-
Notifications
You must be signed in to change notification settings - Fork 3
/
attack_dcgan.py
447 lines (396 loc) · 18.3 KB
/
attack_dcgan.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
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
from __future__ import print_function
import sys
import argparse
import os
import numpy as np
import pandas
from tqdm import tqdm
import matplotlib.pylab as plt
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import torch.optim as optim
import torchvision.utils as vutils
import utils
from utils import mkdir, gaussian_logp
from csv_logger import CSVLogger, plot_csv
from main_aux import save_checkpoint, maybe_load_checkpoint
from experimental import AttackExperiment
from likelihood_model import ReparameterizedMVN, FlowMiner, ReparameterizedGMM2
class LabelSmoothingLoss(nn.Module):
def __init__(self, n_classes, smoothing=0.0, dim=-1):
super(LabelSmoothingLoss, self).__init__()
self.confidence = 1.0 - smoothing
self.smoothing = smoothing
self.n_classes = n_classes
self.dim = dim
def forward(self, lsm, target):
true_dist = torch.zeros_like(lsm)
true_dist.fill_(self.smoothing / (self.n_classes - 1))
true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
return torch.mean(torch.sum(-true_dist * lsm, dim=self.dim))
class Miner(nn.Module):
def __init__(self, nz, nz0, nh):
super(Miner, self).__init__()
self.nz = nz
self.nz0 = nz0
self.nh = nh
layers_ = [
nn.Linear(nz0, nh),
nn.BatchNorm1d(nh),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(nh, nh),
nn.BatchNorm1d(nh),
nn.LeakyReLU(0.2, inplace=True),
nn.Linear(nh, nz)
]
self.main = nn.Sequential(*layers_)
def forward(self, x):
return self.main(x).squeeze(-1).squeeze(-1)
class MineGAN(nn.Module):
def __init__(self, miner, generator):
super(MineGAN, self).__init__()
self.nz = miner.nz0
self.is_conditional = generator.is_conditional
self.miner = miner
self.generator = generator
def forward(self, z0, c=None):
z = self.miner(z0.squeeze(-1).squeeze(-1))
if c is not None: # AuxGAN
x = self.generator(z, c)
else: # GAN
x = self.generator(z)
return x
def main(args):
args.ckpt = os.path.join(args.output_dir, "ckpt.pt")
# db config
if args.db:
pass
# backward compat
# Experiment setup
experiment = AttackExperiment(args.exp_config, device, args.db,
fixed_id=args.fixed_id, run_target_feat_eval=args.run_target_feat_eval)
target_logsoftmax = experiment.target_logsoftmax
target_dataset = experiment.target_dataset
target_eval_runner = experiment.target_eval_runner
generator = experiment.generator
nclass = experiment.target_dataset['nclass']
gan_method = experiment.gan_method
if args.method == 'minegan':
miner = ReparameterizedMVN(generator.nz).to(device)
generator = MineGAN(miner, generator)
elif args.method == 'gmm':
miner = ReparameterizedGMM2(generator.nz, args.gmm_n_components).to(device)
generator = MineGAN(miner, generator)
elif args.method == 'flow':
miner = FlowMiner(generator.nz, args.flow_permutation,
args.flow_K, args.flow_glow, args.flow_coupling, args.flow_L, args.flow_use_actnorm).to(device)
generator = MineGAN(miner, generator)
# Opt
optimizerG = optim.SGD(miner.parameters(), lr=args.lr,
momentum=0.9, weight_decay=args.wd)
# Logging
iteration_fieldnames = ['global_iteration', 'loss', 'train_target_acc']
iteration_logger = CSVLogger(every=args.log_iter_every,
fieldnames=iteration_fieldnames,
filename=os.path.join(
args.output_dir, 'iteration_log.csv'),
resume=args.resume)
epoch_fieldnames = ['global_iteration',
'eval-acc-marginal',
'eval-frechet-marginal',
'eval-feature-l2-dist-marginal',
'eval-feature-cos-sim-marginal',
'eval-top5_acc-marginal',
]
if args.run_target_feat_eval:
epoch_fieldnames += [
'eval-precision@5-marginal',
'eval-recall@5-marginal',
'eval-precision@10-marginal',
'eval-recall@10-marginal',
]
epoch_logger = CSVLogger(every=args.log_epoch_every,
fieldnames=epoch_fieldnames,
filename=os.path.join(
args.output_dir, 'epoch_log.csv'),
resume=args.resume)
# Check for ckpt
ckpt = maybe_load_checkpoint(args)
if ckpt is not None:
start_epoch = ckpt['epoch']
optimizerG.load_state_dict(ckpt['optimizerG'])
generator.load_state_dict(ckpt['generator'])
else:
start_epoch = 0
patience_count = 0
best_marginal_acc = 0
fixed_noise = torch.randn(500, generator.nz, 1, 1, device=device)
attack_criterion = LabelSmoothingLoss(
nclass, smoothing=args.attack_labelsmooth)
save_model_epochs = [int(e) for e in args.save_model_epochs.split(
',')] if len(args.save_model_epochs) > 0 else []
for epoch in range(start_epoch, args.epochs):
noises = torch.randn(1000, generator.nz, 1, 1, device='cpu')
# Ckpt
state = {
"optimizerG": optimizerG.state_dict(),
"generator": generator.state_dict(),
"epoch": epoch,
}
save_checkpoint(args, state)
# Save Models
if epoch in save_model_epochs:
torch.save(generator.state_dict(), os.path.join(args.output_dir, f'generator_{epoch}.pt'))
torch.save(miner.state_dict(), os.path.join(args.output_dir, f'miner_{epoch}.pt'))
if epoch > 0 and epoch % args.save_samples_every == 0:
with torch.no_grad():
fake = generator(fixed_noise)
torch.save(fake[:args.n_save_samples], os.path.join(args.output_dir, f'samples_e{epoch}.pt'))
# Evaluate
# - Sample conditionally
all_ys = torch.arange(1000)
fakes = []
for start in range(0, 1000, 100):
with torch.no_grad():
noise = noises[start:start + 100].to(device)
if gan_method == 'dcgan_aux':
fake_y = torch.ones((100,)) * args.fixed_id
fake_y_onehot = torch.eye(nclass)[fake_y.long()].to(device)
fake = generator(noise, fake_y_onehot)
else:
if generator.is_conditional:
fake_y = all_ys[start:start + 100].to(device)
fake = generator(noise, fake_y)
else:
fake = generator(noise)
fakes.append(fake)
fakes = torch.cat(fakes)
# - Run eval
epoch_log_dict = {'global_iteration': epoch}
name = 'marginal'
fake = fakes[:100]
fake_y = args.fixed_id * torch.ones(len(fake)).to(device).long()
D = target_eval_runner.evaluate(fake, fake_y, None)
for field in D:
if not ("eval-" + field + f"-{name}" in epoch_fieldnames):
continue
epoch_log_dict["eval-" + field + f"-{name}"] = D[field]
epoch_logger.writerow(epoch_log_dict)
if len(epoch_log_dict) > 1:
plot_csv(epoch_logger.filename, os.path.join(
args.output_dir, 'epoch_plots.jpeg'))
# Maybe exit
if epoch_log_dict['eval-acc-marginal'] > best_marginal_acc:
patience_count = 0
best_marginal_acc = epoch_log_dict['eval-acc-marginal']
else:
patience_count += 1
if patience_count >= args.patience:
print("Patience exceeded, exiting")
sys.exit(0)
# Visualize samples
if epoch % args.viz_every == 0:
def _viz_with_corresponding_preds(fake, fpath):
with torch.no_grad():
preds = target_logsoftmax(fake / 2 + .5).max(1)[1]
real_target = []
for c in preds.cpu():
real_target.append(
target_dataset['X_train'][[target_dataset['Y_train'] == c]][0])
real_target = torch.stack(real_target)
preds = target_eval_runner.get_eval_preds(fake)
real_eval = []
for c in preds.cpu():
real_eval.append(target_dataset['X_train'][[
target_dataset['Y_train'] == c]][0])
real_eval = torch.stack(real_eval)
realgrid_target = vutils.make_grid(
real_target[:100], nrow=10, padding=4, pad_value=1, normalize=True)
realgrid_eval = vutils.make_grid(
real_eval[:100], nrow=10, padding=4, pad_value=1, normalize=True)
fakegrid = vutils.make_grid(
fake.cpu()[:100], nrow=10, padding=4, pad_value=1, normalize=True)
fig, axs = plt.subplots(1, 3, figsize=(20, 12))
axs[0].imshow(np.transpose(
realgrid_eval.cpu().numpy(), (1, 2, 0)), interpolation='bilinear')
axs[0].set_title('Real Eval pred')
axs[1].imshow(np.transpose(fakegrid.cpu().numpy(),
(1, 2, 0)), interpolation='bilinear')
axs[1].set_title('Samples')
axs[2].imshow(np.transpose(realgrid_target.cpu(
).numpy(), (1, 2, 0)), interpolation='bilinear')
axs[2].set_title('Real Target pred')
for ax in axs:
plt.subplot(ax)
plt.tight_layout()
plt.grid()
plt.xticks([])
plt.yticks([])
plt.savefig(fpath, bbox_inches='tight',
pad_inches=0, format='jpeg')
# Marginal samples
with torch.no_grad():
if gan_method == 'dcgan_aux':
fake_y = torch.ones((len(fixed_noise),)) * args.fixed_id
fake_y_onehot = torch.eye(nclass)[fake_y.long()].to(device)
fake = generator(fixed_noise, fake_y_onehot)
else:
fake = generator(fixed_noise).detach()
fpath = f'{args.output_dir}/viz_sample/sample_e{epoch:03d}_marginal.jpeg'
if experiment.config['data']['name'] == 'celeba':
_viz_with_corresponding_preds(fake[:100], fpath)
else:
fake = fake[:100]
fakegrid = vutils.make_grid(
fake.cpu()[:100], nrow=10, padding=4, pad_value=1, normalize=True)
fig, ax = plt.subplots(1, 1, figsize=(10, 10))
ax.imshow(np.transpose(
fakegrid.cpu().numpy(), (1, 2, 0)), interpolation='bilinear')
plt.tight_layout()
plt.grid()
plt.xticks([])
plt.yticks([])
plt.savefig(fpath, bbox_inches='tight',
pad_inches=0, format='jpeg')
# Save Sample Tensors
if epoch % 10 == 0:
torch.save(fake[:50].cpu(), os.path.join(args.output_dir, 'samples_pt', f'e{epoch:03d}.pt'))
# Train loop
generator.train()
pbar = tqdm(range(0, 10000, args.batchSize), desc='Train loop')
for i in pbar:
generator.zero_grad()
# Sample from G
noise = torch.randn(
args.batchSize, generator.nz, 1, 1, device=device)
if gan_method == 'dcgan_aux':
fake_y = torch.ones((args.batchSize,)) * args.fixed_id
fake_y_onehot = torch.eye(nclass)[fake_y.long()].to(device)
fake = generator(noise, fake_y_onehot)
else:
if generator.is_conditional:
fake_y = torch.randint(1000, (args.batchSize,)).to(device)
fake = generator(noise, fake_y)
else:
fake = generator(noise)
# Compute loss
lsm = target_logsoftmax(fake / 2 + .5)
fake_y = args.fixed_id * \
torch.ones(args.batchSize).to(device).long()
loss_attack = 0
if args.lambda_attack > 0:
# loss_attack = -lsm.gather(1, fake_y.view(-1,1)).mean()
loss_attack = attack_criterion(lsm, fake_y)
train_target_acc = (lsm.max(1)[1] == fake_y).float().mean().item()
loss_kl = 0
# if True:
if args.lambda_kl > 0:
if args.method == 'minegan':
C = miner.L @ miner.L.T
logdetcov = torch.logdet(C)
samples = miner(torch.randn(
1000, miner.nz0).to(device))
loss_kl = -.5 * logdetcov + .5 * \
(torch.norm(samples, p=2, dim=[-1])).pow(2).mean()
else:
# KL(Flow || N(0,1))
# E_{x ~ Flow}[ log Flow(x) - log N(x; 0,1)]
samples = miner(torch.randn(
args.batchSize, miner.nz0).to(device))
loss_kl = torch.mean(miner.logp(
samples) - gaussian_logp(torch.zeros_like(samples), torch.zeros_like(samples), samples).sum(-1))
loss = (args.lambda_attack * loss_attack
+ args.lambda_kl * loss_kl)
loss.backward()
optimizerG.step()
# Logging
pbar.set_postfix_str(s=f'Loss: {loss.item():.2f}, Acc: {train_target_acc:.3f}', refresh=True)
if i % args.log_iter_every == 0:
stats_dict = {
'global_iteration': iteration_logger.time,
'loss': loss.item(),
'train_target_acc': train_target_acc
}
iteration_logger.writerow(stats_dict)
plot_csv(iteration_logger.filename, os.path.join(
args.output_dir, 'iteration_plots.jpeg'))
iteration_logger.time += 1
if __name__ == '__main__':
import socket
parser = argparse.ArgumentParser()
parser.add_argument('--overwrite', type=int, default=1)
parser.add_argument('--exp_config', type=str, required=True)
parser.add_argument('--patience', type=int, default=20)
parser.add_argument('--save_model_epochs', type=str, default='')
parser.add_argument('--method', type=str,
default='finetune', choices=['minegan', 'flow', 'gmm'])
parser.add_argument('--run_target_feat_eval', type=int, default=0)
parser.add_argument('--attack_labelsmooth', type=float, default=0)
# Miner
parser.add_argument('--miner_nh', type=int, default=100)
parser.add_argument('--miner_z0', type=int, default=50)
parser.add_argument('--miner_init_std', type=float, default=0.2)
# EWC
parser.add_argument('--fixed_id', type=int, default=0)
parser.add_argument('--ewc_type', type=str, default='fisher')
parser.add_argument('--lambda_attack', type=float, default=1)
parser.add_argument('--lambda_kl', type=float, default=0)
parser.add_argument('--lambda_miner_entropy', type=float, default=0)
parser.add_argument('--prior_model', type=str, default='disc',
choices=['disc', 'lep', 'tep', '0', 'hep'])
parser.add_argument('--lep_path', type=str, default='')
parser.add_argument('--flow_permutation', type=str,
default='shuffle', choices=['shuffle', 'reverse'])
parser.add_argument('--flow_K', type=int, default=5)
parser.add_argument('--flow_glow', type=int, default=0)
parser.add_argument('--flow_coupling', type=str, default='additive', choices= ['additive', 'affine', 'invconv'])
parser.add_argument('--flow_L', type=int, default=1)
parser.add_argument('--flow_use_actnorm', type=int, default=1)
# GMM
parser.add_argument('--gmm_n_components', type=int, default=1)
# Optimization arguments
parser.add_argument('--batchSize', type=int,
default=64, help='input batch size')
parser.add_argument('--epochs', type=int, default=1000,
help='number of epochs to train for')
parser.add_argument('--lr', type=float, default=0.0002,
help='learning rate, default=0.0002')
parser.add_argument('--beta1', type=float,
default=0.5, help='beta1 for adam')
parser.add_argument('--wd', type=float, default=0., help='wd for adam')
parser.add_argument('--seed', type=int, default=2019, help='manual seed')
# Checkpointing and Logging arguments
parser.add_argument('--output_dir', required=True, help='')
parser.add_argument('--save_samples_every', type=int, default=10000)
parser.add_argument('--log_iter_every', type=int, default=100)
parser.add_argument('--viz_every', type=int, default=10)
parser.add_argument('--log_epoch_every', type=int, default=1)
parser.add_argument('--resume', type=int, required=True)
parser.add_argument('--user', type=str, default='wangkuan')
parser.add_argument('--n_save_samples', type=int, default=100)
# Dev
parser.add_argument('--db', type=int, default=0)
args = parser.parse_args()
if not args.overwrite and os.path.exists(args.output_dir):
# Check if the previous experiment ran for more than 10 epochs.
if os.path.exists(os.path.join(args.output_dir, 'epoch_log.csv')):
df = pandas.read_csv(os.path.join(
args.output_dir, 'epoch_log.csv'))
if len(df) > 10:
sys.exit(0)
# Discs
mkdir(args.output_dir)
mkdir(os.path.join(args.output_dir, 'viz_sample'))
mkdir(os.path.join(args.output_dir, 'samples_pt'))
args.jobid = os.environ['SLURM_JOB_ID'] if 'SLURM_JOB_ID' in os.environ else -1
args.host = socket.gethostname()
utils.save_args(args, os.path.join(args.output_dir, 'args.json'))
# Global Config
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args.device = device
np.random.seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.manual_seed(args.seed)
cudnn.benchmark = True
main(args)