-
Notifications
You must be signed in to change notification settings - Fork 4
/
fc_gauss.py
227 lines (186 loc) · 8.07 KB
/
fc_gauss.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
from __future__ import division
import torch
import numpy as np
from src.layers import SkipConnection
from src.utils import BaseNet, to_variable, cprint
from src.radam import RAdam
from src.probability import normal_parse_params, GaussianLoglike
import torch.nn as nn
from torch.nn import MSELoss
from torch.distributions import kl_divergence
import torch.backends.cudnn as cudnn
from torch.distributions.normal import Normal
from .models import MLP_prior_net, MLP_recognition_net, MLP_generator_net, \
MLP_preact_prior_net, MLP_preact_recognition_net, MLP_preact_generator_net
class VAEAC_gauss(nn.Module):
def __init__(self, input_dim, width, depth, latent_dim, pred_sig=True):
super(VAEAC_gauss, self).__init__()
self.latent_dim = latent_dim
self.recognition_net = MLP_preact_recognition_net(input_dim, width, depth, latent_dim)
self.prior_net = MLP_preact_prior_net(input_dim, width, depth, latent_dim)
if pred_sig:
self.generator_net = MLP_preact_generator_net(2*input_dim, width, depth, latent_dim)
self.rec_loglike = GaussianLoglike(min_sigma=1e-2)
else:
self.generator_net = MLP_preact_generator_net(input_dim, width, depth, latent_dim)
self.m_rec_loglike = MSELoss(reduction='none')
self.pred_sig = pred_sig
self.sigma_mu = 1e4
self.sigma_sigma = 1e-4
@staticmethod
def apply_mask(x, mask):
"""Positive bits in mask are set to 0 in x (observed)"""
observed = x.clone() # torch.tensor(x)
observed[mask.bool()] = 0
return observed
def recognition_encode(self, x):
approx_post_params = self.recognition_net(x)
approx_post = normal_parse_params(approx_post_params, 1e-3)
return approx_post
def prior_encode(self, x, mask):
x = self.apply_mask(x, mask)
x = torch.cat([x, mask], 1)
prior_params = self.prior_net(x)
prior = normal_parse_params(prior_params, 1e-3)
return prior
def decode(self, z_sample):
rec_params = self.generator_net(z_sample)
return rec_params
def reg_cost(self, prior):
num_objects = prior.mean.shape[0]
mu = prior.mean.view(num_objects, -1)
sigma = prior.scale.view(num_objects, -1)
mu_regularizer = -(mu ** 2).sum(-1) / 2 / (self.sigma_mu ** 2)
sigma_regularizer = (sigma.log() - sigma).sum(-1) * self.sigma_sigma
return mu_regularizer + sigma_regularizer
def vlb(self, prior, approx_post, x, rec_params):
if self.pred_sig:
rec = self.rec_loglike(rec_params, x).view(x.shape[0], -1).sum(-1)
else:
rec = -self.m_rec_loglike(rec_params, x).view(x.shape[0], -1).sum(-1)
prior_regularization = self.reg_cost(prior).view(x.shape[0], -1).sum(-1)
kl = kl_divergence(approx_post, prior).view(x.shape[0], -1).sum(-1)
return rec - kl + prior_regularization
def iwlb(self, prior, approx_post, x, K=50):
estimates = []
for i in range(K):
latent = approx_post.rsample()
rec_params = self.decode(latent)
if self.pred_sig:
rec_loglike = self.rec_loglike(rec_params, x).view(x.shape[0], -1).sum(-1)
else:
rec_loglike = -self.m_rec_loglike(rec_params, x).view(x.shape[0], -1).sum(-1)
prior_log_prob = prior.log_prob(latent)
prior_log_prob = prior_log_prob.view(x.shape[0], -1)
prior_log_prob = prior_log_prob.sum(-1)
proposal_log_prob = approx_post.log_prob(latent)
proposal_log_prob = proposal_log_prob.view(x.shape[0], -1)
proposal_log_prob = proposal_log_prob.sum(-1)
estimate = rec_loglike + prior_log_prob - proposal_log_prob
estimates.append(estimate[:, None])
return torch.logsumexp(torch.cat(estimates, 1), 1) - np.log(K)
class VAEAC_gauss_net(BaseNet):
def __init__(self, input_dim, width, depth, latent_dim, pred_sig=True, lr=1e-3, cuda=True):
super(VAEAC_gauss_net, self).__init__()
cprint('y', 'VAE_gauss_net')
self.cuda = cuda
self.input_dim = input_dim
self.width = width
self.depth = depth
self.latent_dim = latent_dim
self.lr = lr
self.pred_sig = pred_sig
self.create_net()
self.create_opt()
self.epoch = 0
self.schedule = None
self.vlb_scale = 1 / input_dim # scale for dimensions of input so we can use same LR always
def create_net(self):
torch.manual_seed(42)
torch.cuda.manual_seed(42)
self.model = VAEAC_gauss(self.input_dim, self.width, self.depth, self.latent_dim, self.pred_sig)
if self.cuda:
self.model = self.model.cuda()
cudnn.benchmark = True
print(' Total params: %.2fM' % (self.get_nb_parameters() / 1000000.0))
def create_opt(self):
self.optimizer = RAdam(self.model.parameters(), lr=self.lr) # torch.optim.Adam
def fit(self, x, mask):
self.set_mode_train(train=True)
x, mask = to_variable(var=(x, mask), cuda=self.cuda)
self.optimizer.zero_grad()
prior = self.model.prior_encode(x, mask)
approx_post = self.model.recognition_encode(x)
z_sample = approx_post.rsample()
rec_params = self.model.decode(z_sample)
vlb = self.model.vlb(prior, approx_post, x, rec_params)
loss = (- vlb * self.vlb_scale).mean()
loss.backward()
self.optimizer.step()
if self.pred_sig:
rec_return = normal_parse_params(rec_params, 1e-3)
else:
rec_return = rec_params
return vlb.mean().item(), rec_return
def eval(self, x, mask, sample=False):
self.set_mode_train(train=False)
x, mask = to_variable(var=(x, mask), cuda=self.cuda)
prior = self.model.prior_encode(x, mask)
approx_post = self.model.recognition_encode(x)
if sample:
z_sample = approx_post.sample()
else:
z_sample = approx_post.loc
rec_params = self.model.decode(z_sample)
vlb = self.model.vlb(prior, approx_post, x, rec_params)
if self.pred_sig:
rec_return = normal_parse_params(rec_params, 1e-3)
else:
rec_return = rec_params
return vlb.mean().item(), rec_return
def eval_iw(self, x, mask, k=50):
self.set_mode_train(train=False)
x, mask = to_variable(var=(x, mask), cuda=self.cuda)
prior = self.model.prior_encode(x, mask)
approx_post = self.model.recognition_encode(x)
iw_lb = self.model.iwlb(prior, approx_post, x, k)
return iw_lb.mean().item()
def get_prior(self, x, mask):
self.set_mode_train(train=False)
x, mask = to_variable(var=(x, mask), cuda=self.cuda)
prior = self.model.prior_encode(x, mask)
return prior
def get_post(self, x):
self.set_mode_train(train=False)
x, = to_variable(var=(x,), cuda=self.cuda)
approx_post = self.model.recognition_encode(x)
return approx_post
def inpaint(self, x, mask, Nsample=1, z_mean=False):
self.set_mode_train(train=False)
x, mask = to_variable(var=(x, mask), cuda=self.cuda)
prior = self.model.prior_encode(x, mask)
out = []
for i in range(Nsample):
if z_mean:
z_sample = prior.loc.data
else:
z_sample = prior.sample()
rec_params = self.model.decode(z_sample)
out.append(rec_params.data)
out = torch.stack(out, dim=0)
if self.pred_sig:
return [normal_parse_params(out[i], 1e-2) for i in range(Nsample)]
else:
return out
def regenerate(self, z, grad=False):
self.set_mode_train(train=False)
if grad:
if not z.requires_grad:
z.requires_grad = True
else:
z, = to_variable(var=(z,), volatile=True, cuda=self.cuda)
out = self.model.decode(z)
if self.pred_sig:
return normal_parse_params(out, 1e-2)
else:
return out.data