-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathgeneration.py
222 lines (176 loc) · 8.97 KB
/
generation.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
from torch.autograd import Variable
# evaluate.py
import torch
import plugins
import torchvision
import numpy as np
import scipy
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
from skimage.measure import compare_ssim as ssim
from skimage.measure import compare_psnr as psnr
import pyamg
import PIL.Image
from torchvision.utils import save_image
import warnings
warnings.filterwarnings("ignore")
class Generator():
def __init__(self, args, netD, netG, netE):
self.netD = netD
self.netG = netG
self.netE = netE
self.args = args
self.nchannels = args.nchannels
self.resolution_high = args.resolution_high
self.resolution_wide = args.resolution_wide
self.nz = args.nz
self.wcom = args.disc_loss_weight
self.cuda = args.cuda
self.citers = args.citers
self.lr = args.learning_rate_vae
self.momentum = args.momentum
self.batch_size = args.batch_size
self.use_encoder = args.use_encoder
self.input = Variable(torch.FloatTensor(self.batch_size,self.nchannels,self.resolution_high,self.resolution_wide), volatile=True).cuda()
self.epsilon = Variable(torch.randn(self.batch_size, self.nz), volatile=True).cuda()
self.noise = Variable(torch.FloatTensor(self.batch_size, self.nz, 1, 1).normal_(0, 1), volatile=True)
if args.cuda:
self.input = self.input.cuda()
self.epsilon = self.epsilon.cuda()
self.noise = self.noise.cuda()
self.log_eval_loss = plugins.Logger(args.logs, 'Generation.txt')
self.params_eval_loss = ['Image', 'SSIM_1', 'SSIM_2', 'SSIM_3', 'SSIM_4', 'PSNR_1', 'PSNR_2', 'PSNR_3', 'PSNR_4', 'DiscScore_1', 'DiscScore_2', 'DiscScore_3', 'DiscScore_4']
self.log_eval_loss.register(self.params_eval_loss)
self.losses = {}
self.log_eval_monitor = plugins.Monitor()
self.params_eval_monitor = ['Image', 'SSIM_1', 'SSIM_2', 'SSIM_3', 'SSIM_4', 'PSNR_1', 'PSNR_2', 'PSNR_3', 'PSNR_4', 'DiscScore_1', 'DiscScore_2', 'DiscScore_3', 'DiscScore_4']
self.log_eval_monitor.register(self.params_eval_monitor)
self.print = '[%d/%d] '
for item in self.params_eval_loss:
self.print = self.print + item + " %.4f "
def ssim(self, data1, data2):
num = data1.size(0)
nchannels = data1.size(1)
data1 = data1.transpose(1, 2).transpose(2, 3)
data2 = data2.transpose(1, 2).transpose(2, 3)
score = 0
if nchannels > 1:
for i in range(num):
img1 = data1[i].numpy()
img2 = data2[i].numpy()
range1 = img1.max() - img1.min()
range2 = img2.max() - img2.min()
range3 = max(range1, range2)
score += ssim(img1, img2, dynamic_range=range3,
multichannel=True)
else:
for i in range(num):
img1 = data1[i].numpy()
img2 = data2[i].numpy()
score += ssim(img1, img2, dynamic_range=self.range)
return score/num
def psnr(self, data1, data2):
num = data1.size(0)
nchannels = data1.size(1)
data1 = data1.transpose(1, 2).transpose(2, 3)
data2 = data2.transpose(1, 2).transpose(2, 3)
score = 0
for i in range(num):
img1 = data1[i].numpy()
img2 = data2[i].numpy()
range1 = img1.max() - img1.min()
range2 = img2.max() - img2.min()
range3 = max(range1, range2)
score += psnr(img1, img2, dynamic_range=range3)
return score/num
def generate(self, dataloader):
data_iter = iter(dataloader)
data_i = 0
while data_i < len(dataloader):
data_real = data_iter.next()[0]
data_i += 1
batch_size = data_real.size(0)
self.input.data.resize_(data_real.size()).copy_(data_real)
if self.use_encoder:
self.epsilon.data.resize_(batch_size, self.nz).normal_(0, 1)
noise_mu, noise_logvar = self.netE(self.input)
noise_sigma = torch.exp(torch.mul(noise_logvar, 0.5))
latents = noise_mu + torch.mul(noise_sigma, self.epsilon)
latents = latents.unsqueeze(-1).unsqueeze(-1)
self.noise.data.copy_(latents.data)
else:
self.noise.data.normal_(0, 1)
self.output = [self.netG[i].forward(self.noise) for i in range(4)]
ssims = [self.ssim(self.input.data.cpu(), self.output[i].data.cpu()) for i in range(4)]
psnrs = [self.psnr(self.input.data.cpu(), self.output[i].data.cpu()) for i in range(4)]
disc_scores = [self.netD(self.output[i]).mean(0).data[0] for i in range(4)]
save_image(normalize(self.input.data[0].cpu()), "{}/{}_Original.png".format(self.args.save, data_i), padding=0, normalize=True)
for k in range(4):
save_image(normalize(self.output[k].data[0].cpu()), "{}/{}_Fake_Stage_{}.png".format(self.args.save, data_i, k+1), padding=0, normalize=True)
self.losses['Image'] = float(data_i)
for k in range(4):
self.losses['SSIM_{}'.format(k+1)] = ssims[k]
self.losses['PSNR_{}'.format(k+1)] = psnrs[k]
self.losses['DiscScore_{}'.format(k+1)] = disc_scores[k]
self.log_eval_monitor.update(self.losses, batch_size, keepsame=True)
print(self.print % tuple([data_i, len(dataloader)] + [self.losses[key]
for key in self.params_eval_monitor]))
loss = self.log_eval_monitor.getvalues()
self.log_eval_loss.update(loss)
def generate_one(self, dataloader):
data_iter = iter(dataloader)
data_i = 0
while data_i < len(dataloader):
data_real = data_iter.next()[0]
data_i += 1
batch_size = data_real.size(0)
self.input.data.resize_(data_real.size()).copy_(data_real)
if self.use_encoder:
self.epsilon.data.resize_(batch_size, self.nz).normal_(0, 1)
noise_mu, noise_logvar = self.netE(self.input)
noise_sigma = torch.exp(torch.mul(noise_logvar, 0.5))
latents = noise_mu + torch.mul(noise_sigma, self.epsilon)
latents = latents.unsqueeze(-1).unsqueeze(-1)
self.noise.data.copy_(latents.data)
else:
self.noise.data.normal_(0, 1)
self.output = self.netG.forward(self.noise)
# ssims = [self.ssim(self.input.data.cpu(), self.output[i].data.cpu()) for i in range(4)]
# psnrs = [self.psnr(self.input.data.cpu(), self.output[i].data.cpu()) for i in range(4)]
# disc_scores = [self.netD(self.output[i]).mean(0).data[0] for i in range(4)]
# save_image(normalize(self.input.data[0].cpu()), "{}/{}_Original.png".format(self.args.save, data_i), padding=0, normalize=True)
# for k in range(4):
save_image(normalize(self.output.data[0].cpu()), "{}/{}_wgan_fake.png".format(self.args.save, data_i), padding=0, normalize=True)
def interpolate(self, dataloader):
if self.use_encoder:
data_iter = iter(dataloader)
data_real = data_iter.next()[0]
batch_size = data_real.size(0)
self.input.data.resize_(data_real.size()).copy_(data_real)
self.epsilon.data.resize_(batch_size, self.nz).normal_(0, 1)
noise_mu, noise_logvar = self.netE(self.input)
noise_sigma = torch.exp(torch.mul(noise_logvar, 0.5))
latents = noise_mu + torch.mul(noise_sigma, self.epsilon)
latents = latents.unsqueeze(-1).unsqueeze(-1)
self.noise.data.copy_(latents.data)
z1 = self.noise[0:3]
z2 = self.noise[3:6]
for i in range(8):
z = z1 + (z2-z1)*i/7
self.output = [self.netG[i].forward(z) for i in range(4)]
for stage in range(4):
for anchor in range(3):
save_image(normalize(self.output[stage].data[anchor].cpu()), "{}/Montage_{}_Stage_{}_Z_{}.png".format(self.args.save, anchor+1, stage+1, i+1), padding=0, normalize=True)
else:
z1 = Variable(torch.FloatTensor(3, self.nz, 1, 1).normal_(0, 1), volatile=True).cuda()
z2 = Variable(torch.FloatTensor(3, self.nz, 1, 1).normal_(0, 1), volatile=True).cuda()
for i in range(8):
z = z1 + (z2-z1)*i/7
self.output = [self.netG[i].forward(z) for i in range(4)]
for stage in range(4):
for anchor in range(3):
save_image(normalize(self.output[stage].data[anchor].cpu()), "{}/Montage_{}_Stage_{}_Z_{}.png".format(self.args.save, anchor+1, stage+1, i+1), padding=0, normalize=True)
def normalize(image):
image = (image - image.min())
return image / image.max()