forked from CUHK-AIM-Group/DEeR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
259 lines (227 loc) · 12.8 KB
/
main.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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
import time
import torch
from datasets import Data
from nodes import Node
from args import args_parser
from utils import *
import numpy as np
import os
import torch.nn as nn
import copy
import torch.optim as optim
import torch.nn.functional as F
import math
from pyhessian import hessian
from server_funct import *
#import wandb
from client_funct import *
import pprint
import argparse
import warnings
from utils import compute_noise_multiplier
warnings.filterwarnings('ignore')
np.set_printoptions(precision=7, suppress=True)
'''
ViT-L/14 768
ViT-B/16 512
ViT-B/32 512
'''
def generate_matchlist(node_num, ratio = 0.5):
candidate_list = [i for i in range(node_num)]
select_num = int(ratio * node_num)
match_list = np.random.choice(candidate_list, select_num, replace = False).tolist()
return match_list
_utils_pp = pprint.PrettyPrinter()
def pprint(x):
_utils_pp.pprint(x)
if __name__ == '__main__':
import gc
gc.collect() # 清理内存
#args = args_parser()
##### Exp settings #####
##### change it for different exps #####
#args.client_method = 'fedetf'
parser = argparse.ArgumentParser()
# Data
parser.add_argument('--data_dir', type=str, default='./data/Retinal_OCT-C8/',
help='./data/Retinal_OCT-C8/, ./data/kvasir-dataset-v2-processed-224')
parser.add_argument('--iid', type=int, default=0,
help='set 1 for iid, and 0 for noniid (dir. sampling)')
parser.add_argument('--batchsize', type=int, default=128,
help="batchsize")
parser.add_argument('--dirichlet_alpha', type=float, default=0.5,
help="dirichlet_alpha")
parser.add_argument('--num_classes', type=int, default=8,
help="num_classes")
# System
parser.add_argument('--device', type=str, default='0',
help="cuda device: {cuda, cpu}")
parser.add_argument('--node_num', type=int, default=20,
help="Number of nodes")
parser.add_argument('--T', type=int, default=200,
help="Number of communication rounds")
parser.add_argument('--E', type=int, default=3,
help="Number of local epochs: E")
parser.add_argument('--dataset', type=str, default='cifar10', #Kvasir
help="Type of algorithms:{mnist, cifar10,cifar100, fmnist, tinyimagenet}")
parser.add_argument('--local_model', type=str, default='CNN',
help='Type of local model: {CNN, ResNet20, ResNet18}')
parser.add_argument('--random_seed', type=list, default=[10, 100, 1000], #
help="random seed for the whole experiment")
parser.add_argument('--exp_name', type=str, default='FirstTable',
help="experiment name")
# Client function
parser.add_argument('--optimizer', type=str, default='sgd',
help="optimizer: {sgd, adam}")
parser.add_argument('--lr', type=float, default=0.04,
help='learning rate')
parser.add_argument('--local_wd_rate', type=float, default=5e-4,
help='clients local wd rate')
parser.add_argument('--momentum', type=float, default=0.9,
help='SGD momentum')
parser.add_argument('--mu', type=float, default=0.001,
help="proximal term mu")
#add meilu
parser.add_argument('--method', type=str, default='LORA',
help="method") # LORA, FFA-LoRA, DEeR, DP-DyLoRA
parser.add_argument('--lora_r', type=int, default=1,
help="lora_r") #CLIP, BiomedCLIP
# DP noise
parser.add_argument('--is_DP', type=int, default=0,
help="is_DP") # whether use DP
parser.add_argument('--C', type=float, default=2, # 2 5 10
help='the threshold of clipping in DP')
parser.add_argument('--epsilon', type=float, default=1.,
help='the standard deviation of client-level DP noise')
parser.add_argument('--module1', type=int, default=1,
help="module1")
parser.add_argument('--module2', type=int, default=1,
help="module2")
args = parser.parse_args()
all_acc, all_recall, all_prec, all_f1, all_auc = [],[],[],[],[]
if args.dataset == 'Kvasir':
batchsize = 128
elif args.dataset == 'OCT':
batchsize = 512
args.batchsize = batchsize//args.node_num
random_seeds = args.random_seed
lr = args.lr
for random_seed in random_seeds:
args.random_seed = random_seed
args.lr = lr
print('starting run seed', args.random_seed)
setup_seed(random_seed)
now = time.strftime("%Y-%m-%d %H:%M:%S")
print('The starting time :{}'.format(now), flush=True)
pprint(vars(args))
select_list_recorder = [[i for i in range(args.node_num)] for _ in range(args.T)]
setting_name = args.exp_name + '_' + args.method + '_' + args.dataset + '_' + args.local_model + '_nodenum' + str(args.node_num) + '_dir' + str(args.dirichlet_alpha) +'_E'+ str(args.E) \
+ '_' + args.server_method + '_' + args.client_method + '_seed' + str(args.random_seed)
root_path = './'
output_path = 'results/'
if not os.path.exists(os.path.join(root_path, output_path)):
os.makedirs(os.path.join(root_path, output_path))
os.environ['CUDA_VISIBLE_DEVICES'] = args.device
data = Data(args)
size_weights = [1./args.node_num] * args.node_num
print('size-based weights',size_weights, flush=True)
central_node = Node(args, -1, train_loader = None, val_loader=data.val_loader, test_loader=data.test_loader)
# initialize the client nodes
client_nodes = {}
for i in range(args.node_num):
client_nodes[i] = Node(args, i, train_loader=data.train_loaders[i], val_loader=None, test_loader=None)
client_nodes[i].model.load_state_dict(copy.deepcopy(central_node.model.state_dict()))
client_nodes[i].text_features = copy.deepcopy(central_node.text_features.data)
test_acc_recorder = []
best_val_acc = 0
best_test_acc = 0
best_test_recall=0
best_test_prec=0
best_test_f1=0
best_test_auc=0
print(setting_name, flush=True)
##################################
noise_multiplier = None
if args.is_DP:
target_epsilon = args.epsilon
target_delta = 1./args.node_num
global_epoch = args.T
local_epoch = args.E
batch_size = args.batchsize
sample_sizes = []
for i in range(args.node_num):
sample_sizes.append(len(data.train_loader[i]))
client_data_sizes = sample_sizes
noise_multiplier = compute_noise_multiplier(args, target_epsilon, target_delta, global_epoch, local_epoch, batch_size, client_data_sizes)
print(f'noise_multiplier : {noise_multiplier}')
#################################
for rounds in range(0, args.T):
print('===============Stage 1 The {:d}-th round==============='.format(rounds + 1), flush=True)
#lr_scheduler(rounds, client_nodes, args)
# Client selection
select_list = select_list_recorder[rounds]
if args.method == 'DEeR':
if args.module1 == 1:
for i in range(len(client_nodes)):
for name, param in client_nodes[i].model.named_parameters():
if 'linear_a' in name:
param.requires_grad = False
if 'linear_b' in name:
param.requires_grad = True
client_nodes[i].optimizer = torch.optim.SGD(params=filter(lambda p: p.requires_grad, client_nodes[i].model.parameters()), lr=args.lr, momentum=args.momentum, weight_decay=args.local_wd_rate)
client_nodes, train_loss = Client_update(args, client_nodes, central_node, select_list)
print('Train loss is {:.5f}'.format(train_loss), flush=True)
central_node = Server_update(args, central_node, client_nodes, select_list, size_weights,noise_multiplier, open_ab = 'b')
for i in range(len(client_nodes)):
for name, param in client_nodes[i].model.named_parameters():
if 'linear_a' in name:
param.requires_grad = True
if 'linear_b' in name:
param.requires_grad = False
client_nodes[i].optimizer = torch.optim.SGD(params=filter(lambda p: p.requires_grad, client_nodes[i].model.parameters()), lr=args.lr, momentum=args.momentum, weight_decay=args.local_wd_rate)
client_nodes, train_loss = Client_update(args, client_nodes, central_node, select_list)
print('Train loss is {:.5f}'.format(train_loss), flush=True)
central_node = Server_update(args, central_node, client_nodes, select_list, size_weights,noise_multiplier, open_ab = 'a')
else:
client_nodes, train_loss = Client_update(args, client_nodes, central_node, select_list)
print('Train loss is {:.5f}'.format(train_loss), flush=True)
central_node = Server_update(args, central_node, client_nodes, select_list, size_weights,noise_multiplier, open_ab = 'ab')
elif args.method == 'FFA-LoRA':
client_nodes, train_loss = Client_update(args, client_nodes, central_node, select_list)
print('Train loss is {:.5f}'.format(train_loss), flush=True)
central_node = Server_update(args, central_node, client_nodes, select_list, size_weights,noise_multiplier, open_ab = 'b')
elif args.method == 'LoRA' or args.method == 'DP-DyLoRA':
client_nodes, train_loss = Client_update(args, client_nodes, central_node, select_list)
print('Train loss is {:.5f}'.format(train_loss), flush=True)
central_node = Server_update(args, central_node, client_nodes, select_list, size_weights,noise_multiplier, open_ab = 'ab')
else:
assert False
val_acc, val_recall, val_prec, val_f1, val_auc = validate(args, central_node, which_dataset = 'validate')
print('Val acc: {:.3f}'.format(val_acc)+ ', recall: {:.3f}'.format(val_recall)+ ', prec: {:.3f}'.format(val_prec)+ ', f1: {:.3f}'.format(val_f1)+ ', auc: {:.3f}'.format(val_auc), flush=True)
print('Test acc: {:.3f}'.format(best_test_acc)+ ', recall: {:.3f}'.format(best_test_recall)+ ', prec: {:.3f}'.format(best_test_prec)+ ', f1: {:.3f}'.format(best_test_f1)+ ', auc: {:.3f}'.format(best_test_auc), flush=True)
print()
if val_acc+val_recall+val_prec+val_f1+val_auc>best_val_acc:
best_val_acc = val_acc+val_recall+val_prec+val_f1+val_auc
best_test_acc,best_test_recall, best_test_prec, best_test_f1,best_test_auc = validate(args, central_node, which_dataset = 'test')
print('Test acc: {:.3f}'.format(best_test_acc)+ ', recall: {:.3f}'.format(best_test_recall)+ ', prec: {:.3f}'.format(best_test_prec)+ ', f1: {:.3f}'.format(best_test_f1)+ ', auc: {:.3f}'.format(best_test_auc), flush=True)
print()
#torch.save(central_node.model.state_dict(), os.path.join(root_path, output_path, setting_name+'_finalmodel.pth'))
all_acc.append(best_test_acc)
all_recall.append(best_test_recall)
all_prec.append(best_test_prec)
all_f1.append(best_test_f1)
all_auc.append(best_test_auc)
end = time.strftime("%Y-%m-%d %H:%M:%S")
print('The ending time :{}'.format(end))
print('===========================================================')
print('Best test acc:', all_acc)
print('Best test acc mean: {:.5f}'.format(np.mean(all_acc)),'Best test acc std: {:.5f}'.format(np.std(all_acc)) )
print('Best test recall:', all_recall)
print('Best test recall mean: {:.5f}'.format(np.mean(all_recall)),'Best test recall std: {:.5f}'.format(np.std(all_recall)) )
print('Best test prec:', all_prec)
print('Best test prec mean: {:.5f}'.format(np.mean(all_prec)),'Best test prec std: {:.5f}'.format(np.std(all_prec)) )
print('Best test f1:', all_f1)
print('Best test f1 mean: {:.5f}'.format(np.mean(all_f1)),'Best test f1 std: {:.5f}'.format(np.std(all_f1)) )
print('Best test auc:', all_auc)
print('Best test auc mean: {:.5f}'.format(np.mean(all_auc)),'Best test auc std: {:.5f}'.format(np.std(all_auc)) )
print('===========================================================')