-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_federated_server.py
182 lines (160 loc) · 8.07 KB
/
train_federated_server.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
# -*- coding: utf-8 -*-
from flwr.server.client_proxy import ClientProxy
from flwr.common import (
EvaluateIns,
EvaluateRes,
FitIns,
FitRes,
MetricsAggregationFn,
NDArrays,
Parameters,
Scalar,
ndarrays_to_parameters,
parameters_to_ndarrays,
)
from typing import Callable, Dict, List, Optional, Tuple, Union
import flwr as fl
import torch
import numpy as np
import argparse
import os
from collections import OrderedDict
from save_load_util import save_metrics
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
def get_parser():
parser = argparse.ArgumentParser(description='set federated learning server argument')
parser.add_argument('-le', '--local-epoch', type=int, default=2,
help='the number times that the learning algorithm work through the entire training dataset, default 5')
parser.add_argument('-e', '--eval-time', type=int, default=2,
help='the number evaluate times in a epoch, default 2')
parser.add_argument('-b', '--batch-size', type=int, default=32,
help='batch size of training loader, default 32')
parser.add_argument('-l', '--learning-rate', type=float, default=0.001,
help='learning rate, default 0.001')
parser.add_argument('-r', '--num-round', type=int, default=2,
help='num of round of server aggregation, defaults to 2')
parser.add_argument('-ff', '--fraction-fit', type=float, default=1.0,
help='Fraction of clients used during training. Defaults to 1.0.')
parser.add_argument('-fe', '--fraction-evaluate', type=float, default=1.0,
help='Fraction of clients used during validation. Defaults to 1.0.')
parser.add_argument('-mf', '--min-fit-clients', type=int, default=2,
help='Minimum number of clients used during training. Defaults to 2.')
parser.add_argument('-me', '--min-evaluate-clients', type=int, default=2,
help='Minimum number of clients used during validation. Defaults to 2.')
parser.add_argument('-ma', '--min-available-clients', type=int, default=2,
help='Minimum number of total clients in the system. Defaults to 2.')
parser.add_argument('-s', '--saving-directory', type=str, default='.',
help='specify working directory to save model files, default current')
parser.add_argument('-ba', '--best-valid-accuracy', type=float, default=0.0,
help='specify the minimum validation accuracy, default 0.0')
return parser
local_config = None # define global server config set by main argument
metrics_dict = {
'train_acc_list': [],
'valid_acc_list': [],
'train_loss_list': [],
'valid_loss_list': []
}
def get_on_fit_config_fn(args):
'''Return a function which returns training configurations.'''
def fit_config(server_round: int):
'''Return a configuration with static batch size and (local) epochs.'''
config = {
'current_round': server_round,
'learning_rate': args.learning_rate,
'batch_size': args.batch_size,
'eval_time': args.eval_time,
'local_epochs': args.local_epoch
}
return config
return fit_config
def get_on_evaluate_config_fn(args):
'''Return a function which returns testing configurations.'''
def evaluate_config(server_round: int):
'''Return a configuration with static batch size and (local) epochs.'''
config = {
'current_round': server_round,
'batch_size': args.batch_size
}
return config
return evaluate_config
def fit_metrics_aggregation_fn(fit_metrics):
total_data_size = sum(data_size for data_size, _ in fit_metrics)
factors = [data_size / total_data_size for data_size, _ in fit_metrics]
# metrics[0]為client data size，metrics[1]為client train accuracy值及valid accuracy值的dict
agg_train_acc = sum(metrics[1]['train_accuracy'] * factor for metrics, factor in zip(fit_metrics, factors))
agg_valid_acc = sum(metrics[1]['valid_accuracy'] * factor for metrics, factor in zip(fit_metrics, factors))
agg_train_loss = sum(metrics[1]['train_loss'] * factor for metrics, factor in zip(fit_metrics, factors))
agg_valid_loss = sum(metrics[1]['valid_loss'] * factor for metrics, factor in zip(fit_metrics, factors))
metrics_dict['train_acc_list'].append(agg_train_acc)
metrics_dict['valid_acc_list'].append(agg_valid_acc)
metrics_dict['train_loss_list'].append(agg_train_loss)
metrics_dict['valid_loss_list'].append(agg_valid_loss)
return {'train_accuracy': agg_train_acc,
'valid_accuracy': agg_valid_acc,
'train_loss': agg_train_loss,
'valid_loss': agg_valid_loss}
def evaluate_metrics_aggregation_fn(eval_metrics):
total_data_size = sum(data_size for data_size, _ in eval_metrics)
factors = [data_size / total_data_size for data_size, _ in eval_metrics]
# metrics[0]為client data size，metrics[1]為client testing loss值及accuracy的dict
agg_acc = sum(metrics[1]['accuracy'] * factor for metrics, factor in zip(eval_metrics, factors))
agg_loss = sum(metrics[1]['loss'] * factor for metrics, factor in zip(eval_metrics, factors))
return {'accuracy': agg_acc,
'loss': agg_loss}
def evaluate_fn(server_round, parameters, config):
params_dict = zip(MyLSTM().state_dict().keys(), parameters)
state_dict = OrderedDict({k: torch.tensor(v) for k, v in params_dict})
if len(metrics_dict['valid_acc_list']) > 0 and \
local_config['best_valid_acc'] < metrics_dict['valid_acc_list'][-1]:
local_config['best_valid_acc'] = metrics_dict['valid_acc_list'][-1]
model = MyLSTM().to(device)
model.load_state_dict(state_dict, strict=True)
save_model(local_config['saving_directory'] + '/server_model.pt', model, metrics_dict['valid_acc_list'][-1])
if server_round == local_config['num_round']:
save_metrics(local_config['saving_directory'] + '/server_accuracy_metrics.pt',
metrics_dict['train_acc_list'],
metrics_dict['valid_acc_list'],
list(range(1, server_round + 1)))
save_metrics(local_config['saving_directory'] + '/server_loss_metrics.pt',
metrics_dict['train_loss_list'],
metrics_dict['valid_loss_list'],
list(range(1, server_round + 1)))
def main(args):
global local_config
local_config = {
'saving_directory': args.saving_directory,
'num_round': args.num_round,
'best_valid_acc': args.best_valid_accuracy
}
initial_parameters = None
if os.path.exists(local_config['saving_directory'] + '/server_model.pt'):
state_dict = torch.load(local_config['saving_directory'] + '/server_model.pt', map_location=device)
model = MyLSTM().to(device)
model.load_state_dict(state_dict['model_state_dict'], strict=True)
weights = [val.cpu().numpy() for name, val in model.state_dict().items()]
initial_parameters = ndarrays_to_parameters(weights)
# Define strategy
strategy = fl.server.strategy.FedAvg(
initial_parameters=initial_parameters,
fraction_fit=args.fraction_fit,
fraction_evaluate=args.fraction_evaluate,
min_fit_clients=args.min_fit_clients,
min_evaluate_clients=args.min_evaluate_clients,
min_available_clients=args.min_available_clients,
on_fit_config_fn=get_on_fit_config_fn(args),
on_evaluate_config_fn=get_on_evaluate_config_fn(args),
evaluate_fn=evaluate_fn,
fit_metrics_aggregation_fn=fit_metrics_aggregation_fn,
evaluate_metrics_aggregation_fn=evaluate_metrics_aggregation_fn
)
# Start server
fl.server.start_server(
server_address="[::1]:9999",
config=fl.server.ServerConfig(num_rounds=args.num_round),
strategy=strategy
)
if __name__ == '__main__':
parser = get_parser()
args = parser.parse_args()
main(args)