forked from jahanxb/flcode
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain_fed_client_node-fedml-02sep.py
210 lines (197 loc) · 9.78 KB
/
main_fed_client_node-fedml-02sep.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Python version: 3.6
import copy
import sys
import threading
import grpc
import numpy as np
import time, math
import torch
from utils.data_utils import data_setup, DatasetSplit
from utils.model_utils import *
from utils.aggregation import *
from options import call_parser
from models.Update import LocalUpdate
from models.test import test_img
from torch.utils.data import DataLoader
from concurrent import futures
# from utils.rdp_accountant import compute_rdp, get_privacy_spent
import warnings
warnings.filterwarnings("ignore")
torch.cuda.is_available()
import fedml_pb2, fedml_pb2_grpc
def client_node():
pid = os.getpid()
with grpc.insecure_channel("localhost:9999") as channel:
stub = fedml_pb2_grpc.FedMlManagementServiceStub(channel)
try:
while True:
print("Active PID : %i" % pid)
# time.sleep(10)
start = time.time()
args = call_parser()
print("num user is: ", args.num_users)
response = stub.UserIndex(fedml_pb2.FedmlRequest(
user_index=args.num_users,
dataset=args.dataset,
gpu=args.gpu,
round=args.round
))
# response = stub.ping(pingpong_pb2.Ping(count=counter))
# counter = response.count
# if counter % 100 == 0:
# print(
# "%.4f : resp=%s : procid=%i"
# % (time.time() - start, response.count, pid)
# )
# # counter = 0
time.sleep(0.001)
# #########################################
# # Fedml Code
# #######################################
#
# torch.manual_seed(args.seed + args.repeat)
# torch.cuda.manual_seed(args.seed + args.repeat)
# np.random.seed(args.seed + args.repeat)
#
# args, dataset_train, dataset_test, dict_users = data_setup(args)
# print("{:<50}".format("=" * 15 + " data setup " + "=" * 50)[0:60])
# print(
# 'length of dataset:{}'.format(len(dataset_train) + len(dataset_test)))
# print('num. of training data:{}'.format(len(dataset_train)))
# print('num. of testing data:{}'.format(len(dataset_test)))
# print('num. of classes:{}'.format(args.num_classes))
# print('num. of users:{}'.format(len(dict_users)))
# # sample_per_users = int(sum([ len(dict_users[i]) for i in range(len(dict_users))])/len(dict_users))
#
# sample_per_users = 0
# for i in range(len(dict_users)):
# sample_per_users += int(sum([len(dict_users[i]) / len(dict_users)]))
#
# sample_per_users = 5 # for two users , we take 25000 samples as per the loop
#
# print('num. of samples per user:{}'.format(sample_per_users))
# if args.dataset == 'fmnist' or args.dataset == 'cifar':
# dataset_test, val_set = torch.utils.data.random_split(
# dataset_test, [9000, 1000])
# print(len(dataset_test), len(val_set))
# elif args.dataset == 'svhn':
# dataset_test, val_set = torch.utils.data.random_split(
# dataset_test, [len(dataset_test) - 2000, 2000])
# print(len(dataset_test), len(val_set))
#
# print("{:<50}".format("=" * 15 + " log path " + "=" * 50)[0:60])
# log_path = set_log_path(args)
# print(log_path)
#
# args, net_glob = model_setup(args)
# print("{:<50}".format("=" * 15 + " model setup " + "=" * 50)[0:60])
#
# ###################################### model initialization ###########################
# print("{:<50}".format("=" * 15 + " training... " + "=" * 50)[0:60])
# t1 = time.time()
# net_glob.train()
# # copy weights
# global_model = copy.deepcopy(net_glob.state_dict())
# local_m = []
# train_local_loss = []
# test_acc = []
# norm_med = []
# ####################################### run experiment ##########################
#
# # initialize data loader
# data_loader_list = []
# print(len(dict_users))
# for i in range(args.num_users):
# dataset = DatasetSplit(dataset_train, dict_users[i])
# ldr_train = DataLoader(dataset, batch_size=args.batch_size, shuffle=True)
# data_loader_list.append(ldr_train)
# ldr_train_public = DataLoader(val_set, batch_size=args.batch_size, shuffle=True)
#
# m = max(int(args.frac * args.num_users), 1)
# for t in range(args.round):
# args.local_lr = args.local_lr * args.decay_weight
# selected_idxs = list(np.random.choice(range(args.num_users), m, replace=False))
# num_selected_users = len(selected_idxs)
#
# ###################### local training : SGD for selected users ######################
# loss_locals = []
# local_updates = []
# delta_norms = []
# for i in selected_idxs:
# l_solver = LocalUpdate(args=args)
# net_glob.load_state_dict(global_model)
# # choose local solver
# if args.local_solver == 'local_sgd':
# new_model, loss = l_solver.local_sgd(
# net=copy.deepcopy(net_glob).to(args.device),
# ldr_train=data_loader_list[i])
# # compute local delta
# model_update = {k: new_model[k] - global_model[k] for k in global_model.keys()}
#
# # compute local model norm
# delta_norm = torch.norm(
# torch.cat([
# torch.flatten(model_update[k])
# for k in model_update.keys()
# ]))
# delta_norms.append(delta_norm)
#
# # clipping local model or not ? : no clip for cifar10
# # threshold = delta_norm / args.clip
# # if threshold > 1.0:
# # for k in model_update.keys():
# # model_update[k] = model_update[k] / threshold
#
# local_updates.append(model_update)
# loss_locals.append(loss)
# norm_med.append(torch.median(torch.stack(delta_norms)).cpu())
#
# ##################### communication: avg for all groups #######################
# model_update = {
# k: local_updates[0][k] * 0.0
# for k in local_updates[0].keys()
# }
# for i in range(num_selected_users):
# global_model = {
# k: global_model[k] + local_updates[i][k] / num_selected_users
# for k in global_model.keys()
# }
#
# ##################### testing on global model #######################
# net_glob.load_state_dict(global_model)
# net_glob.eval()
# test_acc_, _ = test_img(net_glob, dataset_test, args)
# test_acc.append(test_acc_)
# train_local_loss.append(sum(loss_locals) / len(loss_locals))
# # print('t {:3d}: '.format(t, ))
# print('t {:3d}: train_loss = {:.3f}, norm = {:.3f}, test_acc = {:.3f}'.
# format(t, train_local_loss[-1], norm_med[-1], test_acc[-1]))
#
# if math.isnan(train_local_loss[-1]) or train_local_loss[-1] > 1e8 or t == args.round - 1:
# np.savetxt(log_path + "_test_acc_repeat_" + str(args.repeat) + ".csv",
# test_acc,
# delimiter=",")
# np.savetxt(log_path + "_train_loss_repeat_" + str(args.repeat) + ".csv",
# train_local_loss,
# delimiter=",")
# np.savetxt(log_path + "_norm__repeat_" + str(args.repeat) + ".csv", norm_med, delimiter=",")
# break;
#
# t2 = time.time()
# hours, rem = divmod(t2 - t1, 3600)
# minutes, seconds = divmod(rem, 60)
# print("training time: {:0>2}:{:0>2}:{:05.2f}".format(int(hours), int(minutes), seconds))
#
# ##############################
# ## End of Fedml
# ###############################
except KeyboardInterrupt:
channel.unsubscribe(close)
exit()
def close(channel):
channel.close()
if __name__ == '__main__':
################################### hyperparameter setup ########################################
client_node()