-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
174 lines (153 loc) · 6.88 KB
/
model.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
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset, random_split
import numpy as np
import torch.nn.functional as F
class DenoisingAutoencoder(nn.Module):
def __init__(self, input_dim, hidden_dim):
super(DenoisingAutoencoder, self).__init__()
self.encoder = nn.Sequential(
nn.Linear(input_dim, hidden_dim),
nn.LeakyReLU()
)
self.decoder = nn.Sequential(
nn.Linear(hidden_dim, input_dim),
nn.LeakyReLU()
)
def forward(self, x):
encoded = self.encoder(x)
decoded = self.decoder(encoded)
return decoded
def encode(self, x):
return self.encoder(x)
class AEPretrain(nn.Module):
def __init__(self, encoder, decoder):
super().__init__()
self.encoder = encoder
self.decoder = decoder
def forward(self, x):
encoded = self.encoder(x)
decoded = self.decoder(encoded)
return encoded, decoded
class AE(nn.Module):
def __init__(self, input_dim, hidden_dims):
super().__init__()
self.encoder = nn.Sequential(
nn.Linear(input_dim, hidden_dims[0]),
nn.LeakyReLU(),
nn.Linear(hidden_dims[0], hidden_dims[1]),
nn.LeakyReLU(),
)
self.decoder = nn.Sequential(
nn.Linear(hidden_dims[1], hidden_dims[0]),
nn.LeakyReLU(),
nn.Linear(hidden_dims[0], input_dim),
nn.LeakyReLU(),
)
def forward(self, x):
encoded = self.encoder(x)
decoded = self.decoder(encoded)
return encoded, decoded
class AnomalyScore(nn.Module):
def __init__(self,hidden_dim=8,recon_dim=1,sim_dim=1):
super().__init__()
D=hidden_dim+recon_dim+sim_dim
self.regression=nn.Sequential(
nn.Linear(D,16),
nn.LeakyReLU(),
nn.Linear(16, 1),
nn.Sigmoid(),
)
self.sigmoid = nn.Sigmoid()
def forward(self, x_hidden, x_recon,loss_anchor):
x=torch.cat((x_hidden,x_recon,loss_anchor),1)
score=self.regression(x)
return score
class MNP(nn.Module):
def __init__(self,input_dim, hidden_dims,loss_type='smooth',device='cuda',PretrainAE=None,multinormalprototypes=None,n_prototypes=2,AS_type='AnomalyScore',alpha=1,):
super().__init__()
if multinormalprototypes is None:
self.multinormalprototypes=nn.parameter.Parameter(torch.rand(n_prototypes,hidden_dims[-1]))
self.AE=AE(input_dim=input_dim,hidden_dims=hidden_dims)
else:
self.multinormalprototypes=nn.parameter.Parameter(multinormalprototypes)
self.AE=PretrainAE
self.AS=eval(AS_type)(hidden_dims[-1])
self.device=device
self.alpha=alpha
if loss_type == 'mse':
self.loss_reg = torch.nn.MSELoss(reduction='none')
elif loss_type == 'mae':
self.loss_reg = torch.nn.MAELoss(reduction='none')
elif loss_type == 'smooth':
self.loss_reg = torch.nn.SmoothL1Loss(
reduction='none',
# beta=beta
)
else:
raise ValueError('unsupported loss')
def forward(self, x,output_hidden=False):
x_encoded,x_decoded=self.AE(x)
residual_recon=self.loss_reg(x,x_decoded)
loss_recon=residual_recon.mean(dim=1)
anchors=F.normalize(self.multinormalprototypes,dim=1)
x_encoded=F.normalize(x_encoded,dim=1)
anchors_tmp=anchors.unsqueeze(0)
x_encoded_tmp=x_encoded.unsqueeze(1)
dis_anchor=torch.sum((x_encoded_tmp-anchors_tmp)**2,dim=2)
sim_anchor=(1+dis_anchor/self.alpha)**(-(self.alpha+1)/2)
sim_anchor_max,_=torch.max(sim_anchor,dim=1)
score=self.AS(x_encoded,loss_recon.unsqueeze(1),sim_anchor_max.unsqueeze(1))
if output_hidden:
return loss_recon,score,sim_anchor,x_encoded,anchors
return loss_recon,score,sim_anchor
class Loss_MNP(torch.nn.Module):
def __init__(self,score_loss='mse', device='cuda',T=2,m1=0.02,lambda_kl=1,beta=1):
super(Loss_MNP, self).__init__()
self.device=device
self.T = T
self.m1=m1
self.lambda_kl=lambda_kl
self.beta=beta
if score_loss == 'mse':
self.loss_reg = torch.nn.MSELoss(reduction='none')
elif score_loss == 'mae':
self.loss_reg = torch.nn.MAELoss(reduction='none')
elif score_loss == 'smooth':
self.loss_reg = torch.nn.SmoothL1Loss(
reduction='none',
# beta=beta
)
else:
raise ValueError('unsupported loss')
self.sigmoid = nn.Sigmoid()
self.relu=nn.ReLU()
def forward(self, basenet,x_pos,x_neg,pre_recon_loss, pre_score_loss,pre_anchor_loss):
recon_pos,score_pos,sim_anchor_pos=basenet(x_pos)
recon_neg,score_neg,sim_anchor_neg=basenet(x_neg)
neg_weight=self.sigmoid(self.beta*torch.max(sim_anchor_neg,dim=1)[0])
neg_weight=(neg_weight/neg_weight.sum())*neg_weight.shape[0]
loss_recon_neg=(recon_neg*neg_weight).mean()
loss_recon_pos=self.relu(-recon_pos+loss_recon_neg+self.m1).mean()
loss_recon=loss_recon_neg+loss_recon_pos
loss_score_pos=self.loss_reg(score_pos,torch.ones(score_pos.shape).to(self.device)) # 1
loss_score_neg=self.loss_reg(score_neg,torch.zeros(score_neg.shape).to(self.device)) # 0
loss_score=loss_score_pos.mean()+(loss_score_neg*neg_weight).mean()
rsim_anchor_neg=sim_anchor_neg/torch.sum(sim_anchor_neg,dim=1,keepdim=True)
f_anchors=torch.sum(rsim_anchor_neg,dim=0,keepdim=True)
target_d=torch.square(rsim_anchor_neg)/f_anchors
rtarget_d=target_d/torch.sum(target_d,dim=1,keepdim=True)
loss_kl=torch.sum(rtarget_d * (torch.log(rtarget_d) - torch.log(rsim_anchor_neg)))
max_sim_anchor_pos_mean=(torch.max(sim_anchor_pos,dim=1)[0]).mean()
max_sim_anchor_neg_mean=(neg_weight*(torch.max(sim_anchor_neg,dim=1)[0])).mean()
loss_anchor=-torch.log(self.sigmoid(-max_sim_anchor_pos_mean+max_sim_anchor_neg_mean))
loss_anchor+=self.lambda_kl*loss_kl
#dynamic weight averaging
k1 = torch.exp((loss_recon / pre_recon_loss) / self.T) if pre_recon_loss != 0 else 0
k2 = torch.exp((loss_score / pre_score_loss) / self.T) if pre_score_loss != 0 else 0
k3 = torch.exp((loss_anchor / pre_anchor_loss) / self.T) if pre_anchor_loss != 0 else 0
loss = (k1 / (k1 + k2+k3)) * loss_recon + (k2 / (k1 + k2+k3)) * loss_score+(k3 / (k1 + k2+k3)) * loss_anchor
return loss,(loss_recon,loss_recon_pos,loss_recon_neg),\
(loss_score,loss_score_pos.mean(),(loss_score_neg*neg_weight).mean()),\
(loss_anchor,max_sim_anchor_pos_mean,max_sim_anchor_neg_mean)