-
Notifications
You must be signed in to change notification settings - Fork 1
/
util.py
258 lines (207 loc) · 9.6 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
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
import torch
from torchvision import datasets, transforms
import numpy as np
import argparse
import torch
import torch.utils.data
from torch import nn, optim
from torch.nn import functional as F
from torchvision.utils import save_image
import matplotlib.pyplot as plt
import numpy as np
import scipy.stats
import math
def str2bool(v):
if isinstance(v, bool):
return v
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
def shift_image(x,y,width_shift_val,height_shift_val):
#x is assumed to be a tensor of shape (#batch_size, #channels, width, height) = (#batch size, 1, 28, 28)
#y is assumed to be a tensor of shape (#batch_size)
batch_size = x.size()[0]
shift_aa = transforms.RandomAffine(degrees=0,translate=(width_shift_val,height_shift_val))
x_return = shift_aa(x)
return x_return,y
def log_gaussian_torch(x,mean,var):
std = torch.sqrt(var)
return -torch.log(std*math.sqrt(2*math.pi)+1e-4) - 0.5*((x-mean)/(std+1e-4))**2
def accuracy(y_true, y_pred):
#y_true and y_pred is assumed to be numpy array.
y_true = y_true.reshape(-1).astype(int)
y_pred = y_pred.reshape(-1).astype(int)
correct = 0
for i in range(0,y_true.shape[0]):
if y_true[i] == y_pred[i]:
correct = correct + 1
return correct/y_true.shape[0]
def pred(x,model,device):
#x is assumes to be already in device
batch_size = x.size()[0]
yc = np.zeros((10,batch_size))
for i in range(0,10):
y = i*torch.ones(batch_size).type(torch.int64).to(device)
#print(y)
#x_recon, mu_q1, logvar_q1, mu_q2, logvar_q2 = model(x,y, manipulated=False)
#m~q(m|x) for each batch
mu_q2, logvar_q2 = model.encode2(x)
m = model.reparameterize(mu_q2, logvar_q2)
#m = torch.zeros(x.size()[0], 32)
#calculate the log-ration term for each y = c(as an approximation to log p(x,y=c))
sum = 0
K = 100
for j in range(0,K):
mu_q1, logvar_q1 = model.encode1(x,model.onehot(y),m)
z = model.reparameterize(mu_q1, logvar_q1)
#sum = sum + pred_logRatio(x.to(device), x_recon.to(device), mu_q1.to(device), logvar_q1.to(device), m.to(device))
sum = sum + pred_logRatio(x, y, z, m, mu_q1, logvar_q1, model, device)
log_pxy = torch.log(sum).view(x.size()[0]).detach().cpu().numpy()
#log_pxy = torch.log(sum+1e-4).view(x.size()[0]).detach().cpu().numpy()
yc[i] = log_pxy
#print(yc)
exp_yc = np.exp(yc)
#print(exp_yc)
sum_exp_yc = np.sum(exp_yc,axis=0)
for i in range(0,batch_size):
for j in range(0,10):
exp_yc[j][i] = exp_yc[j][i]/sum_exp_yc[i]
label = np.argmax(exp_yc,axis=0)
return label
def pred_logRatio(x, y, z_sampled, m_sampled, mu_q1, logvar_q1, model, device):
batch_size = x.size()[0]
#sample from z~q(z|x,y,m)
#calculate log q(z|x,y,m)
temp = log_gaussian_torch(z_sampled,mu_q1,torch.exp(logvar_q1))
log_q1 = torch.sum(temp, dim =1)
#calculate log p(x|y,z,m)
#log_pxyzm = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x.view(-1,784)+1e-4)), dim=1)
x_recon = model.decode(model.onehot(y),z_sampled,m_sampled)
log_pxyzm = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x_recon.view(-1,784)+1e-4)), dim=1)
#temp_print = F.binary_cross_entropy(x_recon.view(-1,784),x.view(-1,784),reduction = 'sum')
#calculate p(y)
py = 0.1
#calculate log p(z)
zero_tensor = torch.zeros(z_sampled.size()).to(device)
one_tensor = torch.ones(z_sampled.size()).to(device)
temp = log_gaussian_torch(z_sampled,zero_tensor,one_tensor)
log_pz = torch.sum(temp, dim = 1)
#adding a constant at the end to prevent underflow. The term will not affect the overall calculation due to the softmax.
underflow_const = 100
s = log_pxyzm.reshape(batch_size) + math.log(py) + log_pz.reshape(batch_size) - log_q1.reshape(batch_size) + underflow_const
return torch.exp(s)
def ELBO_x(x,model,device):
#Calculates ELBO(x)
#with torch.no_grad():
x_device = x.to(device)
underflow_const = 600
sum = torch.zeros(x.size()[0]).type(torch.DoubleTensor).to(device)
for i in range(0,10):
yc = i*torch.ones(x.size()[0]).type(torch.int64).to(device)
a=ELBO_xy(x_device,yc,model,device).type(torch.DoubleTensor)
#print(a)
sum = sum + torch.exp(underflow_const+a).to(device)
#print('here')
#print(sum)
#print(torch.log(sum))
return_val = (torch.log(sum) - underflow_const).type(torch.float32)
#print(return_val)
return return_val
def ELBO_xy(x, y, model,device):
#Calculates ELBO(x,y)
#x and y should already be in device.
#x_recon, mu_q1, logvar_q1, mu_q2, logvar_q2 = model(x,y,manipulated=True)
#with torch.no_grad():
#calculate E_q(z,m|x,y)[(log p(x|y,z,m))]
mu_q2, logvar_q2, m_sampled = model(x=x,phase=2)
#print(m_sampled)
#BCE = torch.zeros(x.size()[0]).to(device)
BCE = torch.zeros(x.size()[0]).to(device)
K = 1
#mu_q1, logvar_q1 = model_cpu.encode1(x_cpu,model_cpu.onehot(y_cpu),m_sampled)
for j in range(0,K):
mu_q1, logvar_q1, z_sampled = model(x=x,y=y,m_sampled=m_sampled,phase=1)
x_recon = model(x=x,y=y,z_sampled=z_sampled,m_sampled=m_sampled,phase=3)
log_pxyzm = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x_recon.view(-1,784)+1e-4)), dim=1)
BCE = BCE+ log_pxyzm
BCE = (1/K) * BCE
#KL(q(z,m|x,y))
logvar_cat = torch.cat((logvar_q1, logvar_q2), dim = 1)
mu_cat = torch.cat((mu_q1, mu_q2), dim = 1)
KLD = 0.5 * torch.sum(1 + logvar_cat - mu_cat.pow(2) - logvar_cat.exp(), dim=1)
#print(KLD.size())
#p(y)
py = 0.1
return math.log(py) + BCE + KLD
def ELBO_xy_hp(x, y, model,device):
#Calculates ELBO(x,y)
#x and y should already be in device.
#x_recon, mu_q1, logvar_q1, mu_q2, logvar_q2 = model(x,y,manipulated=True)
#with torch.no_grad():
#calculate E_q(z,m|x,y)[(log p(x|y,z,m))]
#x_device = x.to(device)
mu_q2, logvar_q2, m_sampled = model(x=x,phase=2)
#print(m_sampled)
#BCE = torch.zeros(x.size()[0]).to(device)
BCE = torch.zeros(x.size()[0]).type(torch.DoubleTensor).to(device)
K = 1
#mu_q1, logvar_q1 = model_cpu.encode1(x_cpu,model_cpu.onehot(y_cpu),m_sampled)
for j in range(0,K):
mu_q1, logvar_q1, z_sampled = model(x=x,y=y,m_sampled=m_sampled,phase=1)
x_recon = model(x=x,y=y,z_sampled=z_sampled,m_sampled=m_sampled,phase=3)
log_pxyzm = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x_recon.view(-1,784)+1e-4)), dim=1)
BCE = BCE+ log_pxyzm
BCE = (1/K) * BCE
#KL(q(z,m|x,y))
logvar_cat = torch.cat((logvar_q1, logvar_q2), dim = 1)
mu_cat = torch.cat((mu_q1, mu_q2), dim = 1)
KLD = 0.5 * torch.sum(1 + logvar_cat - mu_cat.pow(2) - logvar_cat.exp(), dim=1)
#print(KLD.size())
#p(y)
py = 0.1
return math.log(py) + BCE + KLD
def ELBO_xy_logterm(x, y, m_sampled, model,device):
mu_q2, logvar_q2, m_sampled = model(x=x,phase=2)
BCE = torch.zeros(x.size()[0]).to(device)
K = 1
for j in range(0,K):
mu_q1, logvar_q1, z_sampled = model(x=x,y=y,m_sampled=m_sampled,phase=1)
x_recon = model(x=x,y=y,z_sampled=z_sampled,m_sampled=m_sampled,phase=3)
log_pxyzm = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x_recon.view(-1,784)+1e-4)), dim=1)
BCE = BCE+ log_pxyzm
#BCE = (1/K) * BCE
#KL(q(z,m|x,y))
logvar_cat = torch.cat((logvar_q1, logvar_q2), dim = 1)
mu_cat = torch.cat((mu_q1, mu_q2), dim = 1)
KLD = 0.5 * torch.sum(1 + logvar_cat - mu_cat.pow(2) - logvar_cat.exp(), dim=1)
#print(KLD.size())
#p(y)
py = 0.1
return math.log(py) + BCE + KLD
def ELBO_xym0(x, y, model):
#Calculates ELBO(x,y,do(m=0))
x_recon, mu_q1, logvar_q1, mu_q2, logvar_q2 = model(x,y,manipulated=False)
BCE = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x_recon.view(-1,784)+1e-4)), dim=1)
KLD = 0.5 * torch.sum(1 + logvar_q1 - mu_q1.pow(2) - logvar_q1.exp(), dim=1)
return BCE + KLD + math.log(0.1)
def ELBO_xym0_logterm(x,y, model,device):
m_sampled = torch.zeros((x.size()[0],32)).type(torch.FloatTensor).to(device)
mu_q1, logvar_q1, z_sampled = model(x=x,y=y,m_sampled=m_sampled,phase=1)
#log p(x|y,z,m=0)
x_recon = model(x=x,y=y,z_sampled=z_sampled,m_sampled=m_sampled,phase=3)
BCE = torch.sum(torch.mul(x.view(-1,784), torch.log(x_recon.view(-1,784)+1e-4)) + torch.mul(1-x.view(-1,784), torch.log(1-x_recon.view(-1,784)+1e-4)), dim=1)
#log p(y)
py = 0.1
log_py = math.log(py)
#calculate log p(z)
zero_tensor = torch.zeros(z_sampled.size()).to(device)
one_tensor = torch.ones(z_sampled.size()).to(device)
temp = log_gaussian_torch(z_sampled,zero_tensor,one_tensor)
log_pz = torch.sum(temp, dim = 1).to(device)
#calculated log q(z|x,y,m=0)
temp = log_gaussian_torch(z_sampled,mu_q1,torch.exp(logvar_q1))
log_q1 = torch.sum(temp, dim =1).to(device)
return BCE + log_py + log_pz - log_q1