-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcaltech_fed.py
171 lines (153 loc) · 7.45 KB
/
caltech_fed.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Python version: 3.6
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import copy
import numpy as np
from torchvision import datasets, transforms
import torch
import sys
sys.path.append('/dataset')
from dataset.caltech_dataset import Custom_Dataset, Server_Dataset
#from dataset.mnist_dataset import Server_Dataset
from utils.sampling import mnist_iid, mnist_noniid, cifar_iid
#from utils.options import args_parser
#from utils.office_options import args_parser
from utils.caltech_options import args_parser
#from models.noisy_update import LocalUpdate
from autodp import rdp_bank, dp_acct, rdp_acct, privacy_calibrator
from models.Update import LocalUpdate, DA_LocalUpdate
from models.Nets import ResNet50M,MLP,Alexnet, CNNMnist, CNNCifar, CNNoffice, Naive, CNNcaltech
from models.Fed import FedAvg, noisy_FedAvg
from models.test import test_img
import models
args = args_parser()
delta = 1e-3
noisy_scale =0.03
clip = 0.08
iteration = args.epochs
prob = args.frac
gaussian = lambda x: rdp_bank.RDP_gaussian({'sigma': noisy_scale /clip}, x)
acct = rdp_acct.anaRDPacct()
if __name__ == '__main__':
# parse args
args = args_parser()
args.device = torch.device('cuda:{}'.format(args.gpu) if torch.cuda.is_available() and args.gpu != -1 else 'cpu')
# load dataset and split users
if args.dataset == 'mnist':
trans_mnist = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
dataset_train = datasets.MNIST(root='/tmp', train=True, download=True, transform=trans_mnist)
dataset_global = Server_Dataset('mnist', transform = trans_mnist, da= True)
dataset_test = datasets.MNIST(root = '/tmp', train=False, download=True, transform=trans_mnist)
# sample users
print('len of target dataset', len(dataset_global))
print('use iid distribution data', args.iid)
if args.iid:
dict_users = mnist_iid(dataset_train, args.num_users)
else:
dict_users = mnist_noniid(dataset_train, args.num_users)
elif args.dataset == 'cifar':
trans_cifar = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
dataset_train = datasets.CIFAR10(root='/tmp', train=True, download=True, transform=trans_cifar)
dataset_test = datasets.CIFAR10(root = '/tmp', train=False, download=True, transform=trans_cifar)
if args.iid:
dict_users = cifar_iid(dataset_train, args.num_users)
else:
exit('Error: only consider IID setting in CIFAR10')
elif args.dataset == 'caltech':
trans_office = transforms.Compose([transforms.Resize(256),transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
dataset_train = Custom_Dataset('caltech', transform = trans_office)
dataset_test = Server_Dataset('caltech', transform = trans_office)
dict_users = dataset_train.dic_user
dataset_global = Server_Dataset('caltech', transform = trans_office, da= True)
print('size of testing data', len(dataset_test))
print('dict_user in caltech', dict_users.keys())
elif args.dataset == 'digit':
#transform_digit = transforms.Compose([transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.485), (0.229))])
transform_digit = transforms.Compose([transforms.Resize(32), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),])
dataset_train = Custom_Dataset('digit', transform = transform_digit)
dataset_global = Custom_Dataset('digit', transform = transform_digit, da= True)
dataset_test = Server_Dataset('digit', transform = transform_digit)
dict_users = dataset_train.dic_user
print('length of training', len(dataset_train))
print('dic_user[0]', dict_users[0])
print('size of testing data', len(dataset_test))
print('dict_user in digit', dict_users.keys())
else:
exit('Error: unrecognized dataset')
img_size = dataset_train[0][0].shape
#privacy_analysis()
# build model
if args.model == 'cnn' and args.dataset == 'cifar':
net_glob = CNNCifar(args=args).to(args.device)
elif args.model =='cnn' and args.dataset =='caltech':
#net_glob = models.init_model(name = 'resnet50', num_classes =31, loss = {'xent'}, use_gpu = True)
net_glob = ResNet50M(args=args).to(args.device)
#net_glob = CNNcaltech(args=args).to(args.device)
#net_glob = Alexnet(args=args).to(args.device)
elif args.model =='cnn' and args.dataset == 'digit':
net_glob = CNNCifar(args=args).to(args.device)
elif args.model == 'cnn' and args.dataset == 'mnist':
net_glob = ResNet50M(args=args).to(args.device)
#net_glob = CNNMnist(args=args).to(args.device)
elif args.model == 'mlp':
len_in = 1
for x in img_size:
len_in *= x
net_glob = MLP(dim_in=len_in, dim_hidden=200, dim_out=args.num_classes).to(args.device)
else:
exit('Error: unrecognized model')
print(net_glob)
net_glob.train()
# copy weights
w_glob = net_glob.state_dict()
# training
loss_train = []
cv_loss, cv_acc = [], []
val_loss_pre, counter = 0, 0
net_best = None
best_loss = None
val_acc_list, net_list = [], []
acct.compose_poisson_subsampled_mechanisms(gaussian, prob,coeff = 10)
import math
nb_batch = int(args.frac * len(dataset_train)/args.local_bs)
for iter in range(args.epochs):
w_locals, loss_locals = [], []
m = max(int(args.frac * args.num_users), 1)
LR= args.lr / math.pow((1 + 10 * (iter*nb_batch - 1) / (args.epochs*nb_batch)), 0.65)
idxs_users = np.random.choice(range(args.num_users), m, replace=False)
for idx in idxs_users:
if args.da:
local = DA_LocalUpdate(args=args, dataset=dataset_train,tgt_dataset = dataset_global, idxs=dict_users[idx])
else:
local = LocalUpdate(args=args, dataset=dataset_train, idxs=dict_users[idx], iter = LR)
w, loss = local.train(net=copy.deepcopy(net_glob).to(args.device))
w_locals.append(copy.deepcopy(w))
loss_locals.append(copy.deepcopy(loss))
# update global weights
w_glob = FedAvg(w_locals)
#w_glob = noisy_FedAvg(w_locals, w_glob, idx=iter)
# copy weight to net_glob
net_glob.load_state_dict(w_glob)
# print loss
loss_avg = sum(loss_locals) / len(loss_locals)
print('Round {:3d}, Average loss {:.3f}'.format(iter, loss_avg))
loss_train.append(loss_avg)
if iter %1 ==0 and iter>0 :
acc_test, loss_test = test_img(net_glob, dataset_test, args)
print("iter {} Testing accuracy: {:.2f}".format(iter, acc_test))
acct.compose_poisson_subsampled_mechanisms(gaussian, prob,coeff = 1)
print('current privacy cost', acct.get_eps(delta))
# plot loss curve
plt.figure()
plt.plot(range(len(loss_train)), loss_train)
plt.ylabel('train_loss')
plt.savefig('./save/fed_{}_{}_{}_C{}_iid{}.png'.format(args.dataset, args.model, args.epochs, args.frac, args.iid))
# testing
net_glob.eval()
acc_train, loss_train = test_img(net_glob, dataset_train, args)
acc_test, loss_test = test_img(net_glob, dataset_test, args)
print("Training accuracy: {:.2f}".format(acc_train))
print("Testing accuracy: {:.2f}".format(acc_test))