-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathcustom_criterions.py
357 lines (292 loc) · 15.1 KB
/
custom_criterions.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
import torch.nn as nn
import torch
from netVggs.netVgg_conv3 import netVgg_conv3
from netVggs.netVgg_conv4 import netVgg_conv4
import pdb
from stn_module import STN
from opts import opt
class MaskedMSELoss(nn.Module):
def __init__(self, reduction=None):
super(MaskedMSELoss, self).__init__()
if reduction:
self.criterion = nn.MSELoss(reduction=reduction)
else:
self.criterion = nn.MSELoss()
def forward(self, ipt, tgt, mask):
self.loss = self.criterion(ipt * mask, tgt)
return self.loss
# https://github.com/jxgu1016/Total_Variation_Loss.pytorch/blob/master/TVLoss.py
class TVLoss(nn.Module):
def __init__(self):
super(TVLoss, self).__init__()
def forward(self, x):
batch_size = x.size()[0]
h_x = x.size()[2]
w_x = x.size()[3]
count_h = self._tensor_size(x[:,:,1:,:])
count_w = self._tensor_size(x[:,:,:,1:])
h_tv = torch.pow((x[:,:,1:,:]-x[:,:,:h_x-1,:]),2).sum()
w_tv = torch.pow((x[:,:,:,1:]-x[:,:,:,:w_x-1]),2).sum()
return (h_tv/count_h+w_tv/count_w)/batch_size
def _tensor_size(self,t):
return t.size()[1]*t.size()[2]*t.size()[3]
class SymLoss(nn.Module):
def __init__(self, C):
super(SymLoss, self).__init__()
self.C = C
def forward(self, grid, sym_axis):
# grid (N,2,H,W)
batch_size = grid.size()[0]
h = grid.size()[2]
w = grid.size()[3]
sym_x = sym_axis[:,0].view(batch_size, 1, 1)
sym_y = sym_axis[:,1].view(batch_size, 1, 1)
delta_grid_x = grid[:,0,:h-self.C,:] - grid[:,0,self.C:,:]
delta_grid_y = grid[:,1,:h-self.C,:] - grid[:,1,self.C:,:]
sym_loss = torch.pow(delta_grid_x * sym_y + delta_grid_y * sym_x, 2).sum()
return sym_loss / ((h - self.C) * w * batch_size)
class FullSymLoss(nn.Module):
def __init__(self, C):
super(FullSymLoss, self).__init__()
self.C = C
def forward(self, grid, gt_sym_axis, gd_sym_axis):
# grid (N,2,H,W)
batch_size = grid.size()[0]
h = grid.size()[2]
w = grid.size()[3]
gt_sym_x = gt_sym_axis[:,0].view(batch_size)
gt_sym_y = gt_sym_axis[:,1].view(batch_size)
gd_sym_x = gd_sym_axis[:,0].view(batch_size)
gd_sym_y = gd_sym_axis[:,1].view(batch_size)
dxs = (-self.C * gt_sym_x).round().int()
dys = (self.C * gt_sym_y).round().int()
# pdb.set_trace()
sym_loss = 0
for b_i, (dx, dy, sym_x, sym_y) in enumerate(zip(dxs, dys, gd_sym_x, gd_sym_y)):
# dx > 0
# pdb.set_trace()
if dx > 0:
# print ('dx > 0')
# print (dx, dy)
# print (h, w)
# print (grid[b_i,0,:h-dy,:w-dx].shape)
# print (grid[b_i,0,dy:,dx:].shape)
delta_grid_x = grid[b_i,0,:h-dy,:w-dx] - grid[b_i,0,dy:,dx:]
delta_grid_y = grid[b_i,1,:h-dy,:w-dx] - grid[b_i,1,dy:,dx:]
else: # dx < 0
# print ('dx < 0')
# print (grid[b_i,0,dy:,:w+dx].shape)
# print (grid[b_i,1,:h-dy,-dx:].shape)
delta_grid_x = grid[b_i,0,dy:,:w+dx] - grid[b_i,0,:h-dy,-dx:]
delta_grid_y = grid[b_i,1,dy:,:w+dx] - grid[b_i,1,:h-dy,-dx:]
# print ('over', b_i)
# print (delta_grid_x.shape)
# print (delta_grid_y.shape)
# pdb.set_trace()
sym_loss += torch.pow(delta_grid_x * sym_y + delta_grid_y * sym_x, 2).mean()
# delta_grid_x = grid[:,0,:h-self.C,:] - grid[:,0,self.C:,:]
# delta_grid_y = grid[:,1,:h-self.C,:] - grid[:,1,self.C:,:]
# sym_loss = torch.pow(delta_grid_x * gd_sym_y + delta_grid_y * gd_sym_x, 2).sum()
# pdb.set_trace()
return sym_loss / batch_size
class BilinearFullSymLoss(nn.Module):
def __init__(self, C):
super(BilinearFullSymLoss, self).__init__()
self.C = C
def forward(self, grid, gt_sym_axis, gd_sym_axis):
# grid (N,2,H,W)
batch_size = grid.size()[0]
h = grid.size()[2]
w = grid.size()[3]
gt_sym_x = gt_sym_axis[:,0].view(batch_size)
gt_sym_y = gt_sym_axis[:,1].view(batch_size)
gd_sym_x = gd_sym_axis[:,0].view(batch_size)
gd_sym_y = gd_sym_axis[:,1].view(batch_size)
dxs = (-self.C * gt_sym_x)
dys = (self.C * gt_sym_y)
# pdb.set_trace()
sym_loss = 0
for b_i, (dx, dy, sym_x, sym_y) in enumerate(zip(dxs, dys, gd_sym_x, gd_sym_y)):
# dx > 0
dy1_f = dy.floor()
dy2_f = dy1_f + 1
dy1 = dy.floor().int()
dy2 = dy1 + 1
dx1_f = dx.floor()
dx2_f = dx1_f + 1
dx1 = dx.floor().int()
dx2 = dx1 + 1
if dx > 0:
# print ('dx > 0')
# print (dx, dy)
# print (h, w)
# print (grid[b_i,0,:h-dy,:w-dx].shape)
# print (grid[b_i,0,dy:,dx:].shape)
# x2,y1 21
# x1 or y1, 会导致多一列/行 (最后一列/行)
delta_grid_x_11 = grid[b_i,0,:h-dy1-1,:w-dx1-1] - grid[b_i,0,dy1:-1,dx1:-1]
delta_grid_y_11 = grid[b_i,1,:h-dy1-1,:w-dx1-1] - grid[b_i,1,dy1:-1,dx1:-1]
delta_grid_x_21 = grid[b_i,0,:h-dy1-1,:w-dx2] - grid[b_i,0,dy1:-1,dx2:]
delta_grid_y_21 = grid[b_i,1,:h-dy1-1,:w-dx2] - grid[b_i,1,dy1:-1,dx2:]
delta_grid_x_12 = grid[b_i,0,:h-dy2,:w-dx1-1] - grid[b_i,0,dy2:,dx1:-1]
delta_grid_y_12 = grid[b_i,1,:h-dy2,:w-dx1-1] - grid[b_i,1,dy2:,dx1:-1]
delta_grid_x_22 = grid[b_i,0,:h-dy2,:w-dx2] - grid[b_i,0,dy2:,dx2:]
delta_grid_y_22 = grid[b_i,1,:h-dy2,:w-dx2] - grid[b_i,1,dy2:,dx2:]
# pdb.set_trace()
# delta_grid_x = (dx - dx1_f) * (dy - dy1_f) * delta_grid_x_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_x_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_x_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_x_11
# delta_grid_y = (dx - dx1_f) * (dy - dy1_f) * delta_grid_y_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_y_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_y_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_y_11
else: # dx <= 0
# print ('dx < 0')
# print (grid[b_i,0,dy:,:w+dx].shape)
# print (grid[b_i,1,:h-dy,-dx:].shape)
# x2 的第一列不能要
# y1 的第一行不能要
delta_grid_x_11 = grid[b_i,0,dy1+1:,:w+dx1] - grid[b_i,0,1:h-dy1,-dx1:]
delta_grid_y_11 = grid[b_i,1,dy1+1:,:w+dx1] - grid[b_i,1,1:h-dy1,-dx1:]
delta_grid_x_21 = grid[b_i,0,dy1+1:,1:w+dx2] - grid[b_i,0,1:h-dy1,-dx2+1:]
delta_grid_y_21 = grid[b_i,1,dy1+1:,1:w+dx2] - grid[b_i,1,1:h-dy1,-dx2+1:]
delta_grid_x_12 = grid[b_i,0,dy2:,:w+dx1] - grid[b_i,0,:h-dy2,-dx1:]
delta_grid_y_12 = grid[b_i,1,dy2:,:w+dx1] - grid[b_i,1,:h-dy2,-dx1:]
delta_grid_x_22 = grid[b_i,0,dy2:,1:w+dx2] - grid[b_i,0,:h-dy2,-dx2+1:]
delta_grid_y_22 = grid[b_i,1,dy2:,1:w+dx2] - grid[b_i,1,:h-dy2,-dx2+1:]
# pdb.set_trace()
# 大小和对称轴的歪转程度有关系
# dx <= 0 : torch.Size([246, 255])
# dx > 0 : torch.Size([246, 255])
# dx > 0: torch.Size([246, 254]) seed 5222
# dx < 0: torch.Size([247, 251]) seed 5221
delta_grid_x = (dx - dx1_f) * (dy - dy1_f) * delta_grid_x_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_x_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_x_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_x_11
delta_grid_y = (dx - dx1_f) * (dy - dy1_f) * delta_grid_y_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_y_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_y_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_y_11
# print ('over', b_i)
# print (delta_grid_x.shape)
# print (delta_grid_y.shape)
# pdb.set_trace()
sym_loss += torch.pow(delta_grid_x * sym_y + delta_grid_y * sym_x, 2).mean()
# delta_grid_x = grid[:,0,:h-self.C,:] - grid[:,0,self.C:,:]
# delta_grid_y = grid[:,1,:h-self.C,:] - grid[:,1,self.C:,:]
# sym_loss = torch.pow(delta_grid_x * gd_sym_y + delta_grid_y * gd_sym_x, 2).sum()
# pdb.set_trace()
return sym_loss / batch_size
class MaskedBilinearFullSymLoss(nn.Module):
def __init__(self, C):
super(MaskedBilinearFullSymLoss, self).__init__()
self.C = C
def forward(self, grid, gt_sym_axis, gd_sym_axis, mask):
# grid (N,2,H,W)
batch_size = grid.size()[0]
h = grid.size()[2]
w = grid.size()[3]
gt_sym_x = gt_sym_axis[:,0].view(batch_size)
gt_sym_y = gt_sym_axis[:,1].view(batch_size)
gd_sym_x = gd_sym_axis[:,0].view(batch_size)
gd_sym_y = gd_sym_axis[:,1].view(batch_size)
dxs = (-self.C * gt_sym_x)
dys = (self.C * gt_sym_y)
# pdb.set_trace()
sym_loss = 0
for b_i, (dx, dy, sym_x, sym_y) in enumerate(zip(dxs, dys, gd_sym_x, gd_sym_y)):
# dx > 0
dy1_f = dy.floor()
dy2_f = dy1_f + 1
dy1 = dy.floor().int()
dy2 = dy1 + 1
dx1_f = dx.floor()
dx2_f = dx1_f + 1
dx1 = dx.floor().int()
dx2 = dx1 + 1
if dx > 0:
# print ('dx > 0')
# print (dx, dy)
# print (h, w)
# print (grid[b_i,0,:h-dy,:w-dx].shape)
# print (grid[b_i,0,dy:,dx:].shape)
# x2,y1 21
# x1 or y1, 会导致多一列/行 (最后一列/行)
delta_grid_x_11 = grid[b_i,0,:h-dy1-1,:w-dx1-1] - grid[b_i,0,dy1:-1,dx1:-1]
delta_grid_y_11 = grid[b_i,1,:h-dy1-1,:w-dx1-1] - grid[b_i,1,dy1:-1,dx1:-1]
delta_grid_x_21 = grid[b_i,0,:h-dy1-1,:w-dx2] - grid[b_i,0,dy1:-1,dx2:]
delta_grid_y_21 = grid[b_i,1,:h-dy1-1,:w-dx2] - grid[b_i,1,dy1:-1,dx2:]
delta_grid_x_12 = grid[b_i,0,:h-dy2,:w-dx1-1] - grid[b_i,0,dy2:,dx1:-1]
delta_grid_y_12 = grid[b_i,1,:h-dy2,:w-dx1-1] - grid[b_i,1,dy2:,dx1:-1]
delta_grid_x_22 = grid[b_i,0,:h-dy2,:w-dx2] - grid[b_i,0,dy2:,dx2:]
delta_grid_y_22 = grid[b_i,1,:h-dy2,:w-dx2] - grid[b_i,1,dy2:,dx2:]
mask_i = mask[b_i, :h-dy2, :w-dx2]
# pdb.set_trace()
# delta_grid_x = (dx - dx1_f) * (dy - dy1_f) * delta_grid_x_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_x_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_x_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_x_11
# delta_grid_y = (dx - dx1_f) * (dy - dy1_f) * delta_grid_y_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_y_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_y_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_y_11
else: # dx <= 0
# print ('dx < 0')
# print (grid[b_i,0,dy:,:w+dx].shape)
# print (grid[b_i,1,:h-dy,-dx:].shape)
# x2 的第一列不能要
# y1 的第一行不能要
delta_grid_x_11 = grid[b_i,0,dy1+1:,:w+dx1] - grid[b_i,0,1:h-dy1,-dx1:]
delta_grid_y_11 = grid[b_i,1,dy1+1:,:w+dx1] - grid[b_i,1,1:h-dy1,-dx1:]
delta_grid_x_21 = grid[b_i,0,dy1+1:,1:w+dx2] - grid[b_i,0,1:h-dy1,-dx2+1:]
delta_grid_y_21 = grid[b_i,1,dy1+1:,1:w+dx2] - grid[b_i,1,1:h-dy1,-dx2+1:]
delta_grid_x_12 = grid[b_i,0,dy2:,:w+dx1] - grid[b_i,0,:h-dy2,-dx1:]
delta_grid_y_12 = grid[b_i,1,dy2:,:w+dx1] - grid[b_i,1,:h-dy2,-dx1:]
delta_grid_x_22 = grid[b_i,0,dy2:,1:w+dx2] - grid[b_i,0,:h-dy2,-dx2+1:]
delta_grid_y_22 = grid[b_i,1,dy2:,1:w+dx2] - grid[b_i,1,:h-dy2,-dx2+1:]
mask_i = mask[b_i, dy2:, -dx1:]
# pdb.set_trace()
# pdb.set_trace()
# 大小和对称轴的歪转程度有关系
# dx <= 0 : torch.Size([246, 255])
# dx > 0 : torch.Size([246, 255])
# dx > 0: torch.Size([246, 254]) seed 5222
# dx < 0: torch.Size([247, 251]) seed 5221
delta_grid_x = (dx - dx1_f) * (dy - dy1_f) * delta_grid_x_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_x_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_x_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_x_11
delta_grid_y = (dx - dx1_f) * (dy - dy1_f) * delta_grid_y_22 + (dx - dx1_f) * (dy2_f - dy) * delta_grid_y_21 + (dx2_f - dx) * (dy - dy1_f) * delta_grid_y_12 + (dx2_f - dx) * (dy2_f - dy) * delta_grid_y_11
# print ('over', b_i)
# print (delta_grid_x.shape)
# print (delta_grid_y.shape)
# pdb.set_trace()
sym_loss += (mask_i * torch.pow(delta_grid_x * sym_y + delta_grid_y * sym_x, 2)).mean()
# delta_grid_x = grid[:,0,:h-self.C,:] - grid[:,0,self.C:,:]
# delta_grid_y = grid[:,1,:h-self.C,:] - grid[:,1,self.C:,:]
# sym_loss = torch.pow(delta_grid_x * gd_sym_y + delta_grid_y * gd_sym_x, 2).sum()
# pdb.set_trace()
return sym_loss / batch_size
# L1Loss version
# MSELoss version
# BCELoss version
class Face2FaceLoss(nn.Module):
def __init__(self):
super(Face2FaceLoss, self).__init__()
self.stn = STN()
self.mse_loss = nn.MSELoss()
def forward(self, grid, l_fm, r_fm):
batch_size = grid.size(0)
grid_NHWC = grid.permute(0,2,3,1)
warp_fm = self.stn(r_fm, grid_NHWC)
if opt.f2f_kind == "l1":
# pdb.set_trace()
l1_loss = torch.abs(warp_fm - l_fm)
final_loss = l1_loss.mean()
elif opt.f2f_kind == "l2":
l2_loss = self.mse_loss(warp_fm, l_fm)
final_loss = l2_loss
return warp_fm, final_loss
class VggFaceLoss(nn.Module):
def __init__(self, ver = 3):
super(VggFaceLoss, self).__init__()
self.netVgg = netVgg_conv3 if ver == 3 else netVgg_conv4
self.netVgg.load_state_dict(torch.load('./netVggs/netVgg_conv%d.pth' % ver))
self.netVgg.eval()
self.criterion = nn.MSELoss()
self.register_parameter("RGB_mean", nn.Parameter(torch.tensor([129.1863,104.7624,93.5940]).view(1, 3, 1, 1)))
# self.RGB_mean = torch.tensor([129.1863,104.7624,93.5940], device=device).view(1, 3, 1, 1)
# for param in self.netVgg.parameters() 会导致perp loss巨大，必须将self.RGB_mean param的requires_grad也设置为False才可以，或者使用tensor.
for param in self.parameters():
param.requires_grad = False
def forward(self, restored, gt):
restored_vgg = restored * 255 - self.RGB_mean
gt_vgg = gt * 255 - self.RGB_mean
# print ("RGB_mean =", self.RGB_mean)
# RGB->BGR
permute = [2, 1, 0]
gt_feat = self.netVgg(gt_vgg[:, permute, ...])
res_feat = self.netVgg(restored_vgg[:, permute, ...])
loss = self.criterion(res_feat, gt_feat)
return loss