-
Notifications
You must be signed in to change notification settings - Fork 1
/
on_imgs_vae.py
148 lines (119 loc) · 4.33 KB
/
on_imgs_vae.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
import torch
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
import torch.optim as optim
from torch.optim import lr_scheduler
from dataloader import ATeX
from models.ae import VAEConv3
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# mean_std = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
# transforms_list = [transforms.ToTensor(), transforms.Normalize(*mean_std)]
transforms_list = [transforms.ToTensor()]
transforms = transforms.Compose(transforms_list)
dataset = ATeX(split="val", transform=transforms)
atex = DataLoader(dataset, batch_size=256, shuffle=False, drop_last=False)
def final_loss(bce_loss, mu, logvar):
"""
This function will add the reconstruction loss (BCELoss) and the
KL-Divergence.
KL-Divergence = 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
:param bce_loss: recontruction loss
:param mu: the mean from the latent vector
:param logvar: log variance from the latent vector
"""
BCE = bce_loss
KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
# KLD = 0.5 * torch.sum(logvar.exp() - logvar - 1 + mu.pow(2))
return BCE + KLD
def lr_poly(base_lr, iter, max_iter, power):
return base_lr * ((1 - float(iter) / max_iter) ** (power))
def adjust_learning_rate(optimizer, lr):
optimizer.param_groups[0]['lr'] = lr
if len(optimizer.param_groups) > 1:
optimizer.param_groups[1]['lr'] = lr * 10
base_lr = 1e-3
model = VAEConv3()
criterion = torch.nn.BCELoss(reduction='sum')
optimizer = optim.Adam(model.parameters(), lr=base_lr, weight_decay=1e-5)
model.to(device)
num_epochs = 1000
i_iter = 0
outputs = []
for epoch in range(num_epochs):
for (img, _, _) in atex:
i_iter += 256
img = img.to(device)
optimizer.zero_grad()
lr = lr_poly(base_lr, i_iter, num_epochs * len(atex) * 256, 0.9)
adjust_learning_rate(optimizer, lr)
# recon = model(img)
# loss = criterion(recon, img)
recon, mu, logvar = model(img)
bce_loss = criterion(recon, img)
loss = final_loss(bce_loss, mu, logvar)
loss.backward()
optimizer.step()
# scheduler.step()
print(f"Epoch: {epoch+1:3d}, Loss: {loss.item():.4f}")
if (epoch % 100) == 0:
outputs.append((epoch, img, recon))
mean = np.array([0.485, 0.456, 0.406]).reshape(1, 1, 3)
std = np.array([0.229, 0.224, 0.225]).reshape(1, 1, 3)
for k in range(len(outputs)):
plt.figure()
imgs = outputs[k][1].detach().cpu().numpy()
recon = outputs[k][2].detach().cpu().numpy()
for i, item in enumerate(imgs):
if i >= 20:
break
plt.subplot(2, 20, i + 1)
item = item.transpose((1, 2, 0))
# item = std * item + mean
plt.axis('off')
plt.imshow(item)
for i, item in enumerate(recon):
if i >= 20:
break
plt.subplot(2, 20, 20 + i + 1)
item = item.transpose((1, 2, 0))
# item = std * item + mean
plt.axis('off')
plt.imshow(item)
plt.show()
# num_epochs = 300
# outputs = []
# for epoch in range(num_epochs):
# for (img, _, _) in atex:
# img = img.to(device)
# optimizer.zero_grad()
# recon, mu, logvar = model(img)
# bce_loss = criterion(recon, img)
# loss = final_loss(bce_loss, mu, logvar)
# loss.backward()
# optimizer.step()
# scheduler.step()
# print(f"Epoch: {epoch+1:3d}, Loss: {loss.item():.4f}")
# outputs.append((epoch, img, recon))
# for k in range(0, num_epochs, 20):
# plt.figure()
# imgs = outputs[k][1].detach().cpu().numpy()
# recon = outputs[k][2].detach().cpu().numpy()
# for i, item in enumerate(imgs):
# if i >= 20:
# break
# plt.subplot(2, 20, i + 1)
# item = item.transpose((1, 2, 0))
# # item = 0.229 * item + 0.485
# plt.axis('off')
# plt.imshow(item)
# for i, item in enumerate(recon):
# if i >= 20:
# break
# plt.subplot(2, 20, 20 + i + 1)
# item = item.transpose((1, 2, 0))
# # item = 0.229 * item + 0.485
# plt.axis('off')
# plt.imshow(item)
# plt.show()