-
Notifications
You must be signed in to change notification settings - Fork 55
/
mi_estimators.py
201 lines (160 loc) · 9.87 KB
/
mi_estimators.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
'''
Modified from: https://github.com/Linear95/CLUB
'''
import torch
import torch.nn as nn
class CLUB(nn.Module): # CLUB: Mutual Information Contrastive Learning Upper Bound
'''
This class provides the CLUB estimation to I(X,Y)
Method:
mi_est() : provides the estimation with input samples
loglikeli() : provides the log-likelihood of the approximation q(Y|X) with input samples
Arguments:
x_dim, y_dim : the dimensions of samples from X, Y respectively
hidden_size : the dimension of the hidden layer of the approximation network q(Y|X)
x_samples, y_samples : samples from X and Y, having shape [sample_size, x_dim/y_dim]
'''
def __init__(self, x_dim, y_dim, hidden_size):
super(CLUB, self).__init__()
# p_mu outputs mean of q(Y|X)
self.p_mu = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim))
# p_logvar outputs log of variance of q(Y|X)
self.p_logvar = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim),
nn.Tanh())
# self.p_logvar = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
# nn.ReLU(),
# nn.Linear(hidden_size//2, y_dim))
def get_mu_logvar(self, x_samples):
mu = self.p_mu(x_samples)
logvar = self.p_logvar(x_samples)
return mu, logvar
def mi_est(self, x_samples, y_samples):
mu, logvar = self.get_mu_logvar(x_samples)
# log of conditional probability of positive sample pairs
positive = - (mu - y_samples)**2 /2./logvar.exp()
prediction_1 = mu.unsqueeze(1) # shape [nsample,1,dim]
y_samples_1 = y_samples.unsqueeze(0) # shape [1,nsample,dim]
# log of conditional probability of negative sample pairs
negative = - ((y_samples_1 - prediction_1)**2).mean(dim=1)/2./logvar.exp()
return (positive.sum(dim = -1) - negative.sum(dim = -1)).mean()
def loglikeli(self, x_samples, y_samples): # unnormalized loglikelihood
mu, logvar = self.get_mu_logvar(x_samples)
return (-(mu - y_samples)**2 /logvar.exp()-logvar).sum(dim=1).mean(dim=0)
class CLUBSample(nn.Module): # Sampled version of the CLUB estimator
def __init__(self, x_dim, y_dim, hidden_size):
super(CLUBSample, self).__init__()
self.p_mu = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim))
self.p_logvar = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim),
nn.Tanh())
def get_mu_logvar(self, x_samples):
mu = self.p_mu(x_samples)
logvar = self.p_logvar(x_samples)
return mu, logvar
def loglikeli(self, x_samples, y_samples):
mu, logvar = self.get_mu_logvar(x_samples)
return (-(mu - y_samples)**2 /logvar.exp()-logvar).sum(dim=1).mean(dim=0)
def mi_est(self, x_samples, y_samples):
mu, logvar = self.get_mu_logvar(x_samples)
sample_size = x_samples.shape[0]
#random_index = torch.randint(sample_size, (sample_size,)).long()
random_index = torch.randperm(sample_size).long()
positive = - (mu - y_samples)**2 / logvar.exp()
negative = - (mu - y_samples[random_index])**2 / logvar.exp()
upper_bound = (positive.sum(dim = -1) - negative.sum(dim = -1)).mean()
return upper_bound/2.
class CLUBSample_reshape(nn.Module): # Sampled version of the CLUB estimator
def __init__(self, x_dim, y_dim, hidden_size):
super(CLUBSample_reshape, self).__init__()
self.p_mu = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim))
self.p_logvar = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim),
nn.Tanh())
def get_mu_logvar(self, x_samples):
mu = self.p_mu(x_samples)
logvar = self.p_logvar(x_samples)
return mu, logvar
def loglikeli(self, x_samples, y_samples):
mu, logvar = self.get_mu_logvar(x_samples)
mu = mu.reshape(-1, mu.shape[-1]) # (bs, y_dim) -> (bs, 1, y_dim) -> (bs, T, y_dim) -> (bs*T, y_dim)
logvar = logvar.reshape(-1, logvar.shape[-1])
y_samples = y_samples.reshape(-1, y_samples.shape[-1]) # (bs, T, y_dim) -> (bs*T, y_dim)
return (-(mu - y_samples)**2 /logvar.exp()-logvar).sum(dim=1).mean(dim=0)
def mi_est(self, x_samples, y_samples):
mu, logvar = self.get_mu_logvar(x_samples)
sample_size = mu.shape[0]
random_index = torch.randperm(sample_size).long()
y_shuffle = y_samples[random_index]
mu = mu.reshape(-1, mu.shape[-1]) # (bs, y_dim) -> (bs, 1, y_dim) -> (bs, T, y_dim) -> (bs*T, y_dim)
logvar = logvar.reshape(-1, logvar.shape[-1])
y_samples = y_samples.reshape(-1, y_samples.shape[-1]) # (bs, T, y_dim) -> (bs*T, y_dim)
y_shuffle = y_shuffle.reshape(-1, y_shuffle.shape[-1]) # (bs, T, y_dim) -> (bs*T, y_dim)
positive = - (mu - y_samples)**2 / logvar.exp()
negative = - (mu - y_shuffle)**2 / logvar.exp()
upper_bound = (positive.sum(dim = -1) - negative.sum(dim = -1)).mean()
return upper_bound/2.
class CLUBSample_group(nn.Module): # Sampled version of the CLUB estimator
def __init__(self, x_dim, y_dim, hidden_size):
super(CLUBSample_group, self).__init__()
self.p_mu = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim))
self.p_logvar = nn.Sequential(nn.Linear(x_dim, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, hidden_size//2),
nn.ReLU(),
nn.Linear(hidden_size//2, y_dim),
nn.Tanh())
def get_mu_logvar(self, x_samples):
mu = self.p_mu(x_samples)
logvar = self.p_logvar(x_samples)
return mu, logvar
def loglikeli(self, x_samples, y_samples): # unnormalized loglikelihood
mu, logvar = self.get_mu_logvar(x_samples) # mu/logvar: (bs, y_dim)
mu = mu.unsqueeze(1).expand(-1, y_samples.shape[1], -1).reshape(-1, mu.shape[-1]) # (bs, y_dim) -> (bs, 1, y_dim) -> (bs, T, y_dim) -> (bs*T, y_dim)
logvar = logvar.unsqueeze(1).expand(-1, y_samples.shape[1], -1).reshape(-1, logvar.shape[-1])
y_samples = y_samples.reshape(-1, y_samples.shape[-1]) # (bs, T, y_dim) -> (bs*T, y_dim)
return (-(mu - y_samples)**2 /logvar.exp()-logvar).sum(dim=1).mean(dim=0) / 2
def mi_est(self, x_samples, y_samples): # x_samples: (bs, x_dim); y_samples: (bs, T, y_dim)
mu, logvar = self.get_mu_logvar(x_samples)
sample_size = x_samples.shape[0]
#random_index = torch.randint(sample_size, (sample_size,)).long()
random_index = torch.randperm(sample_size).long()
# log of conditional probability of positive sample pairs
mu_exp1 = mu.unsqueeze(1).expand(-1, y_samples.shape[1], -1) # (bs, y_dim) -> (bs, T, y_dim)
# logvar_exp1 = logvar.unqueeze(1).expand(-1, y_samples.shape[1], -1).reshape(-1, logvar.shape[-1])
positive = - ((mu_exp1 - y_samples)**2).mean(dim=1) / logvar.exp() # mean along T
negative = - ((mu_exp1 - y_samples[random_index])**2).mean(dim=1) / logvar.exp() # mean along T
return (positive.sum(dim = -1) - negative.sum(dim = -1)).mean() / 2