asociita 0.3.4

An intuitive and modular simulator for assessing the marginal value of a client's contribution in a decentralized setting.

Setting Configuration

In order to run the simulation, the Orchestrator instance must receive a settings object that contains all the necessary parameters. It is possible to store those parameters in a JSON format and load them as the Python dictionary by using asociita.utils.helper.load_from_json function. Below is an exemplary settings object embedded as a json file. All the elements are necessary unless stated otherwise.

{
    "orchestrator":{
        "iterations": int,
        "number_of_nodes": int,
        "local_warm_start": bool,
        "sample_size": int,
        "evaluation": "none" | "full"
        "save_metrics": bool,
	"save_models": bool,
	"save_path": str
	"nodes": [0,
	1,
	2]
    },
    "nodes":{
    "local_epochs": int,
    "model_settings": {
        "optimizer": "RMS",
        "batch_size": int,
        "learning_rate": float}
        }
}

The settings contains two dictionaries: orchestrator and nodes.

orchestrator contains all the settings necessary details of the training:

• iterations is the number of rounds to be performed. Example: iterations:12
• number_of_nodes is the number of nodes that will be included in the training. Example: number_of_nodes: 10
• local_warm_start allows to distribute various pre-trained weights to different local clients. Not implemeneted yet. Example: local_warm_start: false.
• sample_size is the size of the sample that will be taken each round. Example: sample_size : 4.
• evaluation allows to control the evaluation procedure across the clients. Currently, only none or full are supported. Setting the evaluation to full will perform a full evaluation of every client included in the training. Example: evaluation: full
• save_metrics allows to control whether the metrics should be saved in a csv file. Example: save_metrics: true.
• save_models allows to control whether the models should be saved. Not implemeneted yet. Example: save_metrics: false.
• save_path is the system path that will be used when saving the model. It is possible to define a saving_path in a method call.
• nodes is the list containing the ids of all the nodes participating in the training. Length of nodes must be equal number_of_nodes.

nodes contains all the necessary configuration for nodes.

• "local_epochs": the number of local epochs to be performed on the local nodes.
• "model_settings" is a dictionary containing all the parameters for training the model.
  • optimizer is an optimizer that will be used during the training. Example: optimizer: "RMS"
  • batch_size is the batch size that will be used during the training. Example: batch_size: 32
  • learning_rate is the learning rate that will be used during the training. Example: learning_rate: 0.001