-
Notifications
You must be signed in to change notification settings - Fork 38
/
Copy pathstylize_objectives.py
237 lines (152 loc) · 7.77 KB
/
stylize_objectives.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
import math
import torch
from torch.autograd import Variable
import torch.nn.functional as F
from .contextual_loss import *
from . import utils
use_random=True
class objective_class():
def __init__(self, objective='remd_dp'):
self.z_dist = torch.zeros(1).cuda()
self.rand_ixx = {}
self.rand_ixy = {}
self.rand_iy = {}
if objective == 'remd_dp':
self.eval = self.gen_remd_dp_objective
if objective == 'remd_dp_g':
self.eval = self.gen_remd_dp_objective_guided
def gen_remd_dp_objective_guided(self, z_x, z_c, z_s, gz, d_ell, d_weight=10.0, content_weight=4.0, moment_weight=1.0, style_loss_func=remd_loss, content_loss_func=dp_loss, h=1.0):
## Extract Random Subset of Features from Stylized Image & Content Image ##
# (Choose Features from Same locations in Stylized Image & Content Image) #
final_loss = 0.
for ri in range(len(self.rand_ixx.keys())):
xx, xy, yx = self.get_feature_inds(ri=ri)
x_st, c_st = self.spatial_feature_extract(z_x, z_c, xx, xy)
if gz.sum() > 0.:
gxx, gxy = self.get_feature_inds_g()
gx_st, gc_st = self.spatial_feature_extract(z_x, z_c, gxx, gxy)
## Reshape Features from Style Distribution ##
d = z_s[ri][0].size(1)
z_st = z_s[ri][0].view(1,d,-1,1)
## Compute Content Loss ##
fm = 3+2*64+128*2+256*3+512*2
ell_content = content_loss_func(x_st[:,:,:,:], c_st[:,:,:,:])
## Compute Style Loss ##
remd_loss, used_style_feats = style_loss_func(x_st[:,:fm,:,:], z_st[:,:fm,:,:], self.z_dist, splits=[fm])
if gz.sum() > 0.:
for j in range(gz.size(2)):
remd_loss += style_loss_func(gx_st[:,:-2,j:(j+1),:], gz[:,:,j:(j+1),:],self.z_dist[:1]*0.)[0]/gz.size(2)
### Compute Moment Loss (constrains magnitude of features ###
moment_ell = 0.
if gz.sum() > 0.:
moment_ell = moment_loss(torch.cat([x_st,gx_st],2)[:,:-2,:,:],torch.cat([z_st,gz],2),moments=[1,2])
else:
moment_ell = moment_loss(x_st[:,:-2,:,:],z_st,moments=[1,2])
### Add Pallette Matching Loss ###
content_weight_frac = 1./max(content_weight,1.)
moment_ell += content_weight_frac*style_loss_func(x_st[:,:3,:,:], z_st[:,:3,:,:],self.z_dist,splits=[3])[0]
### Combine Terms and Normalize ###
ell_style = remd_loss+moment_weight*moment_ell
style_weight = 1.0 + moment_weight
final_loss += (content_weight*ell_content+ell_style)/(content_weight+style_weight)
return final_loss/len(self.rand_ixx.keys())
def init_inds(self, z_x, z_s_all, r, ri):
const = 128**2
z_s = z_s_all[ri]
try:
temp = self.rand_ixx[ri]
except:
self.rand_ixx[ri]= []
self.rand_ixy[ri]= []
self.rand_iy[ri]= []
for i in range(len(z_s)):
d = z_s[i].size(1)
z_st = z_s[i].view(1,d,-1,1)
x_st = z_x[i]
big_size = x_st.size(3)*x_st.size(2)
global use_random
if use_random:
stride_x = int(max(math.floor(math.sqrt(big_size//const)),1))
offset_x = np.random.randint(stride_x)
stride_y = int(max(math.ceil(math.sqrt(big_size//const)),1))
offset_y = np.random.randint(stride_y)
else:
stride_x = int(max(math.floor(math.sqrt(big_size//const)),1))
offset_x = stride_x//2
stride_y = int(max(math.ceil(math.sqrt(big_size//const)),1))
offset_y = stride_y//2
region_mask = r#.flatten()
xx,xy = np.meshgrid(np.array(range(x_st.size(2)))[offset_x::stride_x], np.array(range(x_st.size(3)))[offset_y::stride_y] )
xx = np.expand_dims(xx.flatten(),1)
xy = np.expand_dims(xy.flatten(),1)
xc = np.concatenate([xx,xy],1)
try:
xc = xc[region_mask[xy[:,0],xx[:,0]],:]
except:
region_mask = region_mask[:,:]
xc = xc[region_mask[xy[:,0],xx[:,0]],:]
self.rand_ixx[ri].append(xc[:,0])
self.rand_ixy[ri].append(xc[:,1])
zx = np.array(range(z_st.size(2))).astype(np.int32)
self.rand_iy[ri].append(zx)
def init_g_inds(self, coords, x_im):
self.g_ixx = (coords[:,0]*x_im.size(2)).astype(np.int64)
self.g_ixy = (coords[:,1]*x_im.size(3)).astype(np.int64)
def spatial_feature_extract(self, z_x, z_c, xx, xy):
l2 = []
l3 = []
for i in range(len(z_x)):
temp = z_x[i]
temp2 = z_c[i]
if i>0 and z_x[i-1].size(2) > z_x[i].size(2):
xx = xx/2.0
xy = xy/2.0
xxm = np.floor(xx).astype(np.float32)
xxr = xx - xxm
xym = np.floor(xy).astype(np.float32)
xyr = xy - xym
w00 = utils.to_device(torch.from_numpy((1.-xxr)*(1.-xyr))).float().unsqueeze(0).unsqueeze(1).unsqueeze(3)
w01 = utils.to_device(torch.from_numpy((1.-xxr)*xyr)).float().unsqueeze(0).unsqueeze(1).unsqueeze(3)
w10 = utils.to_device(torch.from_numpy(xxr*(1.-xyr))).float().unsqueeze(0).unsqueeze(1).unsqueeze(3)
w11 = utils.to_device(torch.from_numpy(xxr*xyr)).float().unsqueeze(0).unsqueeze(1).unsqueeze(3)
xxm = np.clip(xxm.astype(np.int32),0,temp.size(2)-1)
xym = np.clip(xym.astype(np.int32),0,temp.size(3)-1)
s00 = xxm*temp.size(3)+xym
s01 = xxm*temp.size(3)+np.clip(xym+1,0,temp.size(3)-1)
s10 = np.clip(xxm+1,0,temp.size(2)-1)*temp.size(3)+(xym)
s11 = np.clip(xxm+1,0,temp.size(2)-1)*temp.size(3)+np.clip(xym+1,0,temp.size(3)-1)
temp = temp.view(1,temp.size(1),temp.size(2)*temp.size(3),1)
temp = temp[:,:,s00,:].mul_(w00).add_(temp[:,:,s01,:].mul_(w01)).add_(temp[:,:,s10,:].mul_(w10)).add_(temp[:,:,s11,:].mul_(w11))
temp2 = temp2.view(1,temp2.size(1),temp2.size(2)*temp2.size(3),1)
temp2 = temp2[:,:,s00,:].mul_(w00).add_(temp2[:,:,s01,:].mul_(w01)).add_(temp2[:,:,s10,:].mul_(w10)).add_(temp2[:,:,s11,:].mul_(w11))
l2.append(temp)
l3.append(temp2)
x_st = torch.cat([li.contiguous() for li in l2],1)
c_st = torch.cat([li.contiguous() for li in l3],1)
xx = torch.from_numpy(xx).cuda().view(1,1,x_st.size(2),1).float()
yy = torch.from_numpy(xy).cuda().view(1,1,x_st.size(2),1).float()
x_st = torch.cat([x_st,xx,yy],1)
c_st = torch.cat([c_st,xx,yy],1)
return x_st, c_st
def shuffle_feature_inds(self, i=0):
global use_random
if use_random:
for ri in self.rand_ixx.keys():
np.random.shuffle(self.rand_ixx[ri][i])
np.random.shuffle(self.rand_ixy[ri][i])
np.random.shuffle(self.rand_iy[ri][i])
def get_feature_inds(self,ri=0, i=0, cnt=32**2):
global use_random
if use_random:
xx = self.rand_ixx[ri][i][:cnt]
xy = self.rand_ixy[ri][i][:cnt]
yx = self.rand_iy[ri][i][:cnt]
else:
xx = self.rand_ixx[ri][i][::(self.rand_ixx[ri][i].shape[0]//cnt)]
xy = self.rand_ixy[ri][i][::(self.rand_ixy[ri][i].shape[0]//cnt)]
yx = self.rand_iy[ri][i][::(self.rand_iy[ri][i].shape[0]//cnt)]
return xx, xy, yx
def get_feature_inds_g(self, i=0, cnt=32**2):
xx = self.g_ixx
xy = self.g_ixy
return xx, xy