition(
        self,
        dataset: Any,
        alpha: float = 0.5,
        min_samples_size: int = 50,
        balanced: bool = False,
        max_iter: int = 100,
        verbose: bool = True,
    ) -> dict[int, list[int]]:
        """
        Partition the dataset among clients using a Dirichlet distribution.
        This function ensures each client gets at least min_samples_size samples.

        Parameters
        ----------
        dataset : Dataset
            The dataset to partition. Must have a 'targets' attribute.
        alpha : float, default=0.5
            Concentration parameter for the Dirichlet distribution.
        min_samples_size : int, default=50
            Minimum number of samples required in each partition.
        balanced : bool, default=False
            If True, distribute each class evenly among clients.
            Otherwise, allocate according to a Dirichlet distribution.
        max_iter : int, default=100
            Maximum number of iterations to try finding a valid partition.
        verbose : bool, default=True
            If True, print debug information per iteration.

        Returns
        -------
        partitions : dict[int, list[int]]
            Dictionary mapping each client index to a list of sample indices.
        """
        # Extract targets and unique labels.
        y_data = self._get_targets(dataset)
        unique_labels = np.unique(y_data)

        # For each class, get a shuffled list of indices.
        class_indices = {}
        base_rng = np.random.default_rng(self.seed)
        for label in unique_labels:
            idx = np.where(y_data == label)[0]
            base_rng.shuffle(idx)
            class_indices[label] = idx

        # Prepare container for client indices.
        indices_per_partition = [[] for _ in range(self.partitions_number)]

        def allocate_for_label(label_idx: np.ndarray, rng: np.random.Generator) -> np.ndarray:
            num_label_samples = len(label_idx)
            if balanced:
                proportions = np.full(self.partitions_number, 1.0 / self.partitions_number)
            else:
                proportions = rng.dirichlet([alpha] * self.partitions_number)
            sample_counts = (proportions * num_label_samples).astype(int)
            remainder = num_label_samples - sample_counts.sum()
            if remainder > 0:
                extra_indices = rng.choice(self.partitions_number, size=remainder, replace=False)
                for idx in extra_indices:
                    sample_counts[idx] += 1
            return sample_counts

        for iteration in range(1, max_iter + 1):
            rng = np.random.default_rng(self.seed + iteration)
            temp_indices_per_partition = [[] for _ in range(self.partitions_number)]
            for label in unique_labels:
                label_idx = class_indices[label]
                counts = allocate_for_label(label_idx, rng)
                start = 0
                for client_idx, count in enumerate(counts):
                    end = start + count
                    temp_indices_per_partition[client_idx].extend(label_idx[start:end])
                    start = end

            client_sizes = [len(indices) for indices in temp_indices_per_partition]
            if min(client_sizes) >= min_samples_size:
                indices_per_partition = temp_indices_per_partition
                if verbose:
                    print(f"Partition successful at iteration {iteration}. Client sizes: {client_sizes}")
                break
            if verbose:
                print(f"Iteration {iteration}: client sizes {client_sizes}")

        else:
            raise ValueError(
                f"Could not create partitions with at least {min_samples_size} samples per client after {max_iter} iterations."
            )

        initial_partition = {i: indices for i, indices in enumerate(indices_per_partition)}

        final_partition = self.postprocess_partition(initial_partition, y_data)

        return final_partition

    @staticmethod
    def _get_targets(dataset) -> np.ndarray:
        if isinstance(dataset.targets, np.ndarray):
            return dataset.targets
        elif hasattr(dataset.targets, "numpy"):
            return dataset.targets.numpy()
        else:
            return np.asarray(dataset.targets)

    def postprocess_partition(
        self, partition: dict[int, list[int]], y_data: np.ndarray, min_samples_per_class: int = 10
    ) -> dict[int, list[int]]:
        """
        Post-process a partition to remove (and reassign) classes with too few samples per client.

        For each class:
        - For clients with a count > 0 but below min_samples_per_class, remove those samples.
        - Reassign the removed samples to the client that already has the maximum count for that class.

        Parameters
        ----------
        partition : dict[int, list[int]]
            The initial partition mapping client indices to sample indices.
        y_data : np.ndarray
            The array of labels corresponding to the dataset samples.
        min_samples_per_class : int, default=10
            The minimum acceptable number of samples per class for each client.

        Returns
        -------
        new_partition : dict[int, list[int]]
            The updated partition.
        """
        # Copy partition so we can modify it.
        new_partition = {client: list(indices) for client, indices in partition.items()}

        # Iterate over each class.
        for label in np.unique(y_data):
            # For each client, count how many samples of this label exist.
            client_counts = {}
            for client, indices in new_partition.items():
                client_counts[client] = np.sum(np.array(y_data)[indices] == label)

            # Identify clients with fewer than min_samples_per_class but nonzero counts.
            donors = [client for client, count in client_counts.items() if 0 < count < min_samples_per_class]
            # Identify potential recipients: those with at least min_samples_per_class.
            recipients = [client for client, count in client_counts.items() if count >= min_samples_per_class]
            # If no client meets the threshold, choose the one with the highest count.
            if not recipients:
                best_recipient = max(client_counts, key=client_counts.get)
                recipients = [best_recipient]
            # Choose the recipient with the maximum count.
            best_recipient = max(recipients, key=lambda c: client_counts[c])

            # For each donor, remove samples of this label and reassign them.
            for donor in donors:
                donor_indices = new_partition[donor]
                # Identify indices corresponding to this label.
                donor_label_indices = [idx for idx in donor_indices if y_data[idx] == label]
                # Remove these from the donor.
                new_partition[donor] = [idx for idx in donor_indices if y_data[idx] != label]
                # Add these to the best recipient.
                new_partition[best_recipient].extend(donor_label_indices)

        return new_partition

    def homo_partition(self, dataset):
        """
        Homogeneously partition the dataset into multiple subsets.

        This function divides a dataset into a specified number of subsets, where each subset
        is intended to have a roughly equal number of samples. This method aims to ensure a
        homogeneous distribution of data across all subsets. It's particularly useful in
        scenarios where a uniform distribution of data is desired among all federated learning
        clients.

        Args:
            dataset (torch.utils.data.Dataset): The dataset to partition. It should have
                                                'data' and 'targets' attributes.

        Returns:
            dict: A dictionary where keys are subset indices (ranging from 0 to partitions_number-1)
                and values are lists of indices corresponding to the samples in each subset.

        The function randomly shuffles the entire dataset and then splits it into the number
        of subsets specified by `partitions_number`. It ensures that each subset has a similar number
        of samples. The function also prints the class distribution in each subset for reference.

        Example usage:
            federated_data = homo_partition(my_dataset)
            # This creates federated data subsets with homogeneous distribution.
        """
        n_nets = self.partitions_number

        n_train = len(dataset.targets)
        np.random.seed(self.seed)
        idxs = np.random.permutation(n_train)
        batch_idxs = np.array_split(idxs, n_nets)
        net_dataidx_map = {i: batch_idxs[i] for i in range(n_nets)}

        # partitioned_datasets = []
        for i in range(self.partitions_number):
            # subset = torch.utils.data.Subset(dataset, net_dataidx_map[i])
            # partitioned_datasets.append(subset)

            # Print class distribution in the current partition
            class_counts = [0] * self.num_classes
            for idx in net_dataidx_map[i]:
                label = dataset.targets[idx]
                class_counts[label] += 1
            logging.info(f"Partition {i + 1} class distribution: {class_counts}")

        return net_dataidx_map

    def balanced_iid_partition(self, dataset):
        """
        Partition the dataset into balanced and IID (Independent and Identically Distributed)
        subsets for each client.

        This function divides a dataset into a specified number of subsets (federated clients),
        where each subset has an equal class distribution. This makes the partition suitable for
        simulating IID data scenarios in federated learning.

        Args:
            dataset (list): The dataset to partition. It should be a list of tuples where each
                            tuple represents a data sample and its corresponding label.

        Returns:
            dict: A dictionary where keys are client IDs (ranging from 0 to partitions_number-1) and
                    values are lists of indices corresponding to the samples assigned to each client.

        The function ensures that each class is represented equally in each subset. The
        partitioning process involves iterating over each class, shuffling the indices of that class,
        and then splitting them equally among the clients. The function does not print the class
        distribution in each subset.

        Example usage:
            federated_data = balanced_iid_partition(my_dataset)
            # This creates federated data subsets with equal class distributions.
        """
        num_clients = self.partitions_number
        clients_data = {i: [] for i in range(num_clients)}

        # Get the labels from the dataset
        if isinstance(dataset.targets, np.ndarray):
            labels = dataset.targets
        elif hasattr(dataset.targets, "numpy"):  # Check if it's a tensor with .numpy() method
            labels = dataset.targets.numpy()
        else:  # If it's a list
            labels = np.asarray(dataset.targets)

        label_counts = np.bincount(labels)
        min_label = label_counts.argmin()
        min_count = label_counts[min_label]

        for label in range(self.num_classes):
            # Get the indices of the same label samples
            label_indices = np.where(labels == label)[0]
            np.random.seed(self.seed)
            np.random.shuffle(label_indices)

            # Split the data based on their labels
            samples_per_client = min_count // num_clients

            for i in range(num_clients):
                start_idx = i * samples_per_client
                end_idx = (i + 1) * samples_per_client
                clients_data[i].extend(label_indices[start_idx:end_idx])

        return clients_data

    def unbalanced_iid_partition(self, dataset, imbalance_factor=2):
        """
        Partition the dataset into multiple IID (Independent and Identically Distributed)
        subsets with different size.

        This function divides a dataset into a specified number of IID subsets (federated
        clients), where each subset has a different number of samples. The number of samples
        in each subset is determined by an imbalance factor, making the partition suitable
        for simulating imbalanced data scenarios in federated learning.

        Args:
            dataset (list): The dataset to partition. It should be a list of tuples where
                            each tuple represents a data sample and its corresponding label.
            imbalance_factor (float): The factor to determine the degree of imbalance
                                    among the subsets. A lower imbalance factor leads to more
                                    imbalanced partitions.

        Returns:
            dict: A dictionary where keys are client IDs (ranging from 0 to partitions_number-1) and
                    values are lists of indices corresponding to the samples assigned to each client.

        The function ensures that each class is represented in each subset but with varying
        proportions. The partitioning process involves iterating over each class, shuffling
        the indices of that class, and then splitting them according to the calculated subset
        sizes. The function does not print the class distribution in each subset.

        Example usage:
            federated_data = unbalanced_iid_partition(my_dataset, imbalance_factor=2)
            # This creates federated data subsets with varying number of samples based on
            # an imbalance factor of 2.
        """
        num_clients = self.partitions_number
        clients_data = {i: [] for i in range(num_clients)}

        # Get the labels from the dataset
        labels = np.array([dataset.targets[idx] for idx in range(len(dataset))])
        label_counts = np.bincount(labels)

        min_label = label_counts.argmin()
        min_count = label_counts[min_label]

        # Set the initial_subset_size
        initial_subset_size = min_count // num_clients

        # Calculate the number of samples for each subset based on the imbalance factor
        subset_sizes = [initial_subset_size]
        for i in range(1, num_clients):
            subset_sizes.append(int(subset_sizes[i - 1] * ((imbalance_factor - 1) / imbalance_factor)))

        for label in range(self.num_classes):
            # Get the indices of the same label samples
            label_indices = np.where(labels == label)[0]
            np.random.seed(self.seed)
            np.random.shuffle(label_indices)

            # Split the data based on their labels
            start = 0
            for i in range(num_clients):
                end = start + subset_sizes[i]
                clients_data[i].extend(label_indices[start:end])
                start = end

        return clients_data

    def percentage_partition(self, dataset, percentage=20):
        """
        Partition a dataset into multiple subsets with a specified level of non-IID-ness.

        Args:
            dataset (torch.utils.data.Dataset): The dataset to partition. It should have
                                                'data' and 'targets' attributes.
            percentage (int): A value between 0 and 100 that specifies the desired
                                level of non-IID-ness for the labels of the federated data.
                                This percentage controls the imbalance in the class distribution
                                across different subsets.

        Returns:
            dict: A dictionary where keys are subset indices (ranging from 0 to partitions_number-1)
                and values are lists of indices corresponding to the samples in each subset.

        Example usage:
            federated_data = percentage_partition(my_dataset, percentage=20)
            # This creates federated data subsets with varying class distributions based on
            # a percentage of 20.
        """
        if isinstance(dataset.targets, np.ndarray):
            y_train = dataset.targets
        elif hasattr(dataset.targets, "numpy"):  # Check if it's a tensor with .numpy() method
            y_train = dataset.targets.numpy()
        else:  # If it's a list
            y_train = np.asarray(dataset.targets)

        num_classes = self.num_classes
        num_subsets = self.partitions_number
        class_indices = {i: np.where(y_train == i)[0] for i in range(num_classes)}

        # Get the labels from the dataset
        labels = np.array([dataset.targets[idx] for idx in range(len(dataset))])
        label_counts = np.bincount(labels)

        min_label = label_counts.argmin()
        min_count = label_counts[min_label]

        classes_per_subset = int(num_classes * percentage / 100)
        if classes_per_subset < 1:
            raise ValueError("The percentage is too low to assign at least one class to each subset.")

        subset_indices = [[] for _ in range(num_subsets)]
        class_list = list(range(num_classes))
        np.random.seed(self.seed)
        np.random.shuffle(class_list)

        for i in range(num_subsets):
            for j in range(classes_per_subset):
                # Use modulo operation to cycle through the class_list
                class_idx = class_list[(i * classes_per_subset + j) % num_classes]
                indices = class_indices[class_idx]
                np.random.seed(self.seed)
                np.random.shuffle(indices)
                # Select approximately 50% of the indices
                subset_indices[i].extend(indices[: min_count // 2])

            class_counts = np.bincount(np.array([dataset.targets[idx] for idx in subset_indices[i]]))
            logging.info(f"Partition {i + 1} class distribution: {class_counts.tolist()}")

        partitioned_datasets = {i: subset_indices[i] for i in range(num_subsets)}

        return partitioned_datasets

    def plot_all_data_distribution(self, phase, dataset, partitions_map):
        """

        Plot all of the data distribution of the dataset according to the partitions map provided.

        Args:
            dataset: The dataset to plot (torch.utils.data.Dataset).
            partitions_map: The map of the dataset partitions.
        """
        sns.set()
        sns.set_style("whitegrid", {"axes.grid": False})
        sns.set_context("paper", font_scale=1.5)
        sns.set_palette("Set2")

        num_clients = len(partitions_map)
        num_classes = self.num_classes

        # Plot number of samples per class in the dataset
        plt.figure(figsize=(12, 8))

        class_counts = [0] * num_classes
        for target in dataset.targets:
            class_counts[target] += 1

        plt.bar(range(num_classes), class_counts, tick_label=dataset.classes)
        for j, count in enumerate(class_counts):
            plt.text(j, count, str(count), ha="center", va="bottom", fontsize=12)
        plt.title(f"[{phase}] Number of samples per class in the dataset")
        plt.xlabel("Class")
        plt.ylabel("Number of samples")
        plt.tight_layout()

        path_to_save_class_distribution = f"{self.config_dir}/full_data_distribution_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
        plt.savefig(path_to_save_class_distribution, dpi=300, bbox_inches="tight")
        plt.close()

        # Plot distribution of samples across participants
        plt.figure(figsize=(12, 8))

        label_distribution = [[] for _ in range(num_classes)]
        for c_id, idc in partitions_map.items():
            for idx in idc:
                label_distribution[dataset.targets[idx]].append(c_id)

        plt.hist(
            label_distribution,
            stacked=True,
            bins=np.arange(-0.5, num_clients + 1.5, 1),
            label=dataset.classes,
            rwidth=0.5,
        )
        plt.xticks(
            np.arange(num_clients),
            ["Participant %d" % (c_id + 1) for c_id in range(num_clients)],
        )
        plt.title(f"[{phase}] Distribution of splited datasets")
        plt.xlabel("Participant")
        plt.ylabel("Number of samples")
        plt.xticks(range(num_clients), [f" {i}" for i in range(num_clients)])
        plt.legend(loc="upper right")
        plt.tight_layout()

        path_to_save = f"{self.config_dir}/all_data_distribution_HIST_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
        plt.savefig(path_to_save, dpi=300, bbox_inches="tight")
        plt.close()

        plt.figure(figsize=(12, 8))
        max_point_size = 1200
        min_point_size = 0

        for i in range(self.partitions_number):
            class_counts = [0] * self.num_classes
            indices = partitions_map[i]
            for idx in indices:
                label = dataset.targets[idx]
                class_counts[label] += 1

            # Normalize the point sizes for this partition, handling the case where max_samples_partition is 0
            max_samples_partition = max(class_counts)
            if max_samples_partition == 0:
                logging.warning(f"[{phase}] Participant {i} has no samples. Skipping size normalization.")
                sizes = [min_point_size] * self.num_classes
            else:
                sizes = [
                    (size / max_samples_partition) * (max_point_size - min_point_size) + min_point_size
                    for size in class_counts
                ]
            plt.scatter([i] * self.num_classes, range(self.num_classes), s=sizes, alpha=0.5)

        plt.xlabel("Participant")
        plt.ylabel("Class")
        plt.xticks(range(self.partitions_number))
        plt.yticks(range(self.num_classes))
        plt.title(
            f"[{phase}] Data distribution across participants ({'IID' if self.iid else f'Non-IID - {self.partition}'} - {self.partition_parameter})"
        )
        plt.tight_layout()

        path_to_save = f"{self.config_dir}/all_data_distribution_CIRCLES_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
        plt.savefig(path_to_save, dpi=300, bbox_inches="tight")
        plt.close()

balanced_iid_partition(dataset)

Partition the dataset into balanced and IID (Independent and Identically Distributed) subsets for each client.

This function divides a dataset into a specified number of subsets (federated clients), where each subset has an equal class distribution. This makes the partition suitable for simulating IID data scenarios in federated learning.

Parameters:

Name	Type	Description	Default
dataset	list	The dataset to partition. It should be a list of tuples where each tuple represents a data sample and its corresponding label.	required

Returns:

Name	Type	Description
dict		A dictionary where keys are client IDs (ranging from 0 to partitions_number-1) and values are lists of indices corresponding to the samples assigned to each client.

The function ensures that each class is represented equally in each subset. The partitioning process involves iterating over each class, shuffling the indices of that class, and then splitting them equally among the clients. The function does not print the class distribution in each subset.

Example usage

federated_data = balanced_iid_partition(my_dataset)

This creates federated data subsets with equal class distributions.

Source code in nebula/core/datasets/nebuladataset.py

def balanced_iid_partition(self, dataset):
    """
    Partition the dataset into balanced and IID (Independent and Identically Distributed)
    subsets for each client.

    This function divides a dataset into a specified number of subsets (federated clients),
    where each subset has an equal class distribution. This makes the partition suitable for
    simulating IID data scenarios in federated learning.

    Args:
        dataset (list): The dataset to partition. It should be a list of tuples where each
                        tuple represents a data sample and its corresponding label.

    Returns:
        dict: A dictionary where keys are client IDs (ranging from 0 to partitions_number-1) and
                values are lists of indices corresponding to the samples assigned to each client.

    The function ensures that each class is represented equally in each subset. The
    partitioning process involves iterating over each class, shuffling the indices of that class,
    and then splitting them equally among the clients. The function does not print the class
    distribution in each subset.

    Example usage:
        federated_data = balanced_iid_partition(my_dataset)
        # This creates federated data subsets with equal class distributions.
    """
    num_clients = self.partitions_number
    clients_data = {i: [] for i in range(num_clients)}

    # Get the labels from the dataset
    if isinstance(dataset.targets, np.ndarray):
        labels = dataset.targets
    elif hasattr(dataset.targets, "numpy"):  # Check if it's a tensor with .numpy() method
        labels = dataset.targets.numpy()
    else:  # If it's a list
        labels = np.asarray(dataset.targets)

    label_counts = np.bincount(labels)
    min_label = label_counts.argmin()
    min_count = label_counts[min_label]

    for label in range(self.num_classes):
        # Get the indices of the same label samples
        label_indices = np.where(labels == label)[0]
        np.random.seed(self.seed)
        np.random.shuffle(label_indices)

        # Split the data based on their labels
        samples_per_client = min_count // num_clients

        for i in range(num_clients):
            start_idx = i * samples_per_client
            end_idx = (i + 1) * samples_per_client
            clients_data[i].extend(label_indices[start_idx:end_idx])

    return clients_data

clear()

Clear the dataset. This should remove or reset the dataset.

Source code in nebula/core/datasets/nebuladataset.py

def clear(self):
    """
    Clear the dataset. This should remove or reset the dataset.
    """
    if self.train_set is not None and hasattr(self.train_set, "close"):
        self.train_set.close()
    if self.test_set is not None and hasattr(self.test_set, "close"):
        self.test_set.close()

    self.train_set = None
    self.train_indices_map = None
    self.test_set = None
    self.test_indices_map = None
    self.local_test_indices_map = None

data_partitioning(plot=False)

Perform the data partitioning.

Source code in nebula/core/datasets/nebuladataset.py

def data_partitioning(self, plot=False):
    """
    Perform the data partitioning.
    """

    logging.info(
        f"Partitioning data for {self.__class__.__name__} | Partitions: {self.partitions_number} | IID: {self.iid} | Partition: {self.partition} | Partition parameter: {self.partition_parameter}"
    )

    self.train_indices_map = (
        self.generate_iid_map(self.train_set)
        if self.iid
        else self.generate_non_iid_map(self.train_set, self.partition, self.partition_parameter)
    )
    self.test_indices_map = self.get_test_indices_map()
    self.local_test_indices_map = self.get_local_test_indices_map()

    if plot:
        self.plot_data_distribution("train", self.train_set, self.train_indices_map)
        self.plot_all_data_distribution("train", self.train_set, self.train_indices_map)
        self.plot_data_distribution("local_test", self.test_set, self.local_test_indices_map)
        self.plot_all_data_distribution("local_test", self.test_set, self.local_test_indices_map)

    self.save_partitions()

dirichlet_partition(dataset, alpha=0.5, min_samples_size=50, balanced=False, max_iter=100, verbose=True)

Partition the dataset among clients using a Dirichlet distribution. This function ensures each client gets at least min_samples_size samples.

Parameters

dataset : Dataset The dataset to partition. Must have a 'targets' attribute. alpha : float, default=0.5 Concentration parameter for the Dirichlet distribution. min_samples_size : int, default=50 Minimum number of samples required in each partition. balanced : bool, default=False If True, distribute each class evenly among clients. Otherwise, allocate according to a Dirichlet distribution. max_iter : int, default=100 Maximum number of iterations to try finding a valid partition. verbose : bool, default=True If True, print debug information per iteration.

Returns

partitions : dict[int, list[int]] Dictionary mapping each client index to a list of sample indices.

Source code in nebula/core/datasets/nebuladataset.py

def dirichlet_partition(
    self,
    dataset: Any,
    alpha: float = 0.5,
    min_samples_size: int = 50,
    balanced: bool = False,
    max_iter: int = 100,
    verbose: bool = True,
) -> dict[int, list[int]]:
    """
    Partition the dataset among clients using a Dirichlet distribution.
    This function ensures each client gets at least min_samples_size samples.

    Parameters
    ----------
    dataset : Dataset
        The dataset to partition. Must have a 'targets' attribute.
    alpha : float, default=0.5
        Concentration parameter for the Dirichlet distribution.
    min_samples_size : int, default=50
        Minimum number of samples required in each partition.
    balanced : bool, default=False
        If True, distribute each class evenly among clients.
        Otherwise, allocate according to a Dirichlet distribution.
    max_iter : int, default=100
        Maximum number of iterations to try finding a valid partition.
    verbose : bool, default=True
        If True, print debug information per iteration.

    Returns
    -------
    partitions : dict[int, list[int]]
        Dictionary mapping each client index to a list of sample indices.
    """
    # Extract targets and unique labels.
    y_data = self._get_targets(dataset)
    unique_labels = np.unique(y_data)

    # For each class, get a shuffled list of indices.
    class_indices = {}
    base_rng = np.random.default_rng(self.seed)
    for label in unique_labels:
        idx = np.where(y_data == label)[0]
        base_rng.shuffle(idx)
        class_indices[label] = idx

    # Prepare container for client indices.
    indices_per_partition = [[] for _ in range(self.partitions_number)]

    def allocate_for_label(label_idx: np.ndarray, rng: np.random.Generator) -> np.ndarray:
        num_label_samples = len(label_idx)
        if balanced:
            proportions = np.full(self.partitions_number, 1.0 / self.partitions_number)
        else:
            proportions = rng.dirichlet([alpha] * self.partitions_number)
        sample_counts = (proportions * num_label_samples).astype(int)
        remainder = num_label_samples - sample_counts.sum()
        if remainder > 0:
            extra_indices = rng.choice(self.partitions_number, size=remainder, replace=False)
            for idx in extra_indices:
                sample_counts[idx] += 1
        return sample_counts

    for iteration in range(1, max_iter + 1):
        rng = np.random.default_rng(self.seed + iteration)
        temp_indices_per_partition = [[] for _ in range(self.partitions_number)]
        for label in unique_labels:
            label_idx = class_indices[label]
            counts = allocate_for_label(label_idx, rng)
            start = 0
            for client_idx, count in enumerate(counts):
                end = start + count
                temp_indices_per_partition[client_idx].extend(label_idx[start:end])
                start = end

        client_sizes = [len(indices) for indices in temp_indices_per_partition]
        if min(client_sizes) >= min_samples_size:
            indices_per_partition = temp_indices_per_partition
            if verbose:
                print(f"Partition successful at iteration {iteration}. Client sizes: {client_sizes}")
            break
        if verbose:
            print(f"Iteration {iteration}: client sizes {client_sizes}")

    else:
        raise ValueError(
            f"Could not create partitions with at least {min_samples_size} samples per client after {max_iter} iterations."
        )

    initial_partition = {i: indices for i, indices in enumerate(indices_per_partition)}

    final_partition = self.postprocess_partition(initial_partition, y_data)

    return final_partition

generate_iid_map(dataset, plot=False) abstractmethod

Create an iid map of the dataset.

Source code in nebula/core/datasets/nebuladataset.py

@abstractmethod
def generate_iid_map(self, dataset, plot=False):
    """
    Create an iid map of the dataset.
    """
    pass

generate_non_iid_map(dataset, partition='dirichlet', plot=False) abstractmethod

Create a non-iid map of the dataset.

Source code in nebula/core/datasets/nebuladataset.py

@abstractmethod
def generate_non_iid_map(self, dataset, partition="dirichlet", plot=False):
    """
    Create a non-iid map of the dataset.
    """
    pass

get_local_test_indices_map()

Get the indices of the local test set for each participant. Indices whose labels are the same as the training set are selected.

Returns:

Type	Description
	A dictionary mapping participant_id to a list of indices.

Source code in nebula/core/datasets/nebuladataset.py

def get_local_test_indices_map(self):
    """
    Get the indices of the local test set for each participant.
    Indices whose labels are the same as the training set are selected.

    Returns:
        A dictionary mapping participant_id to a list of indices.
    """
    try:
        local_test_indices_map = {}
        test_targets = np.array(self.test_set.targets)
        for participant_id in range(self.partitions_number):
            train_labels = np.array([self.train_set.targets[idx] for idx in self.train_indices_map[participant_id]])
            indices = np.where(np.isin(test_targets, train_labels))[0].tolist()
            local_test_indices_map[participant_id] = indices
        return local_test_indices_map
    except Exception as e:
        logging.exception(f"Error in get_local_test_indices_map: {e}")
        raise

get_test_indices_map()

Get the indices of the test set for each participant.

Returns:

Type	Description
	A dictionary mapping participant_id to a list of indices.

Source code in nebula/core/datasets/nebuladataset.py

def get_test_indices_map(self):
    """
    Get the indices of the test set for each participant.

    Returns:
        A dictionary mapping participant_id to a list of indices.
    """
    try:
        test_indices_map = {}
        for participant_id in range(self.partitions_number):
            test_indices_map[participant_id] = list(range(len(self.test_set)))
        return test_indices_map
    except Exception as e:
        logging.exception(f"Error in get_test_indices_map: {e}")

homo_partition(dataset)

Homogeneously partition the dataset into multiple subsets.

This function divides a dataset into a specified number of subsets, where each subset is intended to have a roughly equal number of samples. This method aims to ensure a homogeneous distribution of data across all subsets. It's particularly useful in scenarios where a uniform distribution of data is desired among all federated learning clients.

Parameters:

Name	Type	Description	Default
dataset	Dataset	The dataset to partition. It should have 'data' and 'targets' attributes.	required

Returns:

Name	Type	Description
dict		A dictionary where keys are subset indices (ranging from 0 to partitions_number-1) and values are lists of indices corresponding to the samples in each subset.

The function randomly shuffles the entire dataset and then splits it into the number of subsets specified by partitions_number. It ensures that each subset has a similar number of samples. The function also prints the class distribution in each subset for reference.

Example usage

federated_data = homo_partition(my_dataset)

This creates federated data subsets with homogeneous distribution.

Source code in nebula/core/datasets/nebuladataset.py

def homo_partition(self, dataset):
    """
    Homogeneously partition the dataset into multiple subsets.

    This function divides a dataset into a specified number of subsets, where each subset
    is intended to have a roughly equal number of samples. This method aims to ensure a
    homogeneous distribution of data across all subsets. It's particularly useful in
    scenarios where a uniform distribution of data is desired among all federated learning
    clients.

    Args:
        dataset (torch.utils.data.Dataset): The dataset to partition. It should have
                                            'data' and 'targets' attributes.

    Returns:
        dict: A dictionary where keys are subset indices (ranging from 0 to partitions_number-1)
            and values are lists of indices corresponding to the samples in each subset.

    The function randomly shuffles the entire dataset and then splits it into the number
    of subsets specified by `partitions_number`. It ensures that each subset has a similar number
    of samples. The function also prints the class distribution in each subset for reference.

    Example usage:
        federated_data = homo_partition(my_dataset)
        # This creates federated data subsets with homogeneous distribution.
    """
    n_nets = self.partitions_number

    n_train = len(dataset.targets)
    np.random.seed(self.seed)
    idxs = np.random.permutation(n_train)
    batch_idxs = np.array_split(idxs, n_nets)
    net_dataidx_map = {i: batch_idxs[i] for i in range(n_nets)}

    # partitioned_datasets = []
    for i in range(self.partitions_number):
        # subset = torch.utils.data.Subset(dataset, net_dataidx_map[i])
        # partitioned_datasets.append(subset)

        # Print class distribution in the current partition
        class_counts = [0] * self.num_classes
        for idx in net_dataidx_map[i]:
            label = dataset.targets[idx]
            class_counts[label] += 1
        logging.info(f"Partition {i + 1} class distribution: {class_counts}")

    return net_dataidx_map

initialize_dataset() abstractmethod

Initialize the dataset. This should load or create the dataset.

Source code in nebula/core/datasets/nebuladataset.py

@abstractmethod
def initialize_dataset(self):
    """
    Initialize the dataset. This should load or create the dataset.
    """
    raise NotImplementedError("Subclasses must implement this method.")

percentage_partition(dataset, percentage=20)

Partition a dataset into multiple subsets with a specified level of non-IID-ness.

Parameters:

Name	Type	Description	Default
dataset	Dataset	The dataset to partition. It should have 'data' and 'targets' attributes.	required
percentage	int	A value between 0 and 100 that specifies the desired level of non-IID-ness for the labels of the federated data. This percentage controls the imbalance in the class distribution across different subsets.	20

Returns:

Name	Type	Description
dict		A dictionary where keys are subset indices (ranging from 0 to partitions_number-1) and values are lists of indices corresponding to the samples in each subset.

Example usage

federated_data = percentage_partition(my_dataset, percentage=20)

This creates federated data subsets with varying class distributions based on

a percentage of 20.

Source code in nebula/core/datasets/nebuladataset.py

def percentage_partition(self, dataset, percentage=20):
    """
    Partition a dataset into multiple subsets with a specified level of non-IID-ness.

    Args:
        dataset (torch.utils.data.Dataset): The dataset to partition. It should have
                                            'data' and 'targets' attributes.
        percentage (int): A value between 0 and 100 that specifies the desired
                            level of non-IID-ness for the labels of the federated data.
                            This percentage controls the imbalance in the class distribution
                            across different subsets.

    Returns:
        dict: A dictionary where keys are subset indices (ranging from 0 to partitions_number-1)
            and values are lists of indices corresponding to the samples in each subset.

    Example usage:
        federated_data = percentage_partition(my_dataset, percentage=20)
        # This creates federated data subsets with varying class distributions based on
        # a percentage of 20.
    """
    if isinstance(dataset.targets, np.ndarray):
        y_train = dataset.targets
    elif hasattr(dataset.targets, "numpy"):  # Check if it's a tensor with .numpy() method
        y_train = dataset.targets.numpy()
    else:  # If it's a list
        y_train = np.asarray(dataset.targets)

    num_classes = self.num_classes
    num_subsets = self.partitions_number
    class_indices = {i: np.where(y_train == i)[0] for i in range(num_classes)}

    # Get the labels from the dataset
    labels = np.array([dataset.targets[idx] for idx in range(len(dataset))])
    label_counts = np.bincount(labels)

    min_label = label_counts.argmin()
    min_count = label_counts[min_label]

    classes_per_subset = int(num_classes * percentage / 100)
    if classes_per_subset < 1:
        raise ValueError("The percentage is too low to assign at least one class to each subset.")

    subset_indices = [[] for _ in range(num_subsets)]
    class_list = list(range(num_classes))
    np.random.seed(self.seed)
    np.random.shuffle(class_list)

    for i in range(num_subsets):
        for j in range(classes_per_subset):
            # Use modulo operation to cycle through the class_list
            class_idx = class_list[(i * classes_per_subset + j) % num_classes]
            indices = class_indices[class_idx]
            np.random.seed(self.seed)
            np.random.shuffle(indices)
            # Select approximately 50% of the indices
            subset_indices[i].extend(indices[: min_count // 2])

        class_counts = np.bincount(np.array([dataset.targets[idx] for idx in subset_indices[i]]))
        logging.info(f"Partition {i + 1} class distribution: {class_counts.tolist()}")

    partitioned_datasets = {i: subset_indices[i] for i in range(num_subsets)}

    return partitioned_datasets

plot_all_data_distribution(phase, dataset, partitions_map)

Plot all of the data distribution of the dataset according to the partitions map provided.

Parameters:

Name	Type	Description	Default
dataset		The dataset to plot (torch.utils.data.Dataset).	required
partitions_map		The map of the dataset partitions.	required

Source code in nebula/core/datasets/nebuladataset.py

def plot_all_data_distribution(self, phase, dataset, partitions_map):
    """

    Plot all of the data distribution of the dataset according to the partitions map provided.

    Args:
        dataset: The dataset to plot (torch.utils.data.Dataset).
        partitions_map: The map of the dataset partitions.
    """
    sns.set()
    sns.set_style("whitegrid", {"axes.grid": False})
    sns.set_context("paper", font_scale=1.5)
    sns.set_palette("Set2")

    num_clients = len(partitions_map)
    num_classes = self.num_classes

    # Plot number of samples per class in the dataset
    plt.figure(figsize=(12, 8))

    class_counts = [0] * num_classes
    for target in dataset.targets:
        class_counts[target] += 1

    plt.bar(range(num_classes), class_counts, tick_label=dataset.classes)
    for j, count in enumerate(class_counts):
        plt.text(j, count, str(count), ha="center", va="bottom", fontsize=12)
    plt.title(f"[{phase}] Number of samples per class in the dataset")
    plt.xlabel("Class")
    plt.ylabel("Number of samples")
    plt.tight_layout()

    path_to_save_class_distribution = f"{self.config_dir}/full_data_distribution_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
    plt.savefig(path_to_save_class_distribution, dpi=300, bbox_inches="tight")
    plt.close()

    # Plot distribution of samples across participants
    plt.figure(figsize=(12, 8))

    label_distribution = [[] for _ in range(num_classes)]
    for c_id, idc in partitions_map.items():
        for idx in idc:
            label_distribution[dataset.targets[idx]].append(c_id)

    plt.hist(
        label_distribution,
        stacked=True,
        bins=np.arange(-0.5, num_clients + 1.5, 1),
        label=dataset.classes,
        rwidth=0.5,
    )
    plt.xticks(
        np.arange(num_clients),
        ["Participant %d" % (c_id + 1) for c_id in range(num_clients)],
    )
    plt.title(f"[{phase}] Distribution of splited datasets")
    plt.xlabel("Participant")
    plt.ylabel("Number of samples")
    plt.xticks(range(num_clients), [f" {i}" for i in range(num_clients)])
    plt.legend(loc="upper right")
    plt.tight_layout()

    path_to_save = f"{self.config_dir}/all_data_distribution_HIST_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
    plt.savefig(path_to_save, dpi=300, bbox_inches="tight")
    plt.close()

    plt.figure(figsize=(12, 8))
    max_point_size = 1200
    min_point_size = 0

    for i in range(self.partitions_number):
        class_counts = [0] * self.num_classes
        indices = partitions_map[i]
        for idx in indices:
            label = dataset.targets[idx]
            class_counts[label] += 1

        # Normalize the point sizes for this partition, handling the case where max_samples_partition is 0
        max_samples_partition = max(class_counts)
        if max_samples_partition == 0:
            logging.warning(f"[{phase}] Participant {i} has no samples. Skipping size normalization.")
            sizes = [min_point_size] * self.num_classes
        else:
            sizes = [
                (size / max_samples_partition) * (max_point_size - min_point_size) + min_point_size
                for size in class_counts
            ]
        plt.scatter([i] * self.num_classes, range(self.num_classes), s=sizes, alpha=0.5)

    plt.xlabel("Participant")
    plt.ylabel("Class")
    plt.xticks(range(self.partitions_number))
    plt.yticks(range(self.num_classes))
    plt.title(
        f"[{phase}] Data distribution across participants ({'IID' if self.iid else f'Non-IID - {self.partition}'} - {self.partition_parameter})"
    )
    plt.tight_layout()

    path_to_save = f"{self.config_dir}/all_data_distribution_CIRCLES_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
    plt.savefig(path_to_save, dpi=300, bbox_inches="tight")
    plt.close()

plot_data_distribution(phase, dataset, partitions_map)

Plot the data distribution of the dataset.

Plot the data distribution of the dataset according to the partitions map provided.

Parameters:

Name	Type	Description	Default
phase		The phase of the dataset (train, test, local_test).	required
dataset		The dataset to plot (torch.utils.data.Dataset).	required
partitions_map		The map of the dataset partitions.	required

Source code in nebula/core/datasets/nebuladataset.py

def plot_data_distribution(self, phase, dataset, partitions_map):
    """
    Plot the data distribution of the dataset.

    Plot the data distribution of the dataset according to the partitions map provided.

    Args:
        phase: The phase of the dataset (train, test, local_test).
        dataset: The dataset to plot (torch.utils.data.Dataset).
        partitions_map: The map of the dataset partitions.
    """
    logging_training.info(f"Plotting data distribution for {phase} dataset")
    # Plot the data distribution of the dataset, one graph per partition
    sns.set()
    sns.set_style("whitegrid", {"axes.grid": False})
    sns.set_context("paper", font_scale=1.5)
    sns.set_palette("Set2")

    # Plot bar charts for each partition
    partition_index = 0
    for partition_index in range(self.partitions_number):
        indices = partitions_map[partition_index]
        class_counts = [0] * self.num_classes
        for idx in indices:
            label = dataset.targets[idx]
            class_counts[label] += 1

        logging_training.info(f"[{phase}] Participant {partition_index} total samples: {len(indices)}")
        logging_training.info(f"[{phase}] Participant {partition_index} data distribution: {class_counts}")

        plt.figure(figsize=(12, 8))
        plt.bar(range(self.num_classes), class_counts)
        for j, count in enumerate(class_counts):
            plt.text(j, count, str(count), ha="center", va="bottom", fontsize=12)
        plt.xlabel("Class")
        plt.ylabel("Number of samples")
        plt.xticks(range(self.num_classes))
        plt.title(
            f"[{phase}] Participant {partition_index} data distribution ({'IID' if self.iid else f'Non-IID - {self.partition}'} - {self.partition_parameter})"
        )
        plt.tight_layout()
        path_to_save = f"{self.config_dir}/participant_{partition_index}_data_distribution_{'iid' if self.iid else 'non_iid'}{'_' + self.partition if not self.iid else ''}_{phase}.pdf"
        logging_training.info(f"Saving data distribution for participant {partition_index} to {path_to_save}")
        plt.savefig(path_to_save, dpi=300, bbox_inches="tight")
        plt.close()

    if hasattr(self, "tsne") and self.tsne:
        self.visualize_tsne(dataset)

postprocess_partition(partition, y_data, min_samples_per_class=10)

Post-process a partition to remove (and reassign) classes with too few samples per client.

For each class: - For clients with a count > 0 but below min_samples_per_class, remove those samples. - Reassign the removed samples to the client that already has the maximum count for that class.

Parameters

partition : dict[int, list[int]] The initial partition mapping client indices to sample indices. y_data : np.ndarray The array of labels corresponding to the dataset samples. min_samples_per_class : int, default=10 The minimum acceptable number of samples per class for each client.

Returns

new_partition : dict[int, list[int]] The updated partition.

Source code in nebula/core/datasets/nebuladataset.py

def postprocess_partition(
    self, partition: dict[int, list[int]], y_data: np.ndarray, min_samples_per_class: int = 10
) -> dict[int, list[int]]:
    """
    Post-process a partition to remove (and reassign) classes with too few samples per client.

    For each class:
    - For clients with a count > 0 but below min_samples_per_class, remove those samples.
    - Reassign the removed samples to the client that already has the maximum count for that class.

    Parameters
    ----------
    partition : dict[int, list[int]]
        The initial partition mapping client indices to sample indices.
    y_data : np.ndarray
        The array of labels corresponding to the dataset samples.
    min_samples_per_class : int, default=10
        The minimum acceptable number of samples per class for each client.

    Returns
    -------
    new_partition : dict[int, list[int]]
        The updated partition.
    """
    # Copy partition so we can modify it.
    new_partition = {client: list(indices) for client, indices in partition.items()}

    # Iterate over each class.
    for label in np.unique(y_data):
        # For each client, count how many samples of this label exist.
        client_counts = {}
        for client, indices in new_partition.items():
            client_counts[client] = np.sum(np.array(y_data)[indices] == label)

        # Identify clients with fewer than min_samples_per_class but nonzero counts.
        donors = [client for client, count in client_counts.items() if 0 < count < min_samples_per_class]
        # Identify potential recipients: those with at least min_samples_per_class.
        recipients = [client for client, count in client_counts.items() if count >= min_samples_per_class]
        # If no client meets the threshold, choose the one with the highest count.
        if not recipients:
            best_recipient = max(client_counts, key=client_counts.get)
            recipients = [best_recipient]
        # Choose the recipient with the maximum count.
        best_recipient = max(recipients, key=lambda c: client_counts[c])

        # For each donor, remove samples of this label and reassign them.
        for donor in donors:
            donor_indices = new_partition[donor]
            # Identify indices corresponding to this label.
            donor_label_indices = [idx for idx in donor_indices if y_data[idx] == label]
            # Remove these from the donor.
            new_partition[donor] = [idx for idx in donor_indices if y_data[idx] != label]
            # Add these to the best recipient.
            new_partition[best_recipient].extend(donor_label_indices)

    return new_partition

save_partitions()

Save each partition data (train, test, and local test) to separate pickle files. The controller saves one file per partition for each data split.

Source code in nebula/core/datasets/nebuladataset.py

def save_partitions(self):
    """
    Save each partition data (train, test, and local test) to separate pickle files.
    The controller saves one file per partition for each data split.
    """
    try:
        logging.info(f"Saving partitions data for ALL participants ({self.partitions_number}) in {self.config_dir}")
        path = self.config_dir
        if not os.path.exists(path):
            raise FileNotFoundError(f"Path {path} does not exist")
        # Check that the partition maps have the expected number of partitions
        if not (
            len(self.train_indices_map)
            == len(self.test_indices_map)
            == len(self.local_test_indices_map)
            == self.partitions_number
        ):
            raise ValueError("One of the partition maps has an unexpected length.")

        # Save global test data
        file_name = os.path.join(path, "global_test.h5")
        with h5py.File(file_name, "w") as f:
            indices = list(range(len(self.test_set)))
            test_data = [self.test_set[i] for i in indices]
            self.save_partition(test_data, f, "test_data")
            f["test_data"].attrs["num_classes"] = self.num_classes
            test_targets = np.array(self.test_set.targets)
            f.create_dataset("test_targets", data=test_targets, compression="gzip")

        for participant in range(self.partitions_number):
            file_name = os.path.join(path, f"participant_{participant}_train.h5")
            with h5py.File(file_name, "w") as f:
                logging.info(f"Saving training data for participant {participant} in {file_name}")
                indices = self.train_indices_map[participant]
                train_data = [self.train_set[i] for i in indices]
                self.save_partition(train_data, f, "train_data")
                f["train_data"].attrs["num_classes"] = self.num_classes
                train_targets = np.array([self.train_set.targets[i] for i in indices])
                f.create_dataset("train_targets", data=train_targets, compression="gzip")
                logging.info(f"Partition saved for participant {participant}.")

        logging.info("Successfully saved all partition files.")

    except Exception as e:
        logging.exception(f"Error in save_partitions: {e}")

    finally:
        self.clear()
        logging.info("Cleared dataset after saving partitions.")

unbalanced_iid_partition(dataset, imbalance_factor=2)

Partition the dataset into multiple IID (Independent and Identically Distributed) subsets with different size.

This function divides a dataset into a specified number of IID subsets (federated clients), where each subset has a different number of samples. The number of samples in each subset is determined by an imbalance factor, making the partition suitable for simulating imbalanced data scenarios in federated learning.

Parameters:

Name	Type	Description	Default
dataset	list	The dataset to partition. It should be a list of tuples where each tuple represents a data sample and its corresponding label.	required
imbalance_factor	float	The factor to determine the degree of imbalance among the subsets. A lower imbalance factor leads to more imbalanced partitions.	2

Returns:

Name	Type	Description
dict		A dictionary where keys are client IDs (ranging from 0 to partitions_number-1) and values are lists of indices corresponding to the samples assigned to each client.

The function ensures that each class is represented in each subset but with varying proportions. The partitioning process involves iterating over each class, shuffling the indices of that class, and then splitting them according to the calculated subset sizes. The function does not print the class distribution in each subset.

Example usage

federated_data = unbalanced_iid_partition(my_dataset, imbalance_factor=2)

This creates federated data subsets with varying number of samples based on

an imbalance factor of 2.

Source code in nebula/core/datasets/nebuladataset.py

def unbalanced_iid_partition(self, dataset, imbalance_factor=2):
    """
    Partition the dataset into multiple IID (Independent and Identically Distributed)
    subsets with different size.

    This function divides a dataset into a specified number of IID subsets (federated
    clients), where each subset has a different number of samples. The number of samples
    in each subset is determined by an imbalance factor, making the partition suitable
    for simulating imbalanced data scenarios in federated learning.

    Args:
        dataset (list): The dataset to partition. It should be a list of tuples where
                        each tuple represents a data sample and its corresponding label.
        imbalance_factor (float): The factor to determine the degree of imbalance
                                among the subsets. A lower imbalance factor leads to more
                                imbalanced partitions.

    Returns:
        dict: A dictionary where keys are client IDs (ranging from 0 to partitions_number-1) and
                values are lists of indices corresponding to the samples assigned to each client.

    The function ensures that each class is represented in each subset but with varying
    proportions. The partitioning process involves iterating over each class, shuffling
    the indices of that class, and then splitting them according to the calculated subset
    sizes. The function does not print the class distribution in each subset.

    Example usage:
        federated_data = unbalanced_iid_partition(my_dataset, imbalance_factor=2)
        # This creates federated data subsets with varying number of samples based on
        # an imbalance factor of 2.
    """
    num_clients = self.partitions_number
    clients_data = {i: [] for i in range(num_clients)}

    # Get the labels from the dataset
    labels = np.array([dataset.targets[idx] for idx in range(len(dataset))])
    label_counts = np.bincount(labels)

    min_label = label_counts.argmin()
    min_count = label_counts[min_label]

    # Set the initial_subset_size
    initial_subset_size = min_count // num_clients

    # Calculate the number of samples for each subset based on the imbalance factor
    subset_sizes = [initial_subset_size]
    for i in range(1, num_clients):
        subset_sizes.append(int(subset_sizes[i - 1] * ((imbalance_factor - 1) / imbalance_factor)))

    for label in range(self.num_classes):
        # Get the indices of the same label samples
        label_indices = np.where(labels == label)[0]
        np.random.seed(self.seed)
        np.random.shuffle(label_indices)

        # Split the data based on their labels
        start = 0
        for i in range(num_clients):
            end = start + subset_sizes[i]
            clients_data[i].extend(label_indices[start:end])
            start = end

    return clients_data

NebulaPartition

A class to handle the partitioning of datasets for federated learning.

Source code in nebula/core/datasets/nebuladataset.py

class NebulaPartition:
    """
    A class to handle the partitioning of datasets for federated learning.
    """

    def __init__(self, handler: NebulaPartitionHandler, config: dict[str, Any] | None = None):
        self.handler = handler
        self.config = config if config is not None else {}

        self.train_set = None
        self.train_indices = None

        self.test_set = None
        self.test_indices = None

        self.local_test_set = None
        self.local_test_indices = None

        enable_deterministic(seed=self.config.participant["scenario_args"]["random_seed"])

    def get_train_indices(self):
        """
        Get the indices of the training set based on the indices map.
        """
        if self.train_indices is None:
            return None
        return self.train_indices

    def get_test_indices(self):
        """
        Get the indices of the test set based on the indices map.
        """
        if self.test_indices is None:
            return None
        return self.test_indices

    def get_local_test_indices(self):
        """
        Get the indices of the local test set based on the indices map.
        """
        if self.local_test_indices is None:
            return None
        return self.local_test_indices

    def get_train_labels(self):
        """
        Get the labels of the training set based on the indices map.
        """
        if self.train_indices is None:
            return None
        return [self.train_set.targets[idx] for idx in self.train_indices]

    def get_test_labels(self):
        """
        Get the labels of the test set based on the indices map.
        """
        if self.test_indices is None:
            return None
        return [self.test_set.targets[idx] for idx in self.test_indices]

    def get_local_test_labels(self):
        """
        Get the labels of the test set based on the indices map.
        """
        if self.local_test_indices is None:
            return None
        return [self.test_set.targets[idx] for idx in self.local_test_indices]

    def set_local_test_indices(self):
        """
        Set the local test indices for the current node.
        """
        test_labels = self.get_test_labels()
        train_labels = self.get_train_labels()

        if test_labels is None or train_labels is None:
            logging_training.warning("Either test_labels or train_labels is None in set_local_test_indices")
            return []

        if self.test_set is None:
            logging_training.warning("test_set is None in set_local_test_indices")
            return []

        return [idx for idx in range(len(self.test_set)) if test_labels[idx] in train_labels]

    def log_partition(self):
        logging_training.info(f"{'=' * 10}")
        logging_training.info(
            f"LOG NEBULA PARTITION DATASET [Participant {self.config.participant['device_args']['idx']}]"
        )
        logging_training.info(f"{'=' * 10}")
        logging_training.info(f"TRAIN - Train labels unique: {set(self.get_train_labels())}")
        logging_training.info(f"TRAIN - Length of train indices map: {len(self.get_train_indices())}")
        logging_training.info(f"{'=' * 10}")
        logging_training.info(f"LOCAL - Test labels unique: {set(self.get_local_test_labels())}")
        logging_training.info(f"LOCAL - Length of test indices map: {len(self.get_local_test_indices())}")
        logging_training.info(f"{'=' * 10}")
        logging_training.info(f"GLOBAL - Test labels unique: {set(self.get_test_labels())}")
        logging_training.info(f"GLOBAL - Length of test indices map: {len(self.get_test_indices())}")
        logging_training.info(f"{'=' * 10}")

    def load_partition(self):
        """
        Load only the partition data corresponding to the current node.
        The node loads its train, test, and local test partition data from HDF5 files.
        """
        try:
            p = self.config.participant["device_args"]["idx"]
            logging_training.info(f"Loading partition data for participant {p}")
            path = self.config.participant["tracking_args"]["config_dir"]

            train_partition_file = os.path.join(path, f"participant_{p}_train.h5")
            wait_for_file(train_partition_file)
            logging_training.info(f"Loading train data from {train_partition_file}")
            self.train_set = self.handler(train_partition_file, "train", config=self.config)
            self.train_indices = list(range(len(self.train_set)))

            test_partition_file = os.path.join(path, "global_test.h5")
            wait_for_file(test_partition_file)
            logging_training.info(f"Loading test data from {test_partition_file}")
            self.test_set = self.handler(test_partition_file, "test", config=self.config)
            self.test_indices = list(range(len(self.test_set)))

            self.local_test_set = self.handler(
                test_partition_file, "local_test", config=self.config, empty=True
            )
            self.local_test_set.set_data(self.test_set.data, self.test_set.targets)
            self.local_test_indices = self.set_local_test_indices()

            logging_training.info(f"Successfully loaded partition data for participant {p}.")
        except Exception as e:
            logging_training.error(f"Error loading partition: {e}")
            raise

get_local_test_indices()

Get the indices of the local test set based on the indices map.

Source code in nebula/core/datasets/nebuladataset.py

def get_local_test_indices(self):
    """
    Get the indices of the local test set based on the indices map.
    """
    if self.local_test_indices is None:
        return None
    return self.local_test_indices

get_local_test_labels()

Get the labels of the test set based on the indices map.

Source code in nebula/core/datasets/nebuladataset.py

def get_local_test_labels(self):
    """
    Get the labels of the test set based on the indices map.
    """
    if self.local_test_indices is None:
        return None
    return [self.test_set.targets[idx] for idx in self.local_test_indices]

get_test_indices()

Get the indices of the test set based on the indices map.

Source code in nebula/core/datasets/nebuladataset.py

def get_test_indices(self):
    """
    Get the indices of the test set based on the indices map.
    """
    if self.test_indices is None:
        return None
    return self.test_indices

get_test_labels()

Get the labels of the test set based on the indices map.

Source code in nebula/core/datasets/nebuladataset.py

def get_test_labels(self):
    """
    Get the labels of the test set based on the indices map.
    """
    if self.test_indices is None:
        return None
    return [self.test_set.targets[idx] for idx in self.test_indices]

get_train_indices()

Get the indices of the training set based on the indices map.

Source code in nebula/core/datasets/nebuladataset.py

def get_train_indices(self):
    """
    Get the indices of the training set based on the indices map.
    """
    if self.train_indices is None:
        return None
    return self.train_indices

get_train_labels()

Get the labels of the training set based on the indices map.

Source code in nebula/core/datasets/nebuladataset.py

def get_train_labels(self):
    """
    Get the labels of the training set based on the indices map.
    """
    if self.train_indices is None:
        return None
    return [self.train_set.targets[idx] for idx in self.train_indices]

load_partition()

Load only the partition data corresponding to the current node. The node loads its train, test, and local test partition data from HDF5 files.

Source code in nebula/core/datasets/nebuladataset.py

def load_partition(self):
    """
    Load only the partition data corresponding to the current node.
    The node loads its train, test, and local test partition data from HDF5 files.
    """
    try:
        p = self.config.participant["device_args"]["idx"]
        logging_training.info(f"Loading partition data for participant {p}")
        path = self.config.participant["tracking_args"]["config_dir"]

        train_partition_file = os.path.join(path, f"participant_{p}_train.h5")
        wait_for_file(train_partition_file)
        logging_training.info(f"Loading train data from {train_partition_file}")
        self.train_set = self.handler(train_partition_file, "train", config=self.config)
        self.train_indices = list(range(len(self.train_set)))

        test_partition_file = os.path.join(path, "global_test.h5")
        wait_for_file(test_partition_file)
        logging_training.info(f"Loading test data from {test_partition_file}")
        self.test_set = self.handler(test_partition_file, "test", config=self.config)
        self.test_indices = list(range(len(self.test_set)))

        self.local_test_set = self.handler(
            test_partition_file, "local_test", config=self.config, empty=True
        )
        self.local_test_set.set_data(self.test_set.data, self.test_set.targets)
        self.local_test_indices = self.set_local_test_indices()

        logging_training.info(f"Successfully loaded partition data for participant {p}.")
    except Exception as e:
        logging_training.error(f"Error loading partition: {e}")
        raise

set_local_test_indices()

Set the local test indices for the current node.

Source code in nebula/core/datasets/nebuladataset.py

def set_local_test_indices(self):
    """
    Set the local test indices for the current node.
    """
    test_labels = self.get_test_labels()
    train_labels = self.get_train_labels()

    if test_labels is None or train_labels is None:
        logging_training.warning("Either test_labels or train_labels is None in set_local_test_indices")
        return []

    if self.test_set is None:
        logging_training.warning("test_set is None in set_local_test_indices")
        return []

    return [idx for idx in range(len(self.test_set)) if test_labels[idx] in train_labels]

NebulaPartitionHandler

Bases: Dataset, ABC

A class to handle the loading of datasets from HDF5 files.

Source code in nebula/core/datasets/nebuladataset.py

class NebulaPartitionHandler(Dataset, ABC):
    """
    A class to handle the loading of datasets from HDF5 files.
    """

    def __init__(
        self,
        file_path: str,
        prefix: str = "train",
        config: dict[str, Any] | None = None,
        empty: bool = False,
    ):
        self.file_path = file_path
        self.prefix = prefix
        self.config = config
        self.empty = empty
        self.transform = None
        self.target_transform = None
        self.file = None

        self.data = None
        self.targets = None
        self.num_classes = None
        self.length = None

        self.load_data()

    def load_data(self):
        if self.empty:
            logging_training.info(
                f"[NebulaPartitionHandler] No data loaded for {self.prefix} partition. Empty dataset."
            )
            return
        with h5py.File(self.file_path, "r") as f:
            prefix = (
                "test" if self.prefix == "local_test" else self.prefix
            )  # Local test uses the test prefix (same data but different split)
            self.data = self.load_partition(f, f"{prefix}_data")
            self.targets = np.array(f[f"{prefix}_targets"])
            self.num_classes = f[f"{prefix}_data"].attrs.get("num_classes", 0)
            self.length = len(self.data)
        logging_training.info(
            f"[NebulaPartitionHandler] [{self.prefix}] Loaded {self.length} samples from {self.file_path} and {self.num_classes} classes."
        )

    def close(self):
        if self.file is not None:
            self.file.close()
            self.file = None
            logging_training.info(f"[NebulaPartitionHandler] Closed file {self.file_path}")

    def __del__(self):
        self.close()

    def __len__(self):
        return self.length

    def __getitem__(self, idx):
        data = self.data[idx]
        target = self.targets[idx]
        return data, target

    def set_data(self, data, targets, data_opt=None, targets_opt=None):
        """
        Set the data and targets for the dataset.
        """
        try:
            # Input validation
            if data is None or targets is None:
                raise ValueError("Primary data and targets cannot be None")

            if len(data) != len(targets):
                raise ValueError(f"Data and targets length mismatch: {len(data)} vs {len(targets)}")

            if data_opt is None or targets_opt is None:
                self.data = data
                self.targets = targets
                self.length = len(data)
                logging_training.info(f"[NebulaPartitionHandler] Set data with {self.length} samples.")
                return

            if len(data_opt) != len(targets_opt):
                raise ValueError(f"Optional data and targets length mismatch: {len(data_opt)} vs {len(targets_opt)}")

            main_count = int(len(data) * 0.8)
            opt_count = min(len(data_opt), int(len(data) * (1 - 0.8)))
            if isinstance(data, np.ndarray):
                self.data = np.concatenate((data[:main_count], data_opt[:opt_count]))
            else:
                self.data = data[:main_count] + data_opt[:opt_count]

            if isinstance(targets, np.ndarray):
                self.targets = np.concatenate((targets[:main_count], targets_opt[:opt_count]))
            else:
                self.targets = targets[:main_count] + targets_opt[:opt_count]
            self.length = len(self.data)

        except Exception as e:
            logging_training.exception(f"Error setting data: {e}")

    def load_partition(self, file, name):
        item = file[name]
        if isinstance(item, h5py.Dataset):
            typ = item.attrs.get("__type__", None)
            if typ == "pickle":
                logging_training.info(f"Loading pickled object from {name}")
                return pickle.loads(item[()].tobytes())
            else:
                logging_training.warning(f"[NebulaPartitionHandler] Unknown type encountered: {typ} for item {name}")
                return item[()]
        else:
            logging_training.warning(f"[NebulaPartitionHandler] Unknown item encountered: {item} for item {name}")
            return item[()]

set_data(data, targets, data_opt=None, targets_opt=None)

Set the data and targets for the dataset.

Source code in nebula/core/datasets/nebuladataset.py

def set_data(self, data, targets, data_opt=None, targets_opt=None):
    """
    Set the data and targets for the dataset.
    """
    try:
        # Input validation
        if data is None or targets is None:
            raise ValueError("Primary data and targets cannot be None")

        if len(data) != len(targets):
            raise ValueError(f"Data and targets length mismatch: {len(data)} vs {len(targets)}")

        if data_opt is None or targets_opt is None:
            self.data = data
            self.targets = targets
            self.length = len(data)
            logging_training.info(f"[NebulaPartitionHandler] Set data with {self.length} samples.")
            return

        if len(data_opt) != len(targets_opt):
            raise ValueError(f"Optional data and targets length mismatch: {len(data_opt)} vs {len(targets_opt)}")

        main_count = int(len(data) * 0.8)
        opt_count = min(len(data_opt), int(len(data) * (1 - 0.8)))
        if isinstance(data, np.ndarray):
            self.data = np.concatenate((data[:main_count], data_opt[:opt_count]))
        else:
            self.data = data[:main_count] + data_opt[:opt_count]

        if isinstance(targets, np.ndarray):
            self.targets = np.concatenate((targets[:main_count], targets_opt[:opt_count]))
        else:
            self.targets = targets[:main_count] + targets_opt[:opt_count]
        self.length = len(self.data)

    except Exception as e:
        logging_training.exception(f"Error setting data: {e}")

wait_for_file(file_path)

Wait until the given file exists, polling every 'interval' seconds.

Source code in nebula/core/datasets/nebuladataset.py

def wait_for_file(file_path):
    """Wait until the given file exists, polling every 'interval' seconds."""
    while not os.path.exists(file_path):
        logging_training.info(f"Waiting for file: {file_path}")
    return