-
Notifications
You must be signed in to change notification settings - Fork 1
/
avgm_consensus.py
160 lines (148 loc) · 5.7 KB
/
avgm_consensus.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
import torch
import random
import numpy as np
import torch.nn.functional as F
import copy
from matplotlib import pyplot as plt
import matplotlib
import wandb
n = 64
T = 1000
x_size = 200000
scaling = 0.1
scalingm = 0.01
W = torch.zeros(n,n).to('cuda')
I = torch.eye(n).to("cuda")
W[0,0] = W[n-1,n-1] = 1.0/3.0
W[0,1] = W[0,n-1] = W[n-1,0] = W[n-1,n-2] = 1.0/3.0
for i in range(1,n-1):
for j in [i-1, i, i+1]:
if i==j: W[i,j] = 1.0/3.0
else: W[i,j] = 1.0/3.0
#print(W, W1)
Error_sg = []
Error_gut = []
Error_qgm = []
Error_gutm = []
#print(W.size(), x_init.size())
for i in range(0,2):
x_init = -2 * torch.rand(x_size, n) + 1.0
x_init = x_init.to('cuda')
x_final = torch.mean(x_init, dim=1).view(-1,1)
x_final = x_final.expand(x_size, n)
#no momentum
x_t = copy.deepcopy(x_init)
x_g = copy.deepcopy(x_init)
x_prev = copy.deepcopy(x_init)
d_prev = torch.zeros_like(x_prev)
#momentum
x_q = copy.deepcopy(x_init)
x_gm = copy.deepcopy(x_init)
x_prevm = copy.deepcopy(x_init)
d_prevm = torch.zeros_like(x_prevm)
m_buff = torch.zeros_like(x_prevm)
mg_buff = torch.zeros_like(x_prevm)
shift = torch.zeros_like(x_prevm)
error_sg = []
error_gut = []
error_qgm = []
error_gutm = []
Time = []
for t in range(0,T):
#simple gossip update
d = torch.matmul(x_t, W-I)
x_t.data.add_(d.data)
####
#update using tracking method (GUT)
dg = torch.matmul(x_g, W-I)
dg_copy = copy.deepcopy(dg)
xg_copy = copy.deepcopy(x_g)
bias = torch.matmul(x_g - x_prev, W) - d_prev + torch.matmul(x_g, W-I)
x_g.data.add_(dg.data)
x_g.data.add_(bias.data, alpha=scaling)
d_prev.data.copy_(dg_copy.data)
x_prev.data.copy_(xg_copy.data)
#gossip update with qgm
dm = torch.matmul(x_q, W-I)
dm.data.add_(torch.matmul(m_buff, W), alpha=0.9)
x_q.data.add_(dm.data)
m_buff.data.mul_(0.9).add_(dm.data, alpha=0.1)
####
#update using tracking method with qgm (GUTm)
dg = torch.matmul(x_gm, W-I)
dg_copy = copy.deepcopy(dg)
xg_copy = copy.deepcopy(x_gm)
#shift.data.mul_(0.9).add_(torch.matmul(x_gm, (W-I)), alpha=0.1)
bias = torch.matmul(x_gm - x_prevm, W) - d_prevm + torch.matmul(x_gm, (W-I))
u = dg.data + (scalingm*bias.data)
u.data.add_(torch.matmul(mg_buff, W), alpha=0.9)
x_gm.data.add_(u.data)
mg_buff.data.mul_(0.9).add_(u.data, alpha=0.1)
d_prevm.data.copy_(dg_copy.data)
x_prevm.data.copy_(xg_copy.data)
#compute the error without momentum
x_copy = copy.deepcopy(x_t)
xg_copy = copy.deepcopy(x_g)
x_copy.data.add_(x_final.data, alpha=-1.0)
xg_copy.data.add_(x_final.data, alpha=-1.0)
x_copy.square_()
xg_copy.square_()
e = (1.0/n)*torch.sum(x_copy.data)
e1 = (1.0/n)*torch.sum(xg_copy.data)
#print("Average consensus error at iteration %d: %.4f, %.4f"%(t, e, e1))
error_sg.append(e.cpu().numpy())
error_gut.append(e1.cpu().numpy())
#compute the error with momentum
x_copy = copy.deepcopy(x_q)
xg_copy = copy.deepcopy(x_gm)
x_copy.data.add_(x_final.data, alpha=-1.0)
xg_copy.data.add_(x_final.data, alpha=-1.0)
x_copy.square_()
xg_copy.square_()
e = (1.0/n)*torch.sum(x_copy.data)
e1 = (1.0/n)*torch.sum(xg_copy.data)
#print("Average consensus error at iteration %d: %.4f, %.4f"%(t, e, e1))
error_qgm.append(e.cpu().numpy())
error_gutm.append(e1.cpu().numpy())
Time.append(t)
Error_sg.append(error_sg)
Error_gut.append(error_gut)
Error_qgm.append(error_qgm)
Error_gutm.append(error_gutm)
Error_sg = np.array(Error_sg)
Error_gut = np.array(Error_gut)
Error_sg_mean = np.mean(Error_sg, axis=0)
Error_gut_mean = np.mean(Error_gut, axis=0)
Error_sg_std = np.std(Error_sg, axis=0)
Error_gut_std = np.std(Error_gut, axis=0)
Error_qgm = np.array(Error_qgm)
Error_gutm = np.array(Error_gutm)
Error_qgm_mean = np.mean(Error_qgm, axis=0)
Error_gutm_mean = np.mean(Error_gutm, axis=0)
Error_qgm_std = np.std(Error_qgm, axis=0)
Error_gutm_std = np.std(Error_gutm, axis=0)
fig,ax = plt.subplots()
#fig.set_size_inches(10,8)
plt.box(False)
ax.plot(Time, Error_sg_mean, label = "simple gossip", linewidth=1.2, color='red', linestyle='dashed')
ax.fill_between(Time, (Error_sg_mean-Error_sg_std), (Error_sg_mean+Error_sg_std), alpha=0.1, color='red')
ax.plot(Time, Error_gut_mean, label = "gossip with GUT", linewidth=1.2, color='dodgerblue', linestyle='dashed')
ax.fill_between(Time, (Error_gut_mean-Error_gut_std), (Error_gut_mean+Error_gut_std), alpha=0.1, color='dodgerblue')
ax.plot(Time, Error_qgm_mean, label = "gossip with QGM", linewidth=1.2, color='red')
ax.fill_between(Time, (Error_qgm_mean-Error_qgm_std), (Error_qgm_mean+Error_qgm_std), alpha=0.1, color='red')
ax.plot(Time, Error_gutm_mean, label = "gossip with QG-GUTm", linewidth=1.2, color='dodgerblue')
ax.fill_between(Time, (Error_gutm_mean-Error_gutm_std), (Error_gutm_mean+Error_gutm_std), alpha=0.1, color='dodgerblue')
#ax.plot(Time, Error_gut_mean, label = "gossip with GUT")
#fig.patch.set_visible(False)
matplotlib.pyplot.yscale("log")
matplotlib.pyplot.grid(True, color='0.8')
ax.set_xlabel("Iteration", fontsize=18)
ax.set_ylabel("Consensus Error", fontsize=18)
#ax.set_title("Average consensus on 64 nodes ring")
plt.xticks(fontsize=15)
plt.yticks(fontsize=15)
ax.set_xticks([0, 250, 500, 750, 1000])
plt.tight_layout()
plt.legend(fontsize=13)
plt.savefig("ac_64.pdf")
plt.show()