fedbox

This package allows running customizable federated simulations on two different dataset families.

  1. FEMNIST (written digits subsets) (see paper)
  2. CIFAR10 and CIFAR100 (see paper)

Three possible algorithms can be chosen.

  1. FedAvg (see paper)
  2. FedProx (see paper)
  3. Our novel framework proposed in the paper alongside this supplemental code
Examples

This example lets you run a simulation on imbalanced FEMNIST dataset using our proposed algorithm with beta equal to 0.70 (perturbation parameter) and a logistic regression model. Moreover, the simulation in scheduled with 200 global rounds and 10 local optimization epochs per each. The number of training and testing clients is 100 (per each group). The simulation saves the data in the './simulations' directory as a JSON file.

>>> n_rounds = 200
>>> n_epochs = 10
>>> class_imbalance = 10
>>> data_imbalance = 1
>>> n_clients = 100
>>> datasets = femnist(
        './data', 
        n_subsets = n_clients,
        transform = transforms.Compose([
            transforms.Normalize((0.5,), (0.5,))
        ]),
        heterogeneity_degree = class_imbalance,
        samples_variance = data_imbalance,
        strict = False,
        seed = 0
    )
>>> model = MultinomialLogisticRegression(n_inputs = 784, n_classes = 10)
>>> scheme = AdjancencyWeightingScheme(datasets)
>>> logger = Logger(
        directory = './simulations', 
        simulation = { 'dataset': 'femnist', 'class_imbalance': class_imbalance, 'data_imbalance': data_imbalance }, 
        convergence = lambda values: values['step'] if values['accuracy.testing'] >= 0.75 else None, 
        enable = True
    )
>>> coordinator = ours.Coordinator(beta = 0.70, model, datasets, scheme, logger)
>>> coordinator.run(n_iterations = n_rounds, n_epochs = n_epochs, batch_size = 256, step_size = 1e-3, l2_penalty = 1e-4)
>>> logger.flush()

 1'''
 2This package allows running customizable federated simulations on two different dataset families.
 3<ol>
 4  <li>FEMNIST (written digits subsets) (see <a href="https://arxiv.org/pdf/1812.01097">paper</a>)</li>
 5  <li>CIFAR10 and CIFAR100 (see <a href="https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf">paper</a>)</li>
 6</ol>
 7
 8Three possible algorithms can be chosen.
 9<ol>
10  <li>FedAvg (see <a href="https://arxiv.org/pdf/1602.05629">paper</a>)</li>
11  <li>FedProx (see <a href="https://arxiv.org/pdf/1812.06127">paper</a>)</li>
12  <li>Our novel framework proposed in the paper alongside this supplemental code
13</ol>
14
15Examples
16--------
17This example lets you run a simulation on imbalanced FEMNIST dataset using our proposed algorithm
18with beta equal to 0.70 (perturbation parameter) and a logistic regression model. Moreover, the simulation 
19in scheduled with 200 global rounds and 10 local optimization epochs per each. The number of training
20and testing clients is 100 (per each group). The simulation saves the data in the './simulations'
21directory as a JSON file.
22
23>>> n_rounds = 200
24>>> n_epochs = 10
25>>> class_imbalance = 10
26>>> data_imbalance = 1
27>>> n_clients = 100
28>>> datasets = femnist(
29        './data', 
30        n_subsets = n_clients,
31        transform = transforms.Compose([
32            transforms.Normalize((0.5,), (0.5,))
33        ]),
34        heterogeneity_degree = class_imbalance,
35        samples_variance = data_imbalance,
36        strict = False,
37        seed = 0
38    )
39>>> model = MultinomialLogisticRegression(n_inputs = 784, n_classes = 10)
40>>> scheme = AdjancencyWeightingScheme(datasets)
41>>> logger = Logger(
42        directory = './simulations', 
43        simulation = { 'dataset': 'femnist', 'class_imbalance': class_imbalance, 'data_imbalance': data_imbalance }, 
44        convergence = lambda values: values['step'] if values['accuracy.testing'] >= 0.75 else None, 
45        enable = True
46    )
47>>> coordinator = ours.Coordinator(beta = 0.70, model, datasets, scheme, logger)
48>>> coordinator.run(n_iterations = n_rounds, n_epochs = n_epochs, batch_size = 256, step_size = 1e-3, l2_penalty = 1e-4)
49>>> logger.flush()
50'''