-
Notifications
You must be signed in to change notification settings - Fork 1
/
cvae.py
210 lines (183 loc) · 7.38 KB
/
cvae.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
import numpy as np
import os
import argparse
import torch
from torchvision import datasets, transforms
from torch import nn, optim
from torch.autograd import Variable
from torch.distributions import Bernoulli
from matplotlib import pyplot as plt
import inspect
import ipdb
parser = argparse.ArgumentParser(description = "categorical VAE with MNIST")
parser.add_argument('--batch_size', type = int, default = 100,
help = 'input batch size for training (default: 100)')
parser.add_argument('--num_iters', type = int, default = 50000,
help = 'number of iterations for training (default: 50000)')
parser.add_argument('--lr', type = float, default = 0.001,
help = 'learning rate (default: 0.001)')
parser.add_argument('--tau', type = float, default = 1,
help = 'initial temperature for gumbel softmax (default: 1)')
parser.add_argument('--anneal_rate', type = float, default = 0.00003,
help = 'anneal rate for temperature (default: 0.00003)')
parser.add_argument('--min_temp', type = float, default = 0.5,
help = 'minimum temperature (default: 0.5)')
parser.add_argument('--no_cuda', action='store_true', default=False,
help='enables CUDA training')
parser.add_argument('--log_interval', type = int, default = 5000,
help = 'how many iterations to wait before logging training status')
parser.add_argument('--K', type = int, default = 10,
help = 'number of classes')
parser.add_argument('--N', type = int, default = 30,
help = 'number of categorical distributions')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
class CVAE(nn.Module):
def __init__(self, N, K, tau):
super().__init__()
self.fc1 = nn.Linear(784, 512)
self.fc2 = nn.Linear(512, 256)
self.fc3 = nn.Linear(256, N*K)
self.fc4 = nn.Linear(N*K, 256)
self.fc5 = nn.Linear(256, 512)
self.fc6 = nn.Linear(512, 784)
self.relu = nn.ReLU()
self.softmax = nn.Softmax(dim=-1)
self.sigmoid = nn.Sigmoid()
self.K = K
self.N = N
self.temperature = tau
def _sample_gumbel(self, shape, eps=1e-20):
U = torch.rand(shape)
return -torch.log(-torch.log(U + eps) + eps)
def _gumbel_softmax_sample(self, logits):
sample = Variable(self._sample_gumbel(logits.size()[-1]))
if logits.is_cuda:
sample = sample.cuda()
y = logits + sample
return self.softmax(y / self.temperature)
def _gumbel_softmax(self, logits, hard=False):
y = self._gumbel_softmax_sample(logits)
return y
def _encoder(self, x):
h1 = self.relu(self.fc1(x))
h2 = self.relu(self.fc2(h1))
h3 = self.fc3(h2)
logits_y = h3.view(-1, self.K)
q_y = self.softmax(logits_y)
return q_y, logits_y
def _decoder(self, y):
h3 = self.relu(self.fc4(y))
h4 = self.relu(self.fc5(h3))
logits_x = self.fc6(h4)
p_x = self.sigmoid(logits_x)
return p_x, logits_x
def forward(self, x):
q_y, logits_y = self._encoder(x.view(-1, 784))
y = self._gumbel_softmax(logits_y).view(-1, self.N * self.K)
p_x, logits = self._decoder(y)
return p_x, q_y, logits
def loss_fn(q_y, p_x, x, N, K, logits):
prior = Variable(torch.log(torch.from_numpy(np.array([1.0/K]))).type(torch.FloatTensor)).cuda()
kl = (q_y * (torch.log(q_y+1e-20) - prior)).view(-1, N, K)
KL = torch.sum(kl)
data = x.view(-1, 784)
bceloss = nn.BCELoss(size_average=False)
elbo = -bceloss(p_x, data) - KL
loss = -(1.0/p_x.size()[0])*elbo
return loss
def train(model, optimizer, train_loader, num_iters, dat, args):
model.train()
train_loss = 0
np_temp = model.temperature
for i, (data, _) in enumerate(train_loader):
num_iters += 1
if args.cuda:
data = data.cuda()
data = Variable(data)
optimizer.zero_grad()
p_x, q_y, logits = model(data)
loss = loss_fn(q_y, p_x, data, args.N, args.K, logits)
loss.backward()
train_loss += loss.data[0]
optimizer.step()
if num_iters % 100 == 1:
dat.append([num_iters, np_temp, loss.data[0]])
if num_iters % 1000 == 1:
np_temp = np.maximum(args.tau * np.exp(-args.anneal_rate * num_iters),
args.min_temp)
for param_group in optimizer.param_groups:
param_group['lr'] *= 0.9
model.temperature = np_temp
if num_iters % args.log_interval == 1:
print('Step %d, ELBO: %0.3f' % (num_iters,-loss.data[0]))
return model, optimizer, num_iters, dat
def test(model, test_loader):
model.eval()
test_loss = 0
for i, (data, _) in enumerate(test_loader):
if args.cuda:
data = data.cuda()
data = Variable(data, volatile=True)
#p_x, q_y, logits = model(data)
#test_loss += loss_fn(q_y, p_x, data, N, K).data[0]
def img_gen(model, args):
model.eval()
M = 100 * args.N
np_y = np.zeros((M, args.K))
np_y[range(M), np.random.choice(args.K,M)] = 1
y = Variable(torch.from_numpy(np.reshape(np_y, [100, args.N*args.K])).type(torch.FloatTensor))
if args.cuda:
y = y.cuda()
p_x, logits_x = model._decoder(y)
np_y = np_y.reshape((10, 10, args.N, args.K))
np_y = np.concatenate(np.split(np_y,10,axis=0),axis=3)
np_y = np.concatenate(np.split(np_y,10,axis=1),axis=2)
y_img = np.squeeze(np_y)
x_p = p_x.view(10, 10, 28, 28).cpu().data.numpy()
x_p = np.concatenate(np.split(x_p, 10, axis=0), axis=3)
x_p = np.concatenate(np.split(x_p, 10, axis=1), axis=2)
x_img = np.squeeze(x_p)
f,axarr=plt.subplots(1, 2, figsize=(15,15))
# samples
axarr[0].matshow(y_img, cmap=plt.cm.gray)
axarr[0].set_title('Z Samples')
# reconstruction
axarr[1].imshow(x_img, cmap=plt.cm.gray, interpolation='none')
axarr[1].set_title('Generated Images')
f.tight_layout()
f.savefig('code_torch.png')
def main(args):
torch.manual_seed(100)
if args.cuda:
torch.cuda.manual_seed(100)
data_path = os.path.expanduser('~/data/mnist')
kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.MNIST(data_path, train = True, download = True,
transform = transforms.ToTensor()),
batch_size = args.batch_size, shuffle = True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(data_path, train = False, transform = transforms.ToTensor()),
batch_size = args.batch_size, shuffle = True, **kwargs)
num_iters = 0
dat = []
model = CVAE(args.N, args.K, args.tau)
if args.cuda:
model.cuda()
optimizer = optim.Adam(model.parameters(), lr = args.lr)
while num_iters <= args.num_iters:
model, optimizer, num_iters, dat = train(model,
optimizer, train_loader, num_iters, dat, args)
#test(model, test_loader)
img_gen(model, args)
dat=np.array(dat).T
f,axarr=plt.subplots(1,2)
axarr[0].plot(dat[0],dat[1])
axarr[0].set_ylabel('Temperature')
axarr[1].plot(dat[0],dat[2])
axarr[1].set_ylabel('-ELBO')
plt.show()
if __name__ == '__main__':
with ipdb.slaunch_ipdb_on_exception():
main(args)