-
Notifications
You must be signed in to change notification settings - Fork 3
/
loss.py
239 lines (209 loc) · 11.9 KB
/
loss.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
from torch import nn
import torch
class STDNLoss:
def __init__(self, **kwargs):
device = kwargs['device']
self.device = 'cpu' if device is None else device
self.l1_loss = nn.L1Loss()
self.l2_loss = nn.MSELoss()
self.esr_labels = kwargs['esr_labels'].split("_")
self.gan_labels = kwargs['gan_labels'].split("_")
self.disc_labels = kwargs['disc_labels'].split("_")
self.step_one = kwargs['step_one'].split("_")
self.step_two = kwargs['step_two'].split("_")
self.step_three = kwargs['step_three'].split("_")
self.spoof_trace_labels = kwargs['spoof_trace_labels'].split("_")
def _create_new_labels(self, labels, dis_features):
# print(len(dis_features))
dls, dss = dis_features
# print(f'dls: {len(dls)}, dss: {len(dss)}')
self.dl_sl = []
self.dl_rl = []
self.ds_ss = []
self.ds_rs = []
for i in range(3):
# print(f'shape of dls: {dls[i].shape}, shape of dss: {dss[i].shape}')
rl, _, sl, _ = torch.split(dls[i], len(dls[i]) // 4)
self.dl_sl.append(
sl) # only live matters and it will have 2 batchs where first batch is images and second batch is synthesized one
self.dl_rl.append(rl)
_, rs, _, ss = torch.split(dss[i], len(dss[i]) // 4)
self.ds_ss.append(ss)
self.ds_rs.append(rs)
self.spoof_labels, self.real_labels = torch.split(labels, len(labels) // 2)
# esr_loss from code
def _loss_ESR(self, esr_feats):
esr_re, esr_sp = torch.split(esr_feats, len(esr_feats) // 2)
# print(f"esr shape: {esr_re.shape}")
if self.esr_labels[0] == "0":
esr_labels_re = torch.zeros(esr_re.shape).to(self.device)
elif self.esr_labels[0] == "1":
esr_labels_re = torch.ones(esr_re.shape).to(self.device)
elif self.esr_labels[0] == "-1":
esr_labels_re = -1 * torch.ones(esr_re.shape).to(self.device)
else:
esr_labels_re = float(self.esr_labels[0]) * torch.ones(esr_re.shape).to(self.device)
if self.esr_labels[1] == "0":
esr_labels_sp = torch.zeros(esr_sp.shape).to(self.device)
elif self.esr_labels[1] == "1":
esr_labels_sp = torch.ones(esr_sp.shape).to(self.device)
elif self.esr_labels[1] == "-1":
esr_labels_sp = -1 * torch.ones(esr_sp.shape).to(self.device)
else:
esr_labels_sp = float(self.esr_labels[1]) * torch.ones(esr_sp.shape).to(self.device)
# esr_labels_sp = torch.ones(esr_sp.shape).to(self.device)
loss = self.l1_loss(esr_re, esr_labels_re) + self.l1_loss(esr_sp, esr_labels_sp)
return loss
# gan_loss from code
def _loss_G(self):
loss = 0
for i in range(3):
# print(f'gan loss: {self.dl_sl[i].shape}, {self.dl_rl[i].shape}')
# ones = torch.ones(self.dl_sl[i].shape).to(self.device)
if self.gan_labels[0] == "0":
gan_labels_re = torch.zeros(self.dl_sl[i].shape).to(self.device)
elif self.gan_labels[0] == "1":
gan_labels_re = torch.ones(self.dl_sl[i].shape).to(self.device)
elif self.gan_labels[0] == "-1":
gan_labels_re = -1 * torch.ones(self.dl_sl[i].shape).to(self.device)
else:
gan_labels_re = float(self.gan_labels[0]) * torch.ones(self.dl_sl[i].shape).to(self.device)
if self.gan_labels[1] == "0":
gan_labels_sp = torch.zeros(self.dl_sl[i].shape).to(self.device)
elif self.gan_labels[1] == "1":
gan_labels_sp = torch.ones(self.dl_sl[i].shape).to(self.device)
elif self.gan_labels[1] == "-1":
gan_labels_sp = -1 * torch.ones(self.dl_sl[i].shape).to(self.device)
else:
gan_labels_sp = float(self.gan_labels[1]) * torch.ones(self.dl_sl[i].shape).to(self.device)
loss += self.l2_loss(self.dl_sl[i], gan_labels_re) + self.l2_loss(self.ds_ss[i], gan_labels_sp)
return loss
# reg_loss from code
def _loss_R(self, spoof_trace_feats):
s, b, C, T = spoof_trace_feats
if self.spoof_trace_labels[1] == "0":
s_re = torch.zeros(s[:len(s) // 2].shape).to(self.device)
b_re = torch.zeros(b[:len(b) // 2].shape).to(self.device)
C_re = torch.zeros(C[:len(C) // 2].shape).to(self.device)
T_re = torch.zeros(T[:len(T) // 2].shape).to(self.device)
elif self.spoof_trace_labels[1] == "1":
s_re = torch.ones(s[:len(s) // 2].shape).to(self.device)
b_re = torch.ones(b[:len(b) // 2].shape).to(self.device)
C_re = torch.ones(C[:len(C) // 2].shape).to(self.device)
T_re = torch.ones(T[:len(T) // 2].shape).to(self.device)
elif self.spoof_trace_labels[1] == "-1":
s_re = -1 * torch.ones(s[:len(s) // 2].shape).to(self.device)
b_re = -1 * torch.ones(b[:len(b) // 2].shape).to(self.device)
C_re = -1 * torch.ones(C[:len(C) // 2].shape).to(self.device)
T_re = -1 * torch.ones(T[:len(T) // 2].shape).to(self.device)
else:
s_re = float(self.spoof_trace_labels[1]) * torch.ones(s[:len(s) // 2].shape).to(self.device)
b_re = float(self.spoof_trace_labels[1]) * torch.ones(b[:len(b) // 2].shape).to(self.device)
C_re = float(self.spoof_trace_labels[1]) * torch.ones(C[:len(C) // 2].shape).to(self.device)
T_re = float(self.spoof_trace_labels[1]) * torch.ones(T[:len(T) // 2].shape).to(self.device)
if self.spoof_trace_labels[1] == "0":
s_sp = torch.zeros(s[len(s) // 2:].shape).to(self.device)
b_sp = torch.zeros(b[len(b) // 2:].shape).to(self.device)
C_sp = torch.zeros(C[len(C) // 2:].shape).to(self.device)
T_sp = torch.zeros(T[len(T) // 2:].shape).to(self.device)
elif self.spoof_trace_labels[1] == "1":
s_sp = torch.ones(s[len(s) // 2:].shape).to(self.device)
b_sp = torch.ones(b[len(b) // 2:].shape).to(self.device)
C_sp = torch.ones(C[len(C) // 2:].shape).to(self.device)
T_sp = torch.ones(T[len(T) // 2:].shape).to(self.device)
elif self.spoof_trace_labels[1] == "-1":
s_sp = -1 * torch.ones(s[len(s) // 2:].shape).to(self.device)
b_sp = -1 * torch.ones(b[len(b) // 2:].shape).to(self.device)
C_sp = -1 * torch.ones(C[len(C) // 2:].shape).to(self.device)
T_sp = -1 * torch.ones(T[len(T) // 2:].shape).to(self.device)
else:
s_sp = float(self.spoof_trace_labels[1]) * torch.ones(s[len(s) // 2:].shape).to(self.device)
b_sp = float(self.spoof_trace_labels[1]) * torch.ones(b[len(b) // 2:].shape).to(self.device)
C_sp = float(self.spoof_trace_labels[1]) * torch.ones(C[len(C) // 2:].shape).to(self.device)
T_sp = float(self.spoof_trace_labels[1]) * torch.ones(T[len(T) // 2:].shape).to(self.device)
reg_loss_spoof = self.l2_loss(s[len(s) // 2:], s_sp) + self.l2_loss(b[len(b) // 2:], b_sp) + self.l2_loss(
C[len(C) // 2:], C_sp) + self.l2_loss(T[len(T) // 2:], T_sp)
reg_loss_real = self.l2_loss(s[:len(s) // 2], s_re) + self.l2_loss(b[:len(b) // 2], b_re) + self.l2_loss(
C[:len(C) // 2], C_re) + self.l2_loss(T[:len(T) // 2], T_re)
reg_loss = reg_loss_real * 10 + reg_loss_spoof * 1e-4
return reg_loss
# d_loss from code
def _loss_D(self):
loss = 0
for i in range(3):
# print(f'disc loss: {self.dl_sl[i].shape}, {self.dl_rl[i].shape}')
if self.disc_labels[0] == "0":
disc_labels_re = torch.zeros(self.dl_rl[i].shape).to(self.device)
elif self.disc_labels[0] == "1":
disc_labels_re = torch.ones(self.dl_rl[i].shape).to(self.device)
elif self.disc_labels[0] == "-1":
disc_labels_re = -1 * torch.ones(self.dl_rl[i].shape).to(self.device)
else:
disc_labels_re = float(self.disc_labels[0]) * torch.ones(self.dl_rl[i].shape).to(self.device)
if self.disc_labels[1] == "0":
disc_labels_sp = torch.zeros(self.dl_sl[i].shape).to(self.device)
elif self.disc_labels[1] == "1":
disc_labels_sp = torch.ones(self.dl_sl[i].shape).to(self.device)
elif self.disc_labels[1] == "-1":
disc_labels_sp = -1 * torch.ones(self.dl_sl[i].shape).to(self.device)
else:
disc_labels_sp = float(self.disc_labels[1]) * torch.ones(self.dl_sl[i].shape).to(self.device)
ones = torch.ones(self.dl_rl[i].shape).to(self.device)
zeros = torch.zeros(self.dl_sl[i].shape).to(self.device)
loss += self.l2_loss(self.dl_rl[i], disc_labels_re) + self.l2_loss(self.dl_sl[i],
disc_labels_sp) + self.l2_loss(
self.ds_rs[i], disc_labels_re) + self.l2_loss(self.ds_ss[i], disc_labels_sp)
return loss
# esr_loss_a from code
def _loss_ESR_syn(self, esr_feats_syn):
# print(f'synth esr loss: ', esr_feats_syn.shape)
esr_sp, esr_re = torch.split(esr_feats_syn, len(esr_feats_syn) // 2)
if self.esr_labels[0] == "0":
esr_labels_re = torch.zeros(esr_re.shape).to(self.device)
elif self.esr_labels[0] == "1":
esr_labels_re = torch.ones(esr_re.shape).to(self.device)
elif self.esr_labels[0] == "-1":
esr_labels_re = -1 * torch.ones(esr_re.shape).to(self.device)
else:
esr_labels_re = float(self.esr_labels[0]) * torch.ones(esr_re.shape).to(self.device)
if self.esr_labels[1] == "0":
esr_labels_sp = torch.zeros(esr_sp.shape).to(self.device)
elif self.esr_labels[1] == "1":
esr_labels_sp = torch.ones(esr_sp.shape).to(self.device)
elif self.esr_labels[1] == "-1":
esr_labels_sp = -1 * torch.ones(esr_sp.shape).to(self.device)
else:
esr_labels_sp = float(self.esr_labels[1]) * torch.ones(esr_sp.shape).to(self.device)
loss = self.l1_loss(esr_re, esr_labels_re) + self.l1_loss(esr_sp, esr_labels_sp)
return loss
# pixel_loss from code
def _loss_P(self, recon_syn, trace_warp_orig):
# print(f'pixel loss: {recon_syn[:len(recon_syn) // 2].shape}, {trace_warp_orig.shape}')
p = self.l1_loss(recon_syn[:len(recon_syn) // 2], trace_warp_orig)
return p
def _loss_for_one(self, esr_feats, spoof_trace_feats):
gan_loss = self._loss_G()
esr_loss = self._loss_ESR(esr_feats)
reg_loss = self._loss_R(spoof_trace_feats)
step_one_loss = esr_loss * float(self.step_one[0]) + gan_loss * float(self.step_one[1]) + reg_loss * float(
self.step_one[2])
return step_one_loss
def _loss_for_two(self):
step_two_loss = self._loss_D() * float(self.step_two[0])
return step_two_loss
def _loss_for_three(self, esr_feats_syn, recon_syn, trace_warp_orig):
synth_loss = self._loss_ESR_syn(esr_feats_syn)
pixel_loss = self._loss_P(recon_syn, trace_warp_orig)
step_three_loss = synth_loss * float(self.step_three[0]) + pixel_loss * float(self.step_three[1])
return step_three_loss
# esr_feat, spoof_trace, dis_features, esr_feat_syn, spoof_trace_syn, recon_syn, trace_warp_orig
def __call__(self, labels, esr_feats, trace_feats, dis_features, esr_feats_syn, trace_feats_syn, recon_syn,
trace_warp_orig, trace_unwarp_orig, synth):
self._create_new_labels(labels, dis_features)
step_1 = self._loss_for_one(esr_feats, trace_feats)
step_2 = self._loss_for_two()
step_3 = self._loss_for_three(esr_feats_syn, recon_syn, trace_warp_orig)
gan_loss = step_3 + step_1
disc_loss = step_2
sum_loss = step_1 + step_2 + step_3
return sum_loss, gan_loss, disc_loss