-
Notifications
You must be signed in to change notification settings - Fork 1
/
invert.py
341 lines (282 loc) · 12.6 KB
/
invert.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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
import os
import time
import argparse
from glob import glob
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.optim as optim
import torch.nn as nn
import torchvision.datasets as dst
import torchvision.transforms as tfs
from torch.utils.data import DataLoader
import model
# define the data_loader
class CustomDataset(torch.utils.data.Dataset):
def __init__(self, data, label, transform=None):
self.transform = transform
self.data = data
self.data_num = len(data)
self.label = label
def __len__(self):
return self.data_num
def __getitem__(self, idx):
out_data = self.data[idx]
out_label = self.label[idx]
if self.transform:
out_data = self.transform(out_data)
return out_data, out_label
def train(dataloader, epoch, params):
net_c.eval()
model1.eval()
std = (1 / 256.0) * (1 - epoch / 256.0)
cum_loss = [0] * 5
for batch_idx, (activations, real_image) in enumerate(dataloader):
activations = activations.type('torch.FloatTensor').to(device)
real_image = real_image.type('torch.FloatTensor').to(device)
# update Generator
optimizer_g.zero_grad()
net_g.train()
net_d.eval()
real_l = torch.argmax(model1(real_image), 1)
fake_image = net_g(activations)
fake_logit = net_d(fake_image, std)
fake_feature = net_c(fake_image)
real_feature = net_c(real_image)
loss_feat = loss_f_feat(real_feature, fake_feature)
loss_img = loss_f_img(real_image, fake_image)
if params.gan_type == 'lsgan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
loss_f_adv = nn.MSELoss()
loss_f_adv.to(device)
loss_adv = loss_f_adv(fake_logit, ones)
elif params.gan_type == 'gan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
loss_f_adv = nn.BCEWithLogitsLoss()
loss_f_adv.to(device)
loss_adv = loss_f_adv(fake_logit, ones)
elif params.gan_type == 'wgan':
loss_adv = -torch.mean(fake_logit)
loss_g = params.lambda_feat * loss_feat + \
params.lambda_adv * loss_adv + \
params.lambda_img * loss_img
loss_g.backward()
optimizer_g.step()
# update Discriminator
optimizer_d.zero_grad()
net_g.eval()
net_d.train()
fake_logit = net_d(fake_image.detach(), std)
real_logit = net_d(real_image, std)
if params.gan_type == 'lsgan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
zeros = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.zeros(real_image.shape[0])]
loss_f_adv = nn.MSELoss()
loss_f_adv.to(device)
loss_d = loss_f_adv(real_logit, ones) + loss_f_adv(fake_logit, zeros)
elif params.gan_type == 'gan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
zeros = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.zeros(real_image.shape[0])]
loss_f_adv = nn.BCEWithLogitsLoss()
loss_f_adv.to(device)
loss_d = loss_f_adv(real_logit, ones) + loss_f_adv(fake_logit, zeros)
elif params.gan_type == 'wgan':
loss_d = -torch.mean(real_logit) + torch.mean(fake_logit)
if not (params.apply_th and loss_d.item() < 0.1 * loss_adv.item()):
loss_d.backward()
optimizer_d.step()
if params.gan_type == 'wgan':
# clip parameters
for param in net_d.parameters():
param.data.clamp_(-params.clip_value, params.clip_value)
loss_list = [loss_img, loss_adv, loss_feat, loss_g, loss_d]
for i, _loss in enumerate(loss_list):
cum_loss[i] += _loss.item()
for i in range(len(cum_loss)):
cum_loss[i] /= (batch_idx + 1)
return cum_loss
def test(dataloader, epoch, params):
net_g.eval()
net_d.eval()
net_c.eval()
model1.eval()
std = 0.0
cum_loss = [0] * 5
with torch.no_grad():
for batch_idx, (activations, real_image) in enumerate(dataloader):
activations = activations.type('torch.FloatTensor').to(device)
real_image = real_image.type('torch.FloatTensor').to(device)
real_l = torch.argmax(model1(real_image), 1)
fake_image = net_g(activations)
fake_logit = net_d(fake_image, std)
real_logit = net_d(real_image, std)
fake_feature = net_c(fake_image)
real_feature = net_c(real_image)
# generator loss
loss_feat = loss_f_feat(real_feature, fake_feature)
loss_img = loss_f_img(real_image, fake_image)
if params.gan_type == 'lsgan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
loss_f_adv = nn.MSELoss()
loss_f_adv.to(device)
loss_adv = loss_f_adv(fake_logit, ones)
elif params.gan_type == 'gan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
loss_f_adv = nn.BCEWithLogitsLoss()
loss_f_adv.to(device)
loss_adv = loss_f_adv(fake_logit, ones)
elif params.gan_type == 'wgan':
loss_adv = -torch.mean(fake_logit)
loss_g = params.lambda_feat * loss_feat + \
params.lambda_adv * loss_adv + \
params.lambda_img * loss_img
# discriminator loss
if params.gan_type == 'lsgan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
zeros = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.zeros(real_image.shape[0])]
loss_f_adv = nn.MSELoss()
loss_f_adv.to(device)
loss_d = loss_f_adv(real_logit, ones) + loss_f_adv(fake_logit, zeros)
elif params.gan_type == 'gan':
ones = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.ones(real_image.shape[0])]
zeros = torch.eye(2, dtype=torch.float32, requires_grad=False,
device=device)[np.zeros(real_image.shape[0])]
loss_f_adv = nn.BCEWithLogitsLoss()
loss_f_adv.to(device)
loss_d = loss_f_adv(real_logit, ones) + loss_f_adv(fake_logit, zeros)
elif params.gan_type == 'wgan':
loss_d = -torch.mean(real_logit) + torch.mean(fake_logit)
loss_list = [loss_img, loss_adv, loss_feat, loss_g, loss_d]
for i, _loss in enumerate(loss_list):
cum_loss[i] += _loss.item()
# plot
real_image = real_image.detach().cpu().numpy()
fake_image = fake_image.detach().cpu().numpy()
plt.figure(figsize=(12, 12))
for i in range(min(50, len(real_image))):
img1 = np.transpose(real_image[i] * 255, (1, 2, 0)).astype('uint8')
ax = plt.subplot(10, 10, 2 * i + 1)
ax.imshow(img1)
ax.set_title('Real')
plt.axis('off')
img2 = np.transpose(fake_image[i] * 255, (1, 2, 0)).astype('uint8')
ax = plt.subplot(10, 10, 2 * i + 2)
ax.imshow(img2)
ax.set_title('Generated')
plt.axis('off')
plt.tight_layout()
plt.savefig(save_dir + 'img_epoch' + str(epoch) + '.png')
plt.close()
for i in range(len(cum_loss)):
cum_loss[i] /= (batch_idx + 1)
return cum_loss
if __name__ == "__main__":
# params
parser = argparse.ArgumentParser()
parser.add_argument('--target_layer', type=str, default='22')
parser.add_argument('--epochs', default=256, type=int)
parser.add_argument('--trainBatch', default=16, type=int)
parser.add_argument('--testBatch', default=16, type=int)
parser.add_argument('--lambda_feat', default=0.01, type=float)
parser.add_argument('--lambda_adv', default=0.001, type=float)
parser.add_argument('--lambda_img', default=1, type=float)
parser.add_argument('--lr', default=0.002, type=float)
parser.add_argument('--lr_decay', default=0.96, type=float)
parser.add_argument('--beta1', default=0.5, type=float)
parser.add_argument('--beta2', default=0.99, type=float)
parser.add_argument('--gan_type', default='lsgan', type=str)
parser.add_argument('--optimizer_type', default='Adam', type=str)
parser.add_argument('--apply_th', action='store_true')
parser.add_argument('--clip_value', default=0.05, type=float)
params = parser.parse_args()
data_dir = 'data/cifar10/'
resp_dir = 'resps/vgg16/'
save_dir = 'generated/vgg16/'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
device = torch.device('cuda')
# load the images and responses
images = dict()
resps = dict()
for state in ['test', 'train']:
# response
files = glob(resp_dir + state + '_' + \
str(params.target_layer) + '_*.npy')
for i, file in enumerate(files):
if i == 0:
_resps = np.load(file)
else:
_resps = np.vstack((_resps, np.load(file)))
print(state + ' resp shape: ' + str(_resps.shape))
resps[state] = _resps
# image
data = dst.CIFAR10(data_dir, download=False,
train=(state=='train'),
transform=tfs.ToTensor())
dataloader = DataLoader(data, batch_size=len(data), shuffle=False)
_images = next(iter(dataloader))[0].numpy()
print(state + ' image shape: ' + str(_images.shape))
images[state] = _images
# data
train_dataset = CustomDataset(resps['train'], images['train'])
test_dataset = CustomDataset(resps['test'], images['test'])
dataloader = DataLoader(train_dataset, batch_size=params.trainBatch, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=params.testBatch, shuffle=False)
# model
net_g = model.Generator()
net_g.to(device)
net_d = model.Discriminator()
net_d.to(device)
model1 = model.VGG()
model1.load_state_dict(torch.load('models/vgg16.pth'))
model1.to(device)
net_c = model.Comparator(model1, 22)
net_c.to(device)
for param in model1.parameters():
param.requires_grad = False
for param in net_c.parameters():
param.requires_grad = False
# optimizer
if params.optimizer_type == 'Adam':
optimizer_g = optim.Adam(net_g.parameters(), lr=params.lr,
betas=(params.beta1, params.beta2))
optimizer_d = optim.Adam(net_d.parameters(), lr=params.lr,
betas=(params.beta1, params.beta2))
elif params.optimizer_type == 'RMSprop':
optimizer_g = optim.RMSprop(net_g.parameters(), lr=params.lr,
alpha=params.beta1)
optimizer_d = optim.RMSprop(net_d.parameters(), lr=params.lr,
alpha=params.beta1)
loss_f_img = nn.MSELoss()
loss_f_img.to(device)
loss_f_feat = nn.MSELoss()
loss_f_feat.to(device)
loss_names = ['img', 'adv', 'feat', 'g', 'd']
losses = np.zeros((2, params.epochs, len(loss_names)))
for epoch in range(params.epochs):
start = time.time()
l1 = train(dataloader, epoch, params)
l2 = test(test_loader, epoch, params)
run_time = time.time() - start
losses[0, epoch, :], losses[1, epoch, :] = l1, l2
print('Epoch {}, Time {:.2f}'.format(epoch, run_time))
print(' Train: {:.6f}, {:.6f}, {:.6f}, {:.6f}, {:.6f}'.format(\
l1[0], l1[1], l1[2], l1[3], l1[4]))
print(' Test: {:.6f}, {:.6f}, {:.6f}, {:.6f}, {:.6f}'.format(\
l2[0], l2[1], l2[2], l2[3], l2[4]))
if epoch % 128 == 0:
params.lr *= params.lr_decay
# save
np.save(save_dir + 'loss_layer' + str(params.target_layer), losses)
torch.save(net_g.state_dict(), save_dir + 'net_g.pth')