-
Notifications
You must be signed in to change notification settings - Fork 0
/
fl_only.py
executable file
·311 lines (227 loc) · 10.1 KB
/
fl_only.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
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
# -*- coding: utf-8 -*-
"""
Created on Tue Nov 10 11:31:43 2020
@author: henry
"""
import numpy as np
import os
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
from copy import deepcopy
import random
import pickle
from Shared_ML_code.neural_nets import MLP, CNN, FedAvg, FPAvg, LocalUpdate, \
LocalUpdate_PFL, FedAvg2, LocalUpdate_FO_PFL, LocalUpdate_HF_PFL
from Shared_ML_code.testing import test_img, test_img2
# %% import neural network data
# seed declarations
init_seed = 1
random.seed(init_seed)
np.random.seed(init_seed)
torch.manual_seed(init_seed)
# data import and device spec
trans_mnist = transforms.Compose([transforms.ToTensor(), \
transforms.Normalize((0.1307,),(0.3081,))])
dataset_train = torchvision.datasets.MNIST('./data/mnist/',train=True,download=False,\
transform=trans_mnist)
dataset_test = torchvision.datasets.MNIST('./data/mnist/',train=False,download=False,\
transform=trans_mnist)
device = torch.device('cuda')
#device = torch.device('cpu')
# %% filtering the ML data
# label split
train = {i: [] for i in range(10)}
for index, (pixels,label) in enumerate(dataset_train):
train[label].append(index)
test = {i: [] for i in range(10)}
for index, (pixels,label) in enumerate(dataset_test):
test[label].append(index)
data_source = 'mnist' # delete once argparse is configured
# assign datasets to nodes
clusters = 3
swarms = 3
swarm_period = 2#5
global_period = 2
cycles = 20
total_time = swarm_period*global_period*cycles
nodes_per_cluster = [np.random.randint(2,6) for i in range(swarms)]
#labels_per_node = [np.random.randint(1,6) for i in range(nodes)] #number of labels changes over time
#labels_set = {i: [] for i in range(nodes)} #randomly determined based on labels_per_node
# labels_per_node (i.e., distribution) changes over time...
# static_lpc = [np.random.randint(2,5) for i in range(swarms)] #static qty of labels per node
static_lpc = [8 for i in range(swarms)]
static_ls = {i: [] for i in range(swarms)} # actual labels at each node
# variable lpn and ls have rows: time, cols: nodes
var_lpc = np.zeros((total_time,swarms))
for i in range(total_time):
var_lpc[i,:] = [np.random.randint(2,5) for i in range(swarms)]
var_ls = {j: {i: [] for i in range(swarms)} for j in range(total_time)}
## TODO: epsilon based changes in distribution - calc KL divergence for that
# %% populating ML label holders
## TODO: epsilon based changes - see KL divergence
def pop_labels(temp_lpn,temp_ls,max_labels=10):
for i,j in enumerate(temp_lpn):
j = int(j)
temp_ls[i] = sorted(random.sample(range(max_labels),j))
return temp_ls
# pop holders
static_ls = pop_labels(static_lpc,static_ls)
for i in range(total_time):
var_ls[i] = pop_labels(var_lpc[i,:],var_ls[i])
# random data qty per label
avg_qty = int(len(dataset_train)/(sum(nodes_per_cluster))) #*total_time))
# need to determine data per device and total data per swarm
static_qty = np.random.normal(avg_qty,avg_qty/10,size=(sum(nodes_per_cluster))).astype(int)
static_data_per_swarm = []
counter = 0
for i,j in enumerate(nodes_per_cluster):
static_data_per_swarm.append(0)
for k in range(j):
static_data_per_swarm[i] += static_qty[counter]
counter += 1
var_qty = [np.random.normal(avg_qty,avg_qty/10,size=(sum(nodes_per_cluster))).astype(int) \
for j in range(total_time)]
# calculate training datasets per node
static_nts = {i: [] for i in range(sum(nodes_per_cluster))} #static_node_train_sets
var_nts = {j:{i:[] for i in range(sum(nodes_per_cluster))} for j in range(total_time)}
def pop_nts(temp_ls,temp_qty,temp_nts,npc,train=train): #nts - node training set
counter = 0
for ind_t_cluster, t_cluster in enumerate(npc): #npc = [3,5,2], nodes per cluster
temp_ls_inner = temp_ls[ind_t_cluster]
for i in range(t_cluster):
for curr_label in temp_ls_inner:
temp_nts[counter] += random.sample(train[curr_label],\
int(temp_qty[counter]/len(temp_ls_inner)))
counter += 1
return temp_nts
static_nts = pop_nts(static_ls,static_qty,static_nts,nodes_per_cluster)
## no variable time at the moment, get static_nts working first
# for j in range(total_time):
# var_nts[j] = pop_nts(var_ls[j],var_qty[j],var_nts[j],nodes_per_cluster)
# # saving the data
# cwd = os.getcwd()
# with open(cwd+'/data/'+str(init_seed)+data_source+str(nodes)+'_lpn','wb') as f:
# pickle.dump()
# %% personalize the testing dataset
## basically just sort the testing dataset into indexes for each cluster
cluster_test_sets = {i:[] for i in range(clusters)} #indexed by cluster
for i in range(clusters):
cluster_ls = static_ls[i]
for j in cluster_ls:
cluster_test_sets[i] += test[j]
# %% create neural networks
## setup FL
d_in = np.prod(dataset_train[0][0].shape)
d_h = 64
d_out = 10
global_net = MLP(d_in,d_h,d_out).to(device)
print(global_net)
global_net.train()
# default_w = deepcopy(global_net.state_dict())
cwd = os.getcwd()
with open(cwd+'/data/default_w','rb') as f:
default_w = pickle.load(f)
# one central model object for each swarm
fl_swarm_models = [MLP(d_in,d_h,d_out).to(device) for i in range(swarms)]
for i in fl_swarm_models:
i.load_state_dict(default_w)
i.train()
## setup PFL - same as FL, just an additional object
pfl_swarm_models = [MLP(d_in,d_h,d_out).to(device) for i in range(swarms)]
for i in pfl_swarm_models:
i.load_state_dict(default_w)
i.train()
FO_hn_pfl_swarm_models = [MLP(d_in,d_h,d_out).to(device) for i in range(swarms)]
for i in FO_hn_pfl_swarm_models:
i.load_state_dict(default_w)
i.train()
HF_hn_pfl_swarm_models = [MLP(d_in,d_h,d_out).to(device) for i in range(swarms)]
for i in HF_hn_pfl_swarm_models:
i.load_state_dict(default_w)
i.train()
## ovr ML params setup
lr = 1e-3
# lr2 = 1e-3
# %% running for all time
fl_acc = []
hn_pfl_acc = []
FO_hn_pfl_acc = []
HF_hn_pfl_acc = []
for t in range(total_time):
swarm_w = {i:[] for i in range(swarms)}
# data_processed = {i:0 for i in range(swarms)}
#### Hierarchical-FL procedure
### 1. create object for each node/device
### 2. after \tau1 = swarm_period iterations, aggregate cluster-wise (weighed)
### 3. after \tau2 = global_period swarm-wide aggregations, aggregate globally (weighted again)
print('iteration:{}'.format(t))
print('hierarchical FL begins here')
uav_counter = 0
for ind_i,val_i in enumerate(nodes_per_cluster):
for j in range(val_i): # each uav in i
local_obj = LocalUpdate(device,bs=10,lr=lr,epochs=swarm_period,\
dataset=dataset_train,indexes=static_nts[uav_counter])
_,w,loss = local_obj.train(net=deepcopy(fl_swarm_models[ind_i]).to(device))
swarm_w[ind_i].append(w)
uav_counter += 1
if (t+1) % (swarm_period*global_period) == 0:
## then a swarm-wide agg followed immediately by a global
# swarm-wide agg
t_static_qty = deepcopy(static_qty).tolist()
t_swarm_total_qty = []
w_swarms = []
for ind_i,val_i in enumerate(nodes_per_cluster):
t2_static_qty = t_static_qty[:val_i]
del t_static_qty[:val_i]
t3_static_qty = [i*swarm_period for i in t2_static_qty]
w_avg_swarm = FedAvg2(swarm_w[ind_i],t3_static_qty)
t_swarm_total_qty.append(sum(t3_static_qty))
w_swarms.append(w_avg_swarm)
# global agg
w_global = FedAvg2(w_swarms,t_swarm_total_qty)
for i in fl_swarm_models:
i.load_state_dict(w_global)
i.train()
else:
## run FL swarm-wide aggregation only
t_static_qty = deepcopy(static_qty).tolist()
for ind_i,val_i in enumerate(nodes_per_cluster):
t2_static_qty = t_static_qty[:val_i]
del t_static_qty[:val_i]
t3_static_qty = [i*swarm_period for i in t2_static_qty]
w_avg_swarm = FedAvg2(swarm_w[ind_i],t3_static_qty)
fl_swarm_models[ind_i].load_state_dict(w_avg_swarm)
fl_swarm_models[ind_i].train()
## for clarity, splitting this outside of the other if-else statement
## evaluate model performance
if (t+1) % (swarm_period*global_period) == 0:
fl_acc_temp = 0
for i,ii in enumerate(fl_swarm_models):
ii.eval()
fl_acc_temp += test_img2(ii,dataset_test,bs=10,\
indexes=cluster_test_sets[i],device=device)[0] * static_data_per_swarm[i] \
/ sum(static_data_per_swarm)
# fl_acc.append(fl_acc_temp/len(fl_swarm_models))
fl_acc.append(fl_acc_temp)
# print(fl_acc_temp)
print(fl_acc[-1])
# %% saving results
cwd = os.getcwd()
with open(cwd+'/data/fl_acc_test_base','wb') as f:
pickle.dump(fl_acc,f)
# %% graveyard
#### HN-PFL procedure
### 1. create object for each node/device
### 2. after \tau1 = swarm_period iterations, aggregate cluster-wise (weighed)
### 3. after \tau2 = global_period swarm-wide aggregations, aggregate globally (unweighted)
# print('HN-PFL begins here')
# uav_counter = 0
# for ind_i,val_i in enumerate(nodes_per_cluster):
# for j in range(val_i): # each uav in i
# local_obj = LocalUpdate_PFL(device,bs=10,lr=lr,epochs=swarm_period,\
# dataset=dataset_train,indexes=static_nts[uav_counter])
# _,w,loss = local_obj.train(net=deepcopy(pfl_swarm_models[ind_i]).to(device))
# swarm_w[ind_i].append(w)
# uav_counter += 1