-
Notifications
You must be signed in to change notification settings - Fork 0
/
solver.py
327 lines (300 loc) · 12 KB
/
solver.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
""" Training implementation for IDEAW
"""
import torch
import torch.nn as nn
import yaml
from data.dataset import AWdataset, get_data_loader, infinite_iter
from models.ideaw import IDEAW
from metrics import calc_acc, signal_noise_ratio
class Solver(object):
def __init__(self, config_data_path, config_model_path, args):
self.config_data_path = config_data_path
self.config_model_path = config_model_path
self.args = args # training config inside
self.device = self.args.device
with open(self.config_data_path) as f:
self.config_data = yaml.load(f, Loader=yaml.FullLoader)
with open(self.config_model_path) as f:
self.config_model = yaml.load(f, Loader=yaml.FullLoader)
with open(self.args.train_config) as f:
self.config_t = yaml.load(f, Loader=yaml.FullLoader)
self.get_inf_train_iter() # prepare data
self.build_model() # prepare model
self.build_optims() # prepare optimizers
self.loss_criterion() # prepare criterions
if self.args.load_model:
self.load_model()
# prepare data, called in Solver.init
def get_inf_train_iter(self):
dataset_dir = self.args.pickle_path
self.dataset = AWdataset(dataset_dir)
self.batch_size = self.config_t["train"]["batch_size"]
self.num_workers = self.config_t["train"]["num_workers"]
self.stage1_ratio = self.config_t["train"]["stage_I_ratio"]
self.shift_ratio = self.config_t["train"]["shift_ratio"]
self.train_iter = infinite_iter(
get_data_loader(
dataset=self.dataset,
batch_size=self.batch_size,
num_workers=self.num_workers,
)
)
print("[IDEAW]infinite dataloader built")
return
# load model/data to cuda
def cc(self, net):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
return net.to(device)
# called in Solver.init
def build_model(self):
self.model = self.cc(IDEAW(self.config_model_path, self.device))
print("[IDEAW]model built")
print(
"[IDEAW]total parameter count: {}".format(
sum(x.numel() for x in self.model.parameters())
)
)
# called in Solver.init
def build_optims(self):
param_hinet1 = list(
filter(lambda p: p.requires_grad, self.model.hinet_1.parameters())
)
param_hinet2 = list(
filter(lambda p: p.requires_grad, self.model.hinet_2.parameters())
)
param_discr = list(
filter(lambda p: p.requires_grad, self.model.discriminator.parameters())
)
param_att = list(
filter(lambda p: p.requires_grad, self.model.attack_layer.parameters())
)
param_balance = list(
filter(lambda p: p.requires_grad, self.model.balance_block.parameters())
)
lr1 = eval(self.config_t["train"]["lr1"])
lr2 = eval(self.config_t["train"]["lr2"])
beta1 = self.config_t["train"]["beta1"]
beta2 = self.config_t["train"]["beta2"]
eps = eval(self.config_t["train"]["eps"])
weight_decay = eval(self.config_t["train"]["weight_decay"])
self.optim_I = torch.optim.Adam(
param_hinet1 + param_hinet2,
lr=lr1,
betas=(beta1, beta2),
eps=eps,
weight_decay=weight_decay,
)
self.optim_II = torch.optim.Adam(
param_att + param_balance,
lr=lr2,
betas=(beta1, beta2),
eps=eps,
weight_decay=weight_decay,
)
self.optim_III = torch.optim.Adam(
param_discr,
lr=lr1,
betas=(beta1, beta2),
eps=eps,
weight_decay=weight_decay,
)
self.weight_scheduler1 = torch.optim.lr_scheduler.StepLR(
self.optim_I,
self.config_t["train"]["weight_step"],
gamma=self.config_t["train"]["gamma"],
)
self.weight_scheduler2 = torch.optim.lr_scheduler.StepLR(
self.optim_II,
self.config_t["train"]["weight_step"],
gamma=self.config_t["train"]["gamma"],
)
self.weight_scheduler3 = torch.optim.lr_scheduler.StepLR(
self.optim_III,
self.config_t["train"]["weight_step"],
gamma=self.config_t["train"]["gamma"],
)
print("[IDEAW]optimizers built")
# autosave, called in training
def save_model(self, robustness):
if robustness:
torch.save(
self.model.state_dict(),
f"{self.args.store_model_path}stage_II/ideaw.ckpt",
)
torch.save(
self.optim_I.state_dict(),
f"{self.args.store_model_path}stage_II/optim1.opt",
)
torch.save(
self.optim_II.state_dict(),
f"{self.args.store_model_path}stage_II/optim2.opt",
)
torch.save(
self.optim_III.state_dict(),
f"{self.args.store_model_path}stage_II/optim3.opt",
)
else:
torch.save(
self.model.state_dict(),
f"{self.args.store_model_path}stage_I/ideaw.ckpt",
)
torch.save(
self.optim_I.state_dict(),
f"{self.args.store_model_path}stage_I/optim1.opt",
)
torch.save(
self.optim_II.state_dict(),
f"{self.args.store_model_path}stage_I/optim2.opt",
)
torch.save(
self.optim_III.state_dict(),
f"{self.args.store_model_path}stage_I/optim3.opt",
)
# load trained model
def load_model(self):
print(f"[IDEAW]load model from {self.args.load_model_path}")
self.model.load_state_dict(torch.load(f"{self.args.load_model_path}ideaw.ckpt"))
self.optim_I.load_state_dict(
torch.load(f"{self.args.load_model_path}optim1.opt")
)
self.optim_II.load_state_dict(
torch.load(f"{self.args.load_model_path}optim2.opt")
)
self.optim_III.load_state_dict(
torch.load(f"{self.args.load_model_path}optim3.opt")
)
return
# loss criterion
def loss_criterion(self):
self.criterion_percept = nn.MSELoss()
self.criterion_integ = nn.MSELoss()
self.criterion_discr = nn.BCELoss()
# training
def train(self, n_iterations):
print("[IDEAW]starting training...")
self.lambda_1 = self.config_t["train"]["lambda_integ"]
self.lambda_2 = self.config_t["train"]["lambda_percept"]
self.lambda_3 = self.config_t["train"]["lambda_ident"]
percept_loss_history = []
integ_loss_history = []
discr_loss_history = []
ident_loss_history = []
for iter in range(n_iterations):
# get data for current iteration
host_audio = next(self.train_iter).to(torch.float32)
msg_len = self.config_model["IDEAW"]["num_bit"]
lcode_len = self.config_model["IDEAW"]["num_lc_bit"]
watermark_msg = torch.randint(
0, 2, (self.batch_size, msg_len), dtype=torch.float32
)
locate_code = torch.randint(
0, 2, (self.batch_size, lcode_len), dtype=torch.float32
)
orig_label = torch.ones((self.batch_size, 1))
wmd_label = torch.zeros((self.batch_size, 1))
## load to cuda
host_audio = self.cc(host_audio)
watermark_msg = self.cc(watermark_msg)
locate_code = self.cc(locate_code)
orig_label = self.cc(orig_label)
wmd_label = self.cc(wmd_label)
# forward
## stage I training
if iter < n_iterations * self.stage1_ratio:
robustness = False
## stage II training (robustness training)
else:
robustness = True
## locating stripe adaptive training strategy
if iter < n_iterations * self.shift_ratio:
shift = False
else:
shift = True
(
_,
audio_wmd1_stft,
audio_wmd2,
audio_wmd2_stft,
msg_extr1,
msg_extr2,
lcode_extr,
orig_output,
wmd_output,
) = self.model(host_audio, watermark_msg, locate_code, robustness, shift)
# loss
## percept. loss
host_audio_stft = self.model.stft(host_audio)
percept_loss_1 = self.criterion_percept(host_audio_stft, audio_wmd1_stft)
percept_loss_2 = self.criterion_percept(host_audio_stft, audio_wmd2_stft)
percept_loss_3 = self.criterion_percept(audio_wmd1_stft, audio_wmd2_stft)
percept_loss = percept_loss_1 + percept_loss_2 + percept_loss_3
percept_loss_history.append(percept_loss.item())
## integrity loss
integ_loss_1 = self.criterion_integ(watermark_msg, msg_extr1)
integ_loss_2 = self.criterion_integ(watermark_msg, msg_extr2)
integ_loss_3 = self.criterion_integ(locate_code, lcode_extr)
integ_loss = integ_loss_1 + integ_loss_2 + integ_loss_3
integ_loss_history.append(integ_loss.item())
## discriminate loss
discr_loss_orig = self.criterion_discr(orig_output, orig_label)
discr_loss_wmd = self.criterion_discr(wmd_output, wmd_label)
discr_loss = discr_loss_orig + discr_loss_wmd
discr_loss_history.append(discr_loss.item())
## identify loss
ident_loss = -torch.sum(torch.log(1 - wmd_output))
ident_loss_history.append(ident_loss)
## total loss
discr_frozen = True
if iter % 2 == 1:
discr_frozen = False
total_loss = (
self.lambda_1 * integ_loss
+ self.lambda_2 * percept_loss
+ self.lambda_3 * discr_loss
)
else:
discr_frozen = True
total_loss = (
self.lambda_1 * integ_loss
+ self.lambda_2 * percept_loss
+ self.lambda_3 * ident_loss
)
# metric
acc_msg = calc_acc(msg_extr2, watermark_msg, 0.5)
acc_lcode = calc_acc(lcode_extr, locate_code, 0.5)
snr = signal_noise_ratio(host_audio, audio_wmd2)
# backward
total_loss.backward()
if eval(self.config_t["train"]["optim1_step"]):
self.optim_I.step()
self.weight_scheduler1.step()
if eval(self.config_t["train"]["optim2_step"]):
self.optim_II.step()
self.weight_scheduler2.step()
if discr_frozen == False:
self.optim_III.step()
self.weight_scheduler3.step()
self.optim_I.zero_grad()
self.optim_II.zero_grad()
self.optim_III.zero_grad()
# logging
print(
f"[IDEAW]:[{iter+1}/{n_iterations}]",
f"Robustness={robustness}",
f"shift={shift}",
f"loss_percept={percept_loss.item():.6f}",
f"loss_integ={integ_loss.item():6f}",
f"loss_discr={discr_loss.item():6f}",
f"loss_ident={ident_loss.item():6f}",
f"SNR={snr:4f}",
f"acc_msg={acc_msg:4f}",
f"acc_lcode={acc_lcode:4f}",
end="\r",
)
# summary
if (iter + 1) % self.args.summary_steps == 0 or iter + 1 == n_iterations:
print()
# autosave
if (iter + 1) % self.args.save_steps == 0 or iter + 1 == n_iterations:
self.save_model(robustness)
return