forked from bits-back/bits-back
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtvae_beta_binomial.py
166 lines (135 loc) · 6.07 KB
/
tvae_beta_binomial.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
import torch
import torch.utils.data
from torch import nn, optim, lgamma
from torch.nn import functional as F
from torchvision import datasets, transforms
from torch.distributions import Normal, Categorical, Beta, Binomial
from torchvision.utils import save_image
import numpy as np
torch.manual_seed(17)
def beta_binomial_log_pdf(k, n, alpha, beta):
numer = lgamma(n+1) + lgamma(k + alpha) + lgamma(n - k + beta) + lgamma(alpha + beta)
denom = lgamma(k+1) + lgamma(n - k + 1) + lgamma(n + alpha + beta) + lgamma(alpha) + lgamma(beta)
return numer - denom
class BetaBinomialVAE(nn.Module):
def __init__(self, hidden_dim=200, latent_dim=50):
super().__init__()
self.hidden_dim = hidden_dim
self.latent_dim = latent_dim
self.register_buffer('prior_mean', torch.zeros(1))
self.register_buffer('prior_std', torch.ones(1))
self.register_buffer('n', torch.ones(100, 784) * 255.)
self.fc1 = nn.Linear(784, self.hidden_dim)
self.bn1 = nn.BatchNorm1d(self.hidden_dim)
self.fc21 = nn.Linear(self.hidden_dim, self.latent_dim)
self.fc22 = nn.Linear(self.hidden_dim, self.latent_dim)
self.bn21 = nn.BatchNorm1d(self.latent_dim)
self.bn22 = nn.BatchNorm1d(self.latent_dim)
self.fc3 = nn.Linear(self.latent_dim, self.hidden_dim)
self.bn3 = nn.BatchNorm1d(self.hidden_dim)
self.fc4 = nn.Linear(self.hidden_dim, 784*2)
def encode(self, x):
"""Return mu, sigma on latent"""
h = x / 255. # otherwise we will have numerical issues
h = F.relu(self.bn1(self.fc1(h)))
return self.bn21(self.fc21(h)), torch.exp(self.bn22(self.fc22(h)))
def reparameterize(self, mu, std):
if self.training:
eps = torch.randn_like(std)
return eps.mul(std).add_(mu)
else:
return mu
def decode(self, z):
h = F.relu(self.bn3(self.fc3(z)))
h = self.fc4(h)
log_alpha, log_beta = torch.split(h, 784, dim=1)
return torch.exp(log_alpha), torch.exp(log_beta)
def loss(self, x):
z_mu, z_std = self.encode(x.view(-1, 784))
z = self.reparameterize(z_mu, z_std) # sample zs
x_alpha, x_beta = self.decode(z)
l = beta_binomial_log_pdf(x.view(-1, 784), self.n,
x_alpha, x_beta)
l = torch.sum(l, dim=1)
p_z = torch.sum(Normal(self.prior_mean, self.prior_std).log_prob(z), dim=1)
q_z = torch.sum(Normal(z_mu, z_std).log_prob(z), dim=1)
return -torch.mean(l + p_z - q_z) * np.log2(np.e) / 784.
def sample(self, device, epoch, num=64):
sample = torch.randn(num, self.latent_dim).to(device)
x_alpha, x_beta = self.decode(sample)
beta = Beta(x_alpha, x_beta)
p = beta.sample()
binomial = Binomial(255, p)
x_sample = binomial.sample()
x_sample = x_sample.float() / 255.
save_image(x_sample.view(num, 1, 28, 28),
'results/epoch_{}_samples.png'.format(epoch))
def reconstruct(self, x, device, epoch):
x = x.view(-1, 784).float().to(device)
z_mu, z_logvar = self.encode(x)
z = self.reparameterize(z_mu, z_logvar) # sample zs
x_alpha, x_beta = self.decode(z)
beta = Beta(x_alpha, x_beta)
p = beta.sample()
binomial = Binomial(255, p)
x_recon = binomial.sample()
x_recon = x_recon.float() / 255.
x_with_recon = torch.cat((x, x_recon))
save_image(x_with_recon.view(64, 1, 28, 28),
'results/epoch_{}_recon.png'.format(epoch))
def train(model, device, epoch, data_loader, optimizer, log_interval=10):
model.train()
losses = []
for batch_idx, (data, _) in enumerate(data_loader):
data = data.to(device)
optimizer.zero_grad()
loss = model.loss(data)
loss.backward()
losses.append(loss.item())
optimizer.step()
if batch_idx % log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(data_loader.dataset),
100. * batch_idx / len(data_loader),
loss.item()))
print('====> Epoch: {} Average loss: {:.4f}'.format(
epoch, np.mean(losses)))
def test(model, device, epoch, data_loader):
model.eval()
losses = []
for data, _ in data_loader:
data = data.to(device)
loss = model.loss(data)
losses.append(loss.item())
print('\nEpoch: {}\tTest loss: {:.6f}\n\n'.format(
epoch, np.mean(losses)
))
if __name__ == '__main__':
epochs = 20
batch_size = 100
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
class ToInt:
def __call__(self, pic):
return pic * 255
transforms.Compose([transforms.ToTensor(), ToInt()])
model = BetaBinomialVAE().to(device)
optimizer = optim.Adam(model.parameters(), lr=5e-4)
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('data/mnist', train=True, download=True,
transform=transforms.Compose([transforms.ToTensor(), ToInt()])),
batch_size=batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('data/mnist', train=False, download=True,
transform=transforms.Compose([transforms.ToTensor(), ToInt()])),
batch_size=batch_size, shuffle=True, **kwargs)
recon_dataset = datasets.MNIST('data/mnist', train=False, download=True,
transform=transforms.Compose([transforms.ToTensor(),
ToInt()])).test_data[:32]
for epoch in range(1, epochs + 1):
train(model, device, epoch, train_loader, optimizer)
test(model, device, epoch, test_loader)
model.reconstruct(recon_dataset, device, epoch)
model.sample(device, epoch)
torch.save(model.state_dict(), 'saved_params/torch_vae_beta_binomial_params_new')