Configuration

Introduction

When starting the program a config file config.json is passed to the application. It defines the core execution parameters.

Attributes and values

The following table lists all attributes and their possible values.

Key	optional	Description	Min	Max	Default Value	Example Value
no_rounds	no	Number of rounds to run the simulation for	1	-	-	10
algorithm	no	List of algorithms to simulate	-	-	-	["PowD", "FedCS", "random", "ActiveFL", "CEP"]
alogrithm_config	yes	Algorithm configuration	-	-	-	See algorithm documentation
no_epochs	yes	Number of epochs to run on each client per round	1	-	1	1
no_clients	no	Number of clients to simulate	1	-	-	1000
batch_size	yes	Batch size, will affect behavior of training and validation. Can lead to errors with batchnorm, etc. if changed	1	-	32	32
verbose	yes	Enables additional logging	-	-	true	true or false
device	yes	Pytorch device string	-	-	cpu	cpu or cuda
num_cpu_per_client	yes	Number of CPU cores to assign to each client. Integer.	1	CPU_COUNT	2	2
num_gpu_per_client	yes	Number of GPU resources to assign to each client. Only used if device is set to cuda. Partial allocation is possible.	0	-	0.1	0.1
timeout	no	Timeout imposed on clients, in seconds. This affects only the simulation.	1	-	-	60
generate_client_states	yes	Whether to generate new client states when running the simulation. If False is selected, errors may occur if not enough states are predefined.	-	-	true	true or false
client_state_file	no	File to use for client states. Will be overwritten if generate_client_states is True.	-	-	-	client_states.csv
distribute_data	yes	Whether to regenerate the data distribution.	-	-	true	true or false
data_distribution_file	no	File containing an existing data distribution, not used when distribute_data is True	-	-	-	data_distribution.csv
output_dir	no	Folder prefix to use for output. Will not be overwritten.	-	-	-	output
client_configuration_file	no	File that contains the available client configurations. See documentation on Client Configuration.	-	-	-	client_configurations.csv
max_workers	yes	Maximum number of workers that should be used when running distributed tasks. (May not always be enforced)	-	-	32	32
data_config	yes	Configuration for data skew, distortion	-	-	-	See data configuration
validation_config	yes	Configuration of the final validation runs	-	-	-	See validation configuration
simulation_config	yes	Configuration relating to performance, reliability and network bandwidth of clients	-	-	-	See simulation configuration
base_strategy	yes	Base strategy to use for aggregation	-	-	FedAvg	FedAvg, FedAvgM, FedMedian