Quickstart XGBoost#

Federated XGBoost#

EXtreme Gradient Boosting (XGBoost) is a robust and efficient implementation of gradient-boosted decision tree (GBDT), that maximises the computational boundaries for boosted tree methods. It’s primarily designed to enhance both the performance and computational speed of machine learning models. In XGBoost, trees are constructed concurrently, unlike the sequential approach taken by GBDT.

Often, for tabular data on medium-sized datasets with fewer than 10k training examples, XGBoost surpasses the results of deep learning techniques.

Why federated XGBoost?#

Indeed, as the demand for data privacy and decentralized learning grows, there’s an increasing requirement to implement federated XGBoost systems for specialised applications, like survival analysis and financial fraud detection.

Federated learning ensures that raw data remains on the local device, making it an attractive approach for sensitive domains where data security and privacy are paramount. Given the robustness and efficiency of XGBoost, combining it with federated learning offers a promising solution for these specific challenges.

In this tutorial we will learn how to train a federated XGBoost model on HIGGS dataset using Flower and xgboost package. We use a simple example (full code xgboost-quickstart) with two clients and one server to demonstrate how federated XGBoost works, and then we dive into a more complex example (full code xgboost-comprehensive) to run various experiments.

Environment Setup#

First of all, it is recommended to create a virtual environment and run everything within a virtualenv.

We first need to install Flower and Flower Datasets. You can do this by running :

$ pip install flwr flwr-datasets

Since we want to use xgboost package to build up XGBoost trees, let’s go ahead and install xgboost:

$ pip install xgboost

Flower Client#

Clients are responsible for generating individual weight-updates for the model based on their local datasets. Now that we have all our dependencies installed, let’s run a simple distributed training with two clients and one server.

In a file called client.py, import xgboost, Flower, Flower Datasets and other related functions:

import argparse
from typing import Union
from logging import INFO
from datasets import Dataset, DatasetDict
import xgboost as xgb

import flwr as fl
from flwr_datasets import FederatedDataset
from flwr.common.logger import log
from flwr.common import (
    Code,
    EvaluateIns,
    EvaluateRes,
    FitIns,
    FitRes,
    GetParametersIns,
    GetParametersRes,
    Parameters,
    Status,
)
from flwr_datasets.partitioner import IidPartitioner

Dataset partition and hyper-parameter selection#

Prior to local training, we require loading the HIGGS dataset from Flower Datasets and conduct data partitioning for FL:

# Load (HIGGS) dataset and conduct partitioning
# We use a small subset (num_partitions=30) of the dataset for demonstration to speed up the data loading process.
partitioner = IidPartitioner(num_partitions=30)
fds = FederatedDataset(dataset="jxie/higgs", partitioners={"train": partitioner})

# Load the partition for this `node_id`
partition = fds.load_partition(node_id=args.node_id, split="train")
partition.set_format("numpy")

In this example, we split the dataset into two partitions with uniform distribution (IidPartitioner(num_partitions=2)). Then, we load the partition for the given client based on node_id:

# We first define arguments parser for user to specify the client/node ID.
parser = argparse.ArgumentParser()
parser.add_argument(
    "--node-id",
    default=0,
    type=int,
    help="Node ID used for the current client.",
)
args = parser.parse_args()

# Load the partition for this `node_id`.
partition = fds.load_partition(idx=args.node_id, split="train")
partition.set_format("numpy")

After that, we do train/test splitting on the given partition (client’s local data), and transform data format for xgboost package.

# Train/test splitting
train_data, valid_data, num_train, num_val = train_test_split(
    partition, test_fraction=0.2, seed=42
)

# Reformat data to DMatrix for xgboost
train_dmatrix = transform_dataset_to_dmatrix(train_data)
valid_dmatrix = transform_dataset_to_dmatrix(valid_data)

The functions of train_test_split and transform_dataset_to_dmatrix are defined as below:

# Define data partitioning related functions
def train_test_split(partition: Dataset, test_fraction: float, seed: int):
    """Split the data into train and validation set given split rate."""
    train_test = partition.train_test_split(test_size=test_fraction, seed=seed)
    partition_train = train_test["train"]
    partition_test = train_test["test"]

    num_train = len(partition_train)
    num_test = len(partition_test)

    return partition_train, partition_test, num_train, num_test


def transform_dataset_to_dmatrix(data: Union[Dataset, DatasetDict]) -> xgb.core.DMatrix:
    """Transform dataset to DMatrix format for xgboost."""
    x = data["inputs"]
    y = data["label"]
    new_data = xgb.DMatrix(x, label=y)
    return new_data

Finally, we define the hyper-parameters used for XGBoost training.

num_local_round = 1
params = {
    "objective": "binary:logistic",
    "eta": 0.1,  # lr
    "max_depth": 8,
    "eval_metric": "auc",
    "nthread": 16,
    "num_parallel_tree": 1,
    "subsample": 1,
    "tree_method": "hist",
}

The num_local_round represents the number of iterations for local tree boost. We use CPU for the training in default. One can shift it to GPU by setting tree_method to gpu_hist. We use AUC as evaluation metric.

Flower client definition for XGBoost#

After loading the dataset we define the Flower client. We follow the general rule to define XgbClient class inherited from fl.client.Client.

class XgbClient(fl.client.Client):
    def __init__(self):
        self.bst = None
        self.config = None

The self.bst is used to keep the Booster objects that remain consistent across rounds, allowing them to store predictions from trees integrated in earlier rounds and maintain other essential data structures for training.

Then, we override get_parameters, fit and evaluate methods insides XgbClient class as follows.

def get_parameters(self, ins: GetParametersIns) -> GetParametersRes:
    _ = (self, ins)
    return GetParametersRes(
        status=Status(
            code=Code.OK,
            message="OK",
        ),
        parameters=Parameters(tensor_type="", tensors=[]),
    )

Unlike neural network training, XGBoost trees are not started from a specified random weights. In this case, we do not use get_parameters and set_parameters to initialise model parameters for XGBoost. As a result, let’s return an empty tensor in get_parameters when it is called by the server at the first round.

def fit(self, ins: FitIns) -> FitRes:
    if not self.bst:
        # First round local training
        log(INFO, "Start training at round 1")
        bst = xgb.train(
            params,
            train_dmatrix,
            num_boost_round=num_local_round,
            evals=[(valid_dmatrix, "validate"), (train_dmatrix, "train")],
        )
        self.config = bst.save_config()
        self.bst = bst
    else:
        for item in ins.parameters.tensors:
            global_model = bytearray(item)

        # Load global model into booster
        self.bst.load_model(global_model)
        self.bst.load_config(self.config)

        bst = self._local_boost()

    local_model = bst.save_raw("json")
    local_model_bytes = bytes(local_model)

    return FitRes(
        status=Status(
            code=Code.OK,
            message="OK",
        ),
        parameters=Parameters(tensor_type="", tensors=[local_model_bytes]),
        num_examples=num_train,
        metrics={},
    )

In fit, at the first round, we call xgb.train() to build up the first set of trees. the returned Booster object and config are stored in self.bst and self.config, respectively. From the second round, we load the global model sent from server to self.bst, and then update model weights on local training data with function local_boost as follows:

def _local_boost(self):
    # Update trees based on local training data.
    for i in range(num_local_round):
        self.bst.update(train_dmatrix, self.bst.num_boosted_rounds())

    # Extract the last N=num_local_round trees for sever aggregation
    bst = self.bst[
        self.bst.num_boosted_rounds()
        - num_local_round : self.bst.num_boosted_rounds()
    ]

Given num_local_round, we update trees by calling self.bst.update method. After training, the last N=num_local_round trees will be extracted to send to the server.

def evaluate(self, ins: EvaluateIns) -> EvaluateRes:
    eval_results = self.bst.eval_set(
        evals=[(valid_dmatrix, "valid")],
        iteration=self.bst.num_boosted_rounds() - 1,
    )
    auc = round(float(eval_results.split("\t")[1].split(":")[1]), 4)

    return EvaluateRes(
        status=Status(
            code=Code.OK,
            message="OK",
        ),
        loss=0.0,
        num_examples=num_val,
        metrics={"AUC": auc},
    )

In evaluate, we call self.bst.eval_set function to conduct evaluation on valid set. The AUC value will be returned.

Now, we can create an instance of our class XgbClient and add one line to actually run this client:

fl.client.start_client(server_address="127.0.0.1:8080", client=XgbClient())

That’s it for the client. We only have to implement Client`and call :code:`fl.client.start_client(). The string "[::]:8080" tells the client which server to connect to. In our case we can run the server and the client on the same machine, therefore we use "[::]:8080". If we run a truly federated workload with the server and clients running on different machines, all that needs to change is the server_address we point the client at.

Flower Server#

These updates are then sent to the server which will aggregate them to produce a better model. Finally, the server sends this improved version of the model back to each client to finish a complete FL round.

In a file named server.py, import Flower and FedXgbBagging from flwr.server.strategy.

We first define a strategy for XGBoost bagging aggregation.

# Define strategy
strategy = FedXgbBagging(
    fraction_fit=1.0,
    min_fit_clients=2,
    min_available_clients=2,
    min_evaluate_clients=2,
    fraction_evaluate=1.0,
    evaluate_metrics_aggregation_fn=evaluate_metrics_aggregation,
)

def evaluate_metrics_aggregation(eval_metrics):
    """Return an aggregated metric (AUC) for evaluation."""
    total_num = sum([num for num, _ in eval_metrics])
    auc_aggregated = (
        sum([metrics["AUC"] * num for num, metrics in eval_metrics]) / total_num
    )
    metrics_aggregated = {"AUC": auc_aggregated}
    return metrics_aggregated

We use two clients for this example. An evaluate_metrics_aggregation function is defined to collect and wighted average the AUC values from clients.

Then, we start the server:

# Start Flower server
fl.server.start_server(
    server_address="0.0.0.0:8080",
    config=fl.server.ServerConfig(num_rounds=num_rounds),
    strategy=strategy,
)

Tree-based bagging aggregation#

You must be curious about how bagging aggregation works. Let’s look into the details.

In file flwr.server.strategy.fedxgb_bagging.py, we define FedXgbBagging inherited from flwr.server.strategy.FedAvg. Then, we override the aggregate_fit, aggregate_evaluate and evaluate methods as follows:

import json
from logging import WARNING
from typing import Any, Callable, Dict, List, Optional, Tuple, Union, cast

from flwr.common import EvaluateRes, FitRes, Parameters, Scalar
from flwr.common.logger import log
from flwr.server.client_proxy import ClientProxy

from .fedavg import FedAvg


class FedXgbBagging(FedAvg):
    """Configurable FedXgbBagging strategy implementation."""

    def __init__(
        self,
        evaluate_function: Optional[
            Callable[
                [int, Parameters, Dict[str, Scalar]],
                Optional[Tuple[float, Dict[str, Scalar]]],
            ]
        ] = None,
        **kwargs: Any,
    ):
        self.evaluate_function = evaluate_function
        self.global_model: Optional[bytes] = None
        super().__init__(**kwargs)

    def aggregate_fit(
        self,
        server_round: int,
        results: List[Tuple[ClientProxy, FitRes]],
        failures: List[Union[Tuple[ClientProxy, FitRes], BaseException]],
    ) -> Tuple[Optional[Parameters], Dict[str, Scalar]]:
        """Aggregate fit results using bagging."""
        if not results:
            return None, {}
        # Do not aggregate if there are failures and failures are not accepted
        if not self.accept_failures and failures:
            return None, {}

        # Aggregate all the client trees
        global_model = self.global_model
        for _, fit_res in results:
            update = fit_res.parameters.tensors
            for bst in update:
                global_model = aggregate(global_model, bst)

        self.global_model = global_model

        return (
            Parameters(tensor_type="", tensors=[cast(bytes, global_model)]),
            {},
        )

    def aggregate_evaluate(
        self,
        server_round: int,
        results: List[Tuple[ClientProxy, EvaluateRes]],
        failures: List[Union[Tuple[ClientProxy, EvaluateRes], BaseException]],
    ) -> Tuple[Optional[float], Dict[str, Scalar]]:
        """Aggregate evaluation metrics using average."""
        if not results:
            return None, {}
        # Do not aggregate if there are failures and failures are not accepted
        if not self.accept_failures and failures:
            return None, {}

        # Aggregate custom metrics if aggregation fn was provided
        metrics_aggregated = {}
        if self.evaluate_metrics_aggregation_fn:
            eval_metrics = [(res.num_examples, res.metrics) for _, res in results]
            metrics_aggregated = self.evaluate_metrics_aggregation_fn(eval_metrics)
        elif server_round == 1:  # Only log this warning once
            log(WARNING, "No evaluate_metrics_aggregation_fn provided")

        return 0, metrics_aggregated

    def evaluate(
        self, server_round: int, parameters: Parameters
    ) -> Optional[Tuple[float, Dict[str, Scalar]]]:
        """Evaluate model parameters using an evaluation function."""
        if self.evaluate_function is None:
            # No evaluation function provided
            return None
        eval_res = self.evaluate_function(server_round, parameters, {})
        if eval_res is None:
            return None
        loss, metrics = eval_res
        return loss, metrics

In aggregate_fit, we sequentially aggregate the clients’ XGBoost trees by calling aggregate() function:

def aggregate(
    bst_prev_org: Optional[bytes],
    bst_curr_org: bytes,
) -> bytes:
    """Conduct bagging aggregation for given trees."""
    if not bst_prev_org:
        return bst_curr_org

    # Get the tree numbers
    tree_num_prev, _ = _get_tree_nums(bst_prev_org)
    _, paral_tree_num_curr = _get_tree_nums(bst_curr_org)

    bst_prev = json.loads(bytearray(bst_prev_org))
    bst_curr = json.loads(bytearray(bst_curr_org))

    bst_prev["learner"]["gradient_booster"]["model"]["gbtree_model_param"][
        "num_trees"
    ] = str(tree_num_prev + paral_tree_num_curr)
    iteration_indptr = bst_prev["learner"]["gradient_booster"]["model"][
        "iteration_indptr"
    ]
    bst_prev["learner"]["gradient_booster"]["model"]["iteration_indptr"].append(
        iteration_indptr[-1] + paral_tree_num_curr
    )

    # Aggregate new trees
    trees_curr = bst_curr["learner"]["gradient_booster"]["model"]["trees"]
    for tree_count in range(paral_tree_num_curr):
        trees_curr[tree_count]["id"] = tree_num_prev + tree_count
        bst_prev["learner"]["gradient_booster"]["model"]["trees"].append(
            trees_curr[tree_count]
        )
        bst_prev["learner"]["gradient_booster"]["model"]["tree_info"].append(0)

    bst_prev_bytes = bytes(json.dumps(bst_prev), "utf-8")

    return bst_prev_bytes


def _get_tree_nums(xgb_model_org: bytes) -> Tuple[int, int]:
    xgb_model = json.loads(bytearray(xgb_model_org))
    # Get the number of trees
    tree_num = int(
        xgb_model["learner"]["gradient_booster"]["model"]["gbtree_model_param"][
            "num_trees"
        ]
    )
    # Get the number of parallel trees
    paral_tree_num = int(
        xgb_model["learner"]["gradient_booster"]["model"]["gbtree_model_param"][
            "num_parallel_tree"
        ]
    )
    return tree_num, paral_tree_num

In this function, we first fetch the number of trees and the number of parallel trees for the current and previous model by calling _get_tree_nums. Then, the fetched information will be aggregated. After that, the trees (containing model weights) are aggregated to generate a new tree model.

After traversal of all clients’ models, a new global model is generated, followed by the serialisation, and sending back to each client.

Launch Federated XGBoost!#

With both client and server ready, we can now run everything and see federated learning in action. FL systems usually have a server and multiple clients. We therefore have to start the server first:

$ python3 server.py

Once the server is running we can start the clients in different terminals. Open a new terminal and start the first client:

$ python3 client.py --node-id=0

Open another terminal and start the second client:

$ python3 client.py --node-id=1

Each client will have its own dataset. You should now see how the training does in the very first terminal (the one that started the server):

INFO flwr 2023-11-20 11:21:56,454 | app.py:163 | Starting Flower server, config: ServerConfig(num_rounds=5, round_timeout=None)
INFO flwr 2023-11-20 11:21:56,473 | app.py:176 | Flower ECE: gRPC server running (5 rounds), SSL is disabled
INFO flwr 2023-11-20 11:21:56,473 | server.py:89 | Initializing global parameters
INFO flwr 2023-11-20 11:21:56,473 | server.py:276 | Requesting initial parameters from one random client
INFO flwr 2023-11-20 11:22:38,302 | server.py:280 | Received initial parameters from one random client
INFO flwr 2023-11-20 11:22:38,302 | server.py:91 | Evaluating initial parameters
INFO flwr 2023-11-20 11:22:38,302 | server.py:104 | FL starting
DEBUG flwr 2023-11-20 11:22:38,302 | server.py:222 | fit_round 1: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,636 | server.py:236 | fit_round 1 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:38,643 | server.py:173 | evaluate_round 1: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,653 | server.py:187 | evaluate_round 1 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:38,653 | server.py:222 | fit_round 2: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,721 | server.py:236 | fit_round 2 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:38,745 | server.py:173 | evaluate_round 2: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,756 | server.py:187 | evaluate_round 2 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:38,756 | server.py:222 | fit_round 3: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,831 | server.py:236 | fit_round 3 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:38,868 | server.py:173 | evaluate_round 3: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,881 | server.py:187 | evaluate_round 3 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:38,881 | server.py:222 | fit_round 4: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:38,960 | server.py:236 | fit_round 4 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:39,012 | server.py:173 | evaluate_round 4: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:39,026 | server.py:187 | evaluate_round 4 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:39,026 | server.py:222 | fit_round 5: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:39,111 | server.py:236 | fit_round 5 received 2 results and 0 failures
DEBUG flwr 2023-11-20 11:22:39,177 | server.py:173 | evaluate_round 5: strategy sampled 2 clients (out of 2)
DEBUG flwr 2023-11-20 11:22:39,193 | server.py:187 | evaluate_round 5 received 2 results and 0 failures
INFO flwr 2023-11-20 11:22:39,193 | server.py:153 | FL finished in 0.8905023969999988
INFO flwr 2023-11-20 11:22:39,193 | app.py:226 | app_fit: losses_distributed [(1, 0), (2, 0), (3, 0), (4, 0), (5, 0)]
INFO flwr 2023-11-20 11:22:39,193 | app.py:227 | app_fit: metrics_distributed_fit {}
INFO flwr 2023-11-20 11:22:39,193 | app.py:228 | app_fit: metrics_distributed {'AUC': [(1, 0.7572), (2, 0.7705), (3, 0.77595), (4, 0.78), (5, 0.78385)]}
INFO flwr 2023-11-20 11:22:39,193 | app.py:229 | app_fit: losses_centralized []
INFO flwr 2023-11-20 11:22:39,193 | app.py:230 | app_fit: metrics_centralized {}

Congratulations! You’ve successfully built and run your first federated XGBoost system. The AUC values can be checked in metrics_distributed. One can see that the average AUC increases over FL rounds.

The full source code for this example can be found in examples/xgboost-quickstart.

Comprehensive Federated XGBoost#

Now that you have known how federated XGBoost work with Flower, it’s time to run some more comprehensive experiments by customising the experimental settings. In the xgboost-comprehensive example (full code), we provide more options to define various experimental setups, including aggregation strategies, data partitioning and centralised/distributed evaluation. We also support Flower simulation making it easy to simulate large client cohorts in a resource-aware manner. Let’s take a look!

Cyclic training#

In addition to bagging aggregation, we offer a cyclic training scheme, which performs FL in a client-by-client fashion. Instead of aggregating multiple clients, there is only one single client participating in the training per round in the cyclic training scenario. The trained local XGBoost trees will be passed to the next client as an initialised model for next round’s boosting.

To do this, we first customise a ClientManager in server_utils.py:

class CyclicClientManager(SimpleClientManager):
    """Provides a cyclic client selection rule."""

    def sample(
        self,
        num_clients: int,
        min_num_clients: Optional[int] = None,
        criterion: Optional[Criterion] = None,
    ) -> List[ClientProxy]:
        """Sample a number of Flower ClientProxy instances."""

        # Block until at least num_clients are connected.
        if min_num_clients is None:
            min_num_clients = num_clients
        self.wait_for(min_num_clients)

        # Sample clients which meet the criterion
        available_cids = list(self.clients)
        if criterion is not None:
            available_cids = [
                cid for cid in available_cids if criterion.select(self.clients[cid])
            ]

        if num_clients > len(available_cids):
            log(
                INFO,
                "Sampling failed: number of available clients"
                " (%s) is less than number of requested clients (%s).",
                len(available_cids),
                num_clients,
            )
            return []

        # Return all available clients
        return [self.clients[cid] for cid in available_cids]

The customised ClientManager samples all available clients in each FL round based on the order of connection to the server. Then, we define a new strategy FedXgbCyclic in flwr.server.strategy.fedxgb_cyclic.py, in order to sequentially select only one client in given round and pass the received model to next client.

class FedXgbCyclic(FedAvg):
    """Configurable FedXgbCyclic strategy implementation."""

    # pylint: disable=too-many-arguments,too-many-instance-attributes, line-too-long
    def __init__(
        self,
        **kwargs: Any,
    ):
        self.global_model: Optional[bytes] = None
        super().__init__(**kwargs)

    def aggregate_fit(
        self,
        server_round: int,
        results: List[Tuple[ClientProxy, FitRes]],
        failures: List[Union[Tuple[ClientProxy, FitRes], BaseException]],
    ) -> Tuple[Optional[Parameters], Dict[str, Scalar]]:
        """Aggregate fit results using bagging."""
        if not results:
            return None, {}
        # Do not aggregate if there are failures and failures are not accepted
        if not self.accept_failures and failures:
            return None, {}

        # Fetch the client model from last round as global model
        for _, fit_res in results:
            update = fit_res.parameters.tensors
            for bst in update:
                self.global_model = bst

        return (
            Parameters(tensor_type="", tensors=[cast(bytes, self.global_model)]),
            {},
        )

Unlike the original FedAvg, we don’t perform aggregation here. Instead, we just make a copy of the received client model as global model by overriding aggregate_fit.

Also, the customised configure_fit and configure_evaluate methods ensure the clients to be sequentially selected given FL round:

def configure_fit(
    self, server_round: int, parameters: Parameters, client_manager: ClientManager
) -> List[Tuple[ClientProxy, FitIns]]:
    """Configure the next round of training."""
    config = {}
    if self.on_fit_config_fn is not None:
        # Custom fit config function provided
        config = self.on_fit_config_fn(server_round)
    fit_ins = FitIns(parameters, config)

    # Sample clients
    sample_size, min_num_clients = self.num_fit_clients(
        client_manager.num_available()
    )
    clients = client_manager.sample(
        num_clients=sample_size,
        min_num_clients=min_num_clients,
    )

    # Sample the clients sequentially given server_round
    sampled_idx = (server_round - 1) % len(clients)
    sampled_clients = [clients[sampled_idx]]

    # Return client/config pairs
    return [(client, fit_ins) for client in sampled_clients]

def configure_evaluate(
    self, server_round: int, parameters: Parameters, client_manager: ClientManager
) -> List[Tuple[ClientProxy, EvaluateIns]]:
    """Configure the next round of evaluation."""
    # Do not configure federated evaluation if fraction eval is 0.
    if self.fraction_evaluate == 0.0:
        return []

    # Parameters and config
    config = {}
    if self.on_evaluate_config_fn is not None:
        # Custom evaluation config function provided
        config = self.on_evaluate_config_fn(server_round)
    evaluate_ins = EvaluateIns(parameters, config)

    # Sample clients
    sample_size, min_num_clients = self.num_evaluation_clients(
        client_manager.num_available()
    )
    clients = client_manager.sample(
        num_clients=sample_size,
        min_num_clients=min_num_clients,
    )

    # Sample the clients sequentially given server_round
    sampled_idx = (server_round - 1) % len(clients)
    sampled_clients = [clients[sampled_idx]]

    # Return client/config pairs
    return [(client, evaluate_ins) for client in sampled_clients]

Customised data partitioning#

In dataset.py, we have a function instantiate_partitioner to instantiate the data partitioner based on the given num_partitions and partitioner_type. Currently, we provide four supported partitioner type to simulate the uniformity/non-uniformity in data quantity (uniform, linear, square, exponential).

from flwr_datasets.partitioner import (
    IidPartitioner,
    LinearPartitioner,
    SquarePartitioner,
    ExponentialPartitioner,
)

CORRELATION_TO_PARTITIONER = {
    "uniform": IidPartitioner,
    "linear": LinearPartitioner,
    "square": SquarePartitioner,
    "exponential": ExponentialPartitioner,
}


def instantiate_partitioner(partitioner_type: str, num_partitions: int):
    """Initialise partitioner based on selected partitioner type and number of
    partitions."""
    partitioner = CORRELATION_TO_PARTITIONER[partitioner_type](
        num_partitions=num_partitions
    )
    return partitioner

Customised centralised/distributed evaluation#

To facilitate centralised evaluation, we define a function in server_utils.py:

def get_evaluate_fn(test_data):
    """Return a function for centralised evaluation."""

    def evaluate_fn(
        server_round: int, parameters: Parameters, config: Dict[str, Scalar]
    ):
        # If at the first round, skip the evaluation
        if server_round == 0:
            return 0, {}
        else:
            bst = xgb.Booster(params=params)
            for para in parameters.tensors:
                para_b = bytearray(para)

            # Load global model
            bst.load_model(para_b)
            # Run evaluation
            eval_results = bst.eval_set(
                evals=[(test_data, "valid")],
                iteration=bst.num_boosted_rounds() - 1,
            )
            auc = round(float(eval_results.split("\t")[1].split(":")[1]), 4)
            log(INFO, f"AUC = {auc} at round {server_round}")

            return 0, {"AUC": auc}

    return evaluate_fn

This function returns a evaluation function which instantiates a Booster object and loads the global model weights to it. The evaluation is conducted by calling eval_set() method, and the tested AUC value is reported.

As for distributed evaluation on the clients, it’s same as the quick-start example by overriding the evaluate() method insides the XgbClient class in client_utils.py.

Flower simulation#

We also provide an example code (sim.py) to use the simulation capabilities of Flower to simulate federated XGBoost training on either a single machine or a cluster of machines.

from logging import INFO
import xgboost as xgb
from tqdm import tqdm

import flwr as fl
from flwr_datasets import FederatedDataset
from flwr.common.logger import log
from flwr.server.strategy import FedXgbBagging, FedXgbCyclic

from dataset import (
    instantiate_partitioner,
    train_test_split,
    transform_dataset_to_dmatrix,
    separate_xy,
    resplit,
)
from utils import (
    sim_args_parser,
    NUM_LOCAL_ROUND,
    BST_PARAMS,
)
from server_utils import (
    eval_config,
    fit_config,
    evaluate_metrics_aggregation,
    get_evaluate_fn,
    CyclicClientManager,
)
from client_utils import XgbClient

After importing all required packages, we define a main() function to perform the simulation process:

def main():
  # Parse arguments for experimental settings
  args = sim_args_parser()

  # Load (HIGGS) dataset and conduct partitioning
  partitioner = instantiate_partitioner(
      partitioner_type=args.partitioner_type, num_partitions=args.pool_size
  )
  fds = FederatedDataset(
      dataset="jxie/higgs",
      partitioners={"train": partitioner},
      resplitter=resplit,
  )

  # Load centralised test set
  if args.centralised_eval or args.centralised_eval_client:
      log(INFO, "Loading centralised test set...")
      test_data = fds.load_full("test")
      test_data.set_format("numpy")
      num_test = test_data.shape[0]
      test_dmatrix = transform_dataset_to_dmatrix(test_data)

  # Load partitions and reformat data to DMatrix for xgboost
  log(INFO, "Loading client local partitions...")
  train_data_list = []
  valid_data_list = []

  # Load and process all client partitions. This upfront cost is amortized soon
  # after the simulation begins since clients wont need to preprocess their partition.
  for node_id in tqdm(range(args.pool_size), desc="Extracting client partition"):
      # Extract partition for client with node_id
      partition = fds.load_partition(node_id=node_id, split="train")
      partition.set_format("numpy")

      if args.centralised_eval_client:
          # Use centralised test set for evaluation
          train_data = partition
          num_train = train_data.shape[0]
          x_test, y_test = separate_xy(test_data)
          valid_data_list.append(((x_test, y_test), num_test))
      else:
          # Train/test splitting
          train_data, valid_data, num_train, num_val = train_test_split(
              partition, test_fraction=args.test_fraction, seed=args.seed
          )
          x_valid, y_valid = separate_xy(valid_data)
          valid_data_list.append(((x_valid, y_valid), num_val))

      x_train, y_train = separate_xy(train_data)
      train_data_list.append(((x_train, y_train), num_train))

We first load the dataset and perform data partitioning, and the pre-processed data is stored in a list. After the simulation begins, the clients won’t need to pre-process their partitions again.

Then, we define the strategies and other hyper-parameters:

# Define strategy
if args.train_method == "bagging":
    # Bagging training
    strategy = FedXgbBagging(
        evaluate_function=get_evaluate_fn(test_dmatrix)
        if args.centralised_eval
        else None,
        fraction_fit=(float(args.num_clients_per_round) / args.pool_size),
        min_fit_clients=args.num_clients_per_round,
        min_available_clients=args.pool_size,
        min_evaluate_clients=args.num_evaluate_clients
        if not args.centralised_eval
        else 0,
        fraction_evaluate=1.0 if not args.centralised_eval else 0.0,
        on_evaluate_config_fn=eval_config,
        on_fit_config_fn=fit_config,
        evaluate_metrics_aggregation_fn=evaluate_metrics_aggregation
        if not args.centralised_eval
        else None,
    )
else:
    # Cyclic training
    strategy = FedXgbCyclic(
        fraction_fit=1.0,
        min_available_clients=args.pool_size,
        fraction_evaluate=1.0,
        evaluate_metrics_aggregation_fn=evaluate_metrics_aggregation,
        on_evaluate_config_fn=eval_config,
        on_fit_config_fn=fit_config,
    )

# Resources to be assigned to each virtual client
# In this example we use CPU by default
client_resources = {
    "num_cpus": args.num_cpus_per_client,
    "num_gpus": 0.0,
}

# Hyper-parameters for xgboost training
num_local_round = NUM_LOCAL_ROUND
params = BST_PARAMS

# Setup learning rate
if args.train_method == "bagging" and args.scaled_lr:
    new_lr = params["eta"] / args.pool_size
    params.update({"eta": new_lr})

After that, we start the simulation by calling fl.simulation.start_simulation:

# Start simulation
fl.simulation.start_simulation(
    client_fn=get_client_fn(
        train_data_list,
        valid_data_list,
        args.train_method,
        params,
        num_local_round,
    ),
    num_clients=args.pool_size,
    client_resources=client_resources,
    config=fl.server.ServerConfig(num_rounds=args.num_rounds),
    strategy=strategy,
    client_manager=CyclicClientManager() if args.train_method == "cyclic" else None,
)

One of key parameters for start_simulation is client_fn which returns a function to construct a client. We define it as follows:

def get_client_fn(
    train_data_list, valid_data_list, train_method, params, num_local_round
):
    """Return a function to construct a client.

    The VirtualClientEngine will execute this function whenever a client is sampled by
    the strategy to participate.
    """

    def client_fn(cid: str) -> fl.client.Client:
        """Construct a FlowerClient with its own dataset partition."""
        x_train, y_train = train_data_list[int(cid)][0]
        x_valid, y_valid = valid_data_list[int(cid)][0]

        # Reformat data to DMatrix
        train_dmatrix = xgb.DMatrix(x_train, label=y_train)
        valid_dmatrix = xgb.DMatrix(x_valid, label=y_valid)

        # Fetch the number of examples
        num_train = train_data_list[int(cid)][1]
        num_val = valid_data_list[int(cid)][1]

        # Create and return client
        return XgbClient(
            train_dmatrix,
            valid_dmatrix,
            num_train,
            num_val,
            num_local_round,
            params,
            train_method,
        )

    return client_fn

Arguments parser#

In utils.py, we define the arguments parsers for clients, server and simulation, allowing users to specify different experimental settings. Let’s first see the sever side:

import argparse


def server_args_parser():
  """Parse arguments to define experimental settings on server side."""
  parser = argparse.ArgumentParser()

  parser.add_argument(
      "--train-method",
      default="bagging",
      type=str,
      choices=["bagging", "cyclic"],
      help="Training methods selected from bagging aggregation or cyclic training.",
  )
  parser.add_argument(
      "--pool-size", default=2, type=int, help="Number of total clients."
  )
  parser.add_argument(
      "--num-rounds", default=5, type=int, help="Number of FL rounds."
  )
  parser.add_argument(
      "--num-clients-per-round",
      default=2,
      type=int,
      help="Number of clients participate in training each round.",
  )
  parser.add_argument(
      "--num-evaluate-clients",
      default=2,
      type=int,
      help="Number of clients selected for evaluation.",
  )
  parser.add_argument(
      "--centralised-eval",
      action="store_true",
      help="Conduct centralised evaluation (True), or client evaluation on hold-out data (False).",
  )

  args = parser.parse_args()
  return args

This allows user to specify training strategies / the number of total clients / FL rounds / participating clients / clients for evaluation, and evaluation fashion. Note that with --centralised-eval, the sever will do centralised evaluation and all functionalities for client evaluation will be disabled.

Then, the argument parser on client side:

def client_args_parser():
  """Parse arguments to define experimental settings on client side."""
  parser = argparse.ArgumentParser()

  parser.add_argument(
      "--train-method",
      default="bagging",
      type=str,
      choices=["bagging", "cyclic"],
      help="Training methods selected from bagging aggregation or cyclic training.",
  )
  parser.add_argument(
      "--num-partitions", default=10, type=int, help="Number of partitions."
  )
  parser.add_argument(
      "--partitioner-type",
      default="uniform",
      type=str,
      choices=["uniform", "linear", "square", "exponential"],
      help="Partitioner types.",
  )
  parser.add_argument(
      "--node-id",
      default=0,
      type=int,
      help="Node ID used for the current client.",
  )
  parser.add_argument(
      "--seed", default=42, type=int, help="Seed used for train/test splitting."
  )
  parser.add_argument(
      "--test-fraction",
      default=0.2,
      type=float,
      help="Test fraction for train/test splitting.",
  )
  parser.add_argument(
      "--centralised-eval",
      action="store_true",
      help="Conduct evaluation on centralised test set (True), or on hold-out data (False).",
  )
  parser.add_argument(
      "--scaled-lr",
      action="store_true",
      help="Perform scaled learning rate based on the number of clients (True).",
  )

  args = parser.parse_args()
  return args

This defines various options for client data partitioning. Besides, clients also have an option to conduct evaluation on centralised test set by setting --centralised-eval, as well as an option to perform scaled learning rate based on the number of clients by setting --scaled-lr.

We also have an argument parser for simulation:

def sim_args_parser():
    """Parse arguments to define experimental settings on server side."""
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--train-method",
        default="bagging",
        type=str,
        choices=["bagging", "cyclic"],
        help="Training methods selected from bagging aggregation or cyclic training.",
    )

    # Server side
    parser.add_argument(
        "--pool-size", default=5, type=int, help="Number of total clients."
    )
    parser.add_argument(
        "--num-rounds", default=30, type=int, help="Number of FL rounds."
    )
    parser.add_argument(
        "--num-clients-per-round",
        default=5,
        type=int,
        help="Number of clients participate in training each round.",
    )
    parser.add_argument(
        "--num-evaluate-clients",
        default=5,
        type=int,
        help="Number of clients selected for evaluation.",
    )
    parser.add_argument(
        "--centralised-eval",
        action="store_true",
        help="Conduct centralised evaluation (True), or client evaluation on hold-out data (False).",
    )
    parser.add_argument(
        "--num-cpus-per-client",
        default=2,
        type=int,
        help="Number of CPUs used for per client.",
    )

    # Client side
    parser.add_argument(
        "--partitioner-type",
        default="uniform",
        type=str,
        choices=["uniform", "linear", "square", "exponential"],
        help="Partitioner types.",
    )
    parser.add_argument(
        "--seed", default=42, type=int, help="Seed used for train/test splitting."
    )
    parser.add_argument(
        "--test-fraction",
        default=0.2,
        type=float,
        help="Test fraction for train/test splitting.",
    )
    parser.add_argument(
        "--centralised-eval-client",
        action="store_true",
        help="Conduct evaluation on centralised test set (True), or on hold-out data (False).",
    )
    parser.add_argument(
        "--scaled-lr",
        action="store_true",
        help="Perform scaled learning rate based on the number of clients (True).",
    )

    args = parser.parse_args()
    return args

This integrates all arguments for both client and server sides.

Example commands#

To run a centralised evaluated experiment with bagging strategy on 5 clients with exponential distribution for 50 rounds, we first start the server as below:

$ python3 server.py --train-method=bagging --pool-size=5 --num-rounds=50 --num-clients-per-round=5 --centralised-eval

Then, on each client terminal, we start the clients:

$ python3 clients.py --train-method=bagging --num-partitions=5 --partitioner-type=exponential --node-id=NODE_ID

To run the same experiment with Flower simulation:

$ python3 sim.py --train-method=bagging --pool-size=5 --num-rounds=50 --num-clients-per-round=5 --partitioner-type=exponential --centralised-eval

The full code for this comprehensive example can be found in examples/xgboost-comprehensive.