-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutil.py
175 lines (130 loc) · 4.32 KB
/
util.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
import torch
import math
import numpy as np
from torch.distributions.normal import Normal
device = 'cuda'
def temp_calc(args, temp):
param_factor = 4
means, sigmas, weights = args
if temp == 0.05:
temps = []
for i in range(60):
if i <= 3:
temperature = 0.05
elif i >= 8:
temperature = 0.5
else:
temperature = 0.05 + (i - 3) * 0.09
temps.append(temperature)
temps = torch.FloatTensor(temps).to('cuda')
temps = temps.view(means[0].size())
temp_factor = torch.sqrt(temps)
else:
temps = temp
temp_factor = np.sqrt(temps)
alt_means = 0
for k in range(param_factor):
alt_means = alt_means + weights[k] * means[k]
for k in range(param_factor):
means[k] = means[k] + (alt_means - means[k]) * (1 - temps)
sigmas[k] = sigmas[k] * temp_factor
return means, sigmas, weights
def cal_para(out, temperature):
sqrt = math.sqrt
cgm_factor = 4
r_u = 1.6
r_s = 1.1
r_w = 1 / 1.75
out = out.permute(0, 2, 1).contiguous()
out = out.view(out.shape[0], out.shape[1], -1, cgm_factor)
a0 = out[:, :, :, 0]
a1 = out[:, :, :, 1]
a2 = out[:, :, :, 2]
a3 = out[:, :, :, 3]
xi = 2 * torch.sigmoid(a0) - 1
omega = torch.exp(4 * torch.sigmoid(a1)) * 2 / 255
alpha = 2 * torch.sigmoid(a2) - 1
beta = 2 * torch.sigmoid(a3)
# cal temperature
use_t = False
if temperature != 0:
use_t = True
tempers = []
for i in range(xi.shape[-1]):
if i<=3:
temper = 0.05
elif i>=8:
temper = 0.5
else:
temper = 0.05 + (i-3)*0.09
tempers.append(temper)
#tempers = tempers[::-1]
tempers = torch.Tensor(tempers)
tempers = tempers.expand(xi.shape).to(device)
# if temperature != 0.01 mean it is for harmonic so it will be piecewise linear
if temperature != 0.01:
temperature = tempers
sqrt = torch.sqrt
# end cal temperature
sigmas = []
for k in range(cgm_factor):
sigma = omega * torch.exp(k * (torch.abs(alpha) * r_s - 1))
sigmas.append(sigma)
mus = []
for k in range(cgm_factor):
temp_sum = 0
for i in range(k):
temp_sum += sigmas[i] * r_u * alpha
mu = xi + temp_sum
mus.append(mu)
ws = []
temp_sum = 0
for i in range(cgm_factor):
temp_sum += alpha.pow(2 * i) * beta.pow(i) * (r_w ** i)
for k in range(cgm_factor):
w = (alpha.pow(2 * k) * beta.pow(k) * (r_w ** k)) / temp_sum
ws.append(w)
if use_t:
_mus = 0
for k in range(cgm_factor):
_mus += ws[k]*mus[k]
for k in range(cgm_factor):
mus[k] = mus[k] + (_mus - mus[k])*(1 - temperature)
sigmas[k] *= sqrt(temperature)
return sigmas, mus, ws
def CGM_loss(out, y):
y = y.permute(0, 2, 1)
sigmas, mus, ws = cal_para(out, 0)
#print(torch.mean(sigmas[0]))
# 验证w之和是1
sum = 0
for k in range(4):
tw = ws[k].view(-1)
sum += tw
# alternative： torch.distributions.normal.Normal
probs = 0
for k in range(4):
dist = Normal(mus[k].to(device), sigmas[k].to(device))
log_prob = dist.log_prob(y.to(device))
x = dist.sample()
# prob = log_prob * log_prob
probs += ws[k] * log_prob
return -torch.mean(probs)
def sample_from_CGM(out, mydevice, temperature=0.01):
#temperature = 0.01
out = out.unsqueeze(1)
out = out.unsqueeze(0)
sigmas, mus, ws = cal_para(out, temperature)
value = 0
rand = torch.rand(ws[0].shape).to(device)
for k in range(4):
mask_btm = torch.zeros(ws[k].shape).to(device)
for i in range(k):
mask_btm += ws[i]
mask = (rand < (ws[k] + mask_btm)) * (rand >= mask_btm)
mask = mask.float()
gaussian_dist = Normal(loc=mus[k], scale=sigmas[k])
x = gaussian_dist.sample()
value += mask * x
# value shape (batch, length, channel'60')
return value