/
cnn_vae_module_mnist.py
203 lines (177 loc) · 8.94 KB
/
cnn_vae_module_mnist.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
from __future__ import print_function
import torch
from torch import nn, optim
from torch.nn import functional as F
from torchvision.utils import save_image
import numpy as np
import matplotlib.pyplot as plt
from tool import visualize_ls, sample, get_param
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
x_dim = 12
h_dim = 1024
image_channels=1
image_size = 28
class Flatten(nn.Module):
def forward(self, input):
return input.view(input.size(0), -1) # [10,1024]
class UnFlatten(nn.Module):
def forward(self, input, size=h_dim):
return input.view(input.size(0), size, 1, 1)
class VAE(nn.Module):
def __init__(self, image_channels=image_channels, h_dim=h_dim, x_dim=x_dim):
super(VAE, self).__init__()
self.encoder = nn.Sequential(
nn.Conv2d(in_channels=image_channels, out_channels=64, kernel_size=4, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=4, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(in_channels=256, out_channels=1024, kernel_size=4, stride=1, padding=0),
nn.ReLU(),
Flatten()
)
self.fc1 = nn.Linear(h_dim, x_dim)
self.fc2 = nn.Linear(h_dim, x_dim)
self.fc3 = nn.Linear(x_dim, h_dim)
self.decoder = nn.Sequential(
UnFlatten(),
nn.ConvTranspose2d(in_channels=h_dim, out_channels=512, kernel_size=4, stride=1, padding=0),
nn.ReLU(),
nn.ConvTranspose2d(in_channels=512, out_channels=256, kernel_size=3, stride=2, padding=1),
nn.ReLU(),
nn.ConvTranspose2d(in_channels=256, out_channels=128, kernel_size=4, stride=2, padding=1),
nn.ReLU(),
nn.ConvTranspose2d(in_channels=128, out_channels=image_channels, kernel_size=4, stride=2, padding=1),
#nn.ReLU(),
#nn.ConvTranspose2d(in_channels=16, out_channels=image_channels, kernel_size=2, stride=2),
#nn.ReLU(),
#nn.ConvTranspose2d(in_channels=8, out_channels=image_channels, kernel_size=2, stride=2),
#nn.ReLU(),
#nn.ConvTranspose2d(in_channels=16, out_channels=image_channels, kernel_size=3, stride=2),
nn.Sigmoid(),
)
self.fc1 = nn.Linear(h_dim, x_dim)
self.fc2 = nn.Linear(h_dim, x_dim)
self.fc3 = nn.Linear(x_dim, h_dim)
self.softplus= nn.Softplus()
# prior (0,I)
self.prior_var = nn.Parameter(torch.Tensor(1, x_dim).float().fill_(1.0))
self.prior_logvar = nn.Parameter(self.prior_var.log())
self.prior_var.requires_grad = False
self.prior_logvar.requires_grad = False
def reparameterize(self, mu, logvar):
std = torch.exp(0.5*logvar)
eps = torch.randn_like(std)
return mu + eps*std
def encode(self, o_d):
h = self.encoder(o_d)
z, mu, logvar = self.bottleneck(h)
return z, mu, logvar
def decode(self, z):
z = self.fc3(z)
z = self.decoder(z)
return z
def bottleneck(self, h):
mu, logvar = self.fc1(h), self.fc2(h)
z = self.reparameterize(mu, logvar)
return z, mu, logvar
def forward(self, o_d):
z, mu, logvar = self.encode(o_d)
return self.decode(z), mu, logvar, z
# Reconstruction + KL divergence losses summed over all elements and batch
def loss_function(self, recon_o, o_d, en_mu, en_logvar, gmm_mu, gmm_var, iteration):
BCE = F.binary_cross_entropy(recon_o, o_d, reduction='sum')
beta = 1.0
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
if iteration != 0:
gmm_mu = nn.Parameter(gmm_mu)
prior_mu = gmm_mu
prior_mu.requires_grad = False
prior_mu = prior_mu.expand_as(en_mu).to(device)
gmm_var = nn.Parameter(gmm_var)
prior_var = gmm_var
prior_var.requires_grad = False
prior_var = prior_var.expand_as(en_logvar).to(device)
prior_logvar = nn.Parameter(prior_var.log())
prior_logvar.requires_grad = False
prior_logvar = prior_logvar.expand_as(en_logvar).to(device)
var_division = en_logvar.exp() / prior_var # Σ_0 / Σ_1
diff = en_mu - prior_mu # μ_１ - μ_0
diff_term = diff *diff / prior_var # (μ_1 - μ_0)(μ_1 - μ_0)/Σ_1
logvar_division = prior_logvar - en_logvar # log|Σ_1| - log|Σ_0| = log(|Σ_1|/|Σ_2|)
KLD = 0.5 * ((var_division + diff_term + logvar_division).sum(1) - x_dim)
else:
KLD = -0.5 * torch.sum(1 + en_logvar - en_mu.pow(2) - en_logvar.exp())
return BCE + KLD
def train(iteration, gmm_mu, gmm_var, epoch, train_loader, batch_size, all_loader, model_dir, agent):
prior_mean = torch.Tensor(len(train_loader), x_dim).float().fill_(0.0) # 最初のVAEの事前分布の\mu
model = VAE().to(device)
print(f"VAE_Agent {agent} Training Start({iteration}): Epoch:{epoch}, Batch_size:{batch_size}")
optimizer = optim.Adam(model.parameters(), lr=1e-3)
loss_list = np.zeros((epoch))
for i in range(epoch):
model.train()
train_loss = 0
for batch_idx, (data, _) in enumerate(train_loader):
data = data.to(device)
optimizer.zero_grad()
recon_batch, mu, logvar, x_d = model(data)
if iteration==0:
loss = model.loss_function(recon_batch, data, mu, logvar, gmm_mu=None, gmm_var=None, iteration=iteration)
else:
loss = model.loss_function(recon_batch, data, mu, logvar, gmm_mu[batch_idx*batch_size:(batch_idx+1)*batch_size], gmm_var[batch_idx*batch_size:(batch_idx+1)*batch_size], iteration=iteration)
loss = loss.mean()
loss.backward()
train_loss += loss.item()
optimizer.step()
if i == 0 or (i+1) % 25 == 0 or i == (epoch-1):
print('====> Epoch: {} Average loss: {:.4f}'.format(i+1, train_loss / len(train_loader.dataset)))
loss_list[i] = -(train_loss / len(train_loader.dataset))
# plot
plt.figure()
plt.plot(range(0,epoch), loss_list, color="blue", label="ELBO")
if iteration!=0:
loss_0 = np.load(model_dir+'/npy/loss'+agent+'_0.npy')
plt.plot(range(0,epoch), loss_0, color="red", label="ELBO_I0")
plt.xlabel('epoch'); plt.ylabel('ELBO'); plt.legend(loc='lower right')
plt.savefig(model_dir+'/graph'+agent+'/vae_loss_'+str(iteration)+'.png')
plt.close()
np.save(model_dir+'/npy/loss'+agent+'_'+str(iteration)+'.npy', np.array(loss_list))
torch.save(model.state_dict(), model_dir+"/pth/vae"+agent+"_"+str(iteration)+".pth")
x_d, label = send_all_x(iteration=iteration, all_loader=all_loader, model_dir=model_dir, agent=agent)
# Send latent variable x_d to GMM
return x_d, label, loss_list
def decode(iteration, decode_k, sample_num, sample_d, manual, model_dir, agent):
print(f"Reconstruct image on Agent: {agent}, category: {decode_k}")
model = VAE().to("cpu")
model.load_state_dict(torch.load(str(model_dir)+"/pth/vae"+agent+"_"+str(iteration)+".pth",map_location=torch.device('cpu'))); model.eval()
mu_gmm_kd, lambda_gmm_kdd, pi_gmm_k = get_param(iteration, model_dir=model_dir, agent=agent)
sample_d = torch.from_numpy(sample_d.astype(np.float32)).clone()
with torch.no_grad():
sample_d = sample_d.to("cpu")
sample_d = model.decode(sample_d).cpu()
save_image(sample_d.view(sample_num, image_channels, image_size, image_size),model_dir+'/recon'+agent+'/random_'+str(decode_k)+'.png') if manual != True else save_image(sample_d.view(sample_num, image_channels, image_size, image_size),model_dir+'/recon'+agent+'/manual_'+str(decode_k)+'.png')
def plot_latent(iteration, all_loader, model_dir, agent):
print(f"Plot latent space on Agent: {agent}")
model = VAE().to("cpu")
model.load_state_dict(torch.load(model_dir+"/pth/vae"+agent+"_"+str(iteration)+".pth", map_location=torch.device('cpu')))
model.eval()
for batch_idx, (data, label) in enumerate(all_loader):
data = data.to("cpu")
recon_batch, mu, logvar, x_d = model(data)
x_d = x_d.cpu()
visualize_ls(iteration, x_d.detach().numpy(), label, model_dir,agent=agent)
break
def send_all_x(iteration, all_loader, model_dir, agent):
model = VAE().to(device)
model.load_state_dict(torch.load(model_dir+"/pth/vae"+agent+"_"+str(iteration)+".pth"))
model.eval()
for batch_idx, (data, label) in enumerate(all_loader):
data = data.to(device)
recon_batch, mu, logvar, x_d = model(data)
x_d = x_d.cpu()
label = label.cpu()
return x_d.detach().numpy(), label.detach().numpy()