-
Notifications
You must be signed in to change notification settings - Fork 49
/
model_video.py
280 lines (233 loc) · 8.66 KB
/
model_video.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
import torch
from torch import nn
from torch.nn import init
import torch.nn.functional as F
from torch.optim import Adam
import numpy
from einops import rearrange
import time
from transformer import Transformer
from Intra_MLP import index_points,knn_l2
# vgg choice
base = {'vgg': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M']}
# vgg16
def vgg(cfg, i=3, batch_norm=True):
layers = []
in_channels = i
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)]
else:
layers += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return layers
def hsp(in_channel, out_channel):
layers = nn.Sequential(nn.Conv2d(in_channel, out_channel, 1, 1),
nn.ReLU())
return layers
def cls_modulation_branch(in_channel, hiden_channel):
layers = nn.Sequential(nn.Linear(in_channel, hiden_channel),
nn.ReLU())
return layers
def cls_branch(hiden_channel, class_num):
layers = nn.Sequential(nn.Linear(hiden_channel, class_num),
nn.Sigmoid())
return layers
def intra():
layers = []
layers += [nn.Conv2d(512, 512, 1, 1)]
layers += [nn.Sigmoid()]
return layers
def concat_r():
layers = []
layers += [nn.Conv2d(512, 512, 1, 1)]
layers += [nn.ReLU()]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.ReLU()]
layers += [nn.ConvTranspose2d(512, 512, 4, 2, 1)]
return layers
def concat_1():
layers = []
layers += [nn.Conv2d(512, 512, 1, 1)]
layers += [nn.ReLU()]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.ReLU()]
return layers
def mask_branch():
layers = []
layers += [nn.Conv2d(512, 2, 3, 1, 1)]
layers += [nn.ConvTranspose2d(2, 2, 8, 4, 2)]
layers += [nn.Sigmoid()]
return layers
def incr_channel():
layers = []
layers += [nn.Conv2d(128, 512, 3, 1, 1)]
layers += [nn.Conv2d(256, 512, 3, 1, 1)]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
return layers
def incr_channel2():
layers = []
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.Conv2d(512, 512, 3, 1, 1)]
layers += [nn.ReLU()]
return layers
def norm(x, dim):
squared_norm = (x ** 2).sum(dim=dim, keepdim=True)
normed = x / torch.sqrt(squared_norm)
return normed
def fuse_hsp(x, p,group_size=5):
t = torch.zeros(group_size, x.size(1))
for i in range(x.size(0)):
tmp = x[i, :]
if i == 0:
nx = tmp.expand_as(t)
else:
nx = torch.cat(([nx, tmp.expand_as(t)]), dim=0)
nx = nx.view(x.size(0)*group_size, x.size(1), 1, 1)
y = nx.expand_as(p)
return y
class Model(nn.Module):
def __init__(self, device, base, incr_channel, incr_channel2, hsp1, hsp2, cls_m, cls, concat_r, concat_1, mask_branch, intra,demo_mode=False):
super(Model, self).__init__()
self.base = nn.ModuleList(base)
self.sp1 = hsp1
self.sp2 = hsp2
self.cls_m = cls_m
self.cls = cls
self.incr_channel1 = nn.ModuleList(incr_channel)
self.incr_channel2 = nn.ModuleList(incr_channel2)
self.concat4 = nn.ModuleList(concat_r)
self.concat3 = nn.ModuleList(concat_r)
self.concat2 = nn.ModuleList(concat_r)
self.concat1 = nn.ModuleList(concat_1)
self.mask = nn.ModuleList(mask_branch)
self.extract = [13, 23, 33, 43]
self.device = device
self.group_size = 5
self.intra = nn.ModuleList(intra)
self.transformer_1=Transformer(512,4,4,782,group=self.group_size)
self.transformer_2=Transformer(512,4,4,782,group=self.group_size)
self.demo_mode=demo_mode
def forward(self, x):
# backbone, p is the pool2, 3, 4, 5
p = list()
for k in range(len(self.base)):
x = self.base[k](x)
if k in self.extract:
p.append(x)
# increase the channel
newp = list()
newp_T=list()
for k in range(len(p)):
np = self.incr_channel1[k](p[k])
np = self.incr_channel2[k](np)
newp.append(self.incr_channel2[4](np))
if k==3:
tmp_newp_T3=self.transformer_1(newp[k])
newp_T.append(tmp_newp_T3)
if k==2:
newp_T.append(self.transformer_2(newp[k]))
if k<2:
newp_T.append(None)
# intra-MLP
point = newp[3].view(newp[3].size(0), newp[3].size(1), -1)
point = point.permute(0,2,1)
idx = knn_l2(self.device, point, 4, 1)
feat=idx
new_point = index_points(self.device, point,idx)
group_point = new_point.permute(0, 3, 2, 1)
group_point = self.intra[0](group_point)
group_point = torch.max(group_point, 2)[0] # [B, D', S]
intra_mask = group_point.view(group_point.size(0), group_point.size(1), 7, 7)
intra_mask = intra_mask + newp[3]
spa_mask = self.intra[1](intra_mask)
# hsp
x = newp[3]
x = self.sp1(x)
x = x.view(-1, x.size(1), x.size(2) * x.size(3))
x = torch.bmm(x, x.transpose(1, 2))
x = x.view(-1, x.size(1) * x.size(2))
x = x.view(x.size(0) // self.group_size, x.size(1), -1, 1)
x = self.sp2(x)
x = x.view(-1, x.size(1), x.size(2) * x.size(3))
x = torch.bmm(x, x.transpose(1, 2))
x = x.view(-1, x.size(1) * x.size(2))
#cls pred
cls_modulated_vector = self.cls_m(x)
cls_pred = self.cls(cls_modulated_vector)
#semantic and spatial modulator
g1 = fuse_hsp(cls_modulated_vector, newp[0],self.group_size)
g2 = fuse_hsp(cls_modulated_vector, newp[1],self.group_size)
g3 = fuse_hsp(cls_modulated_vector, newp[2],self.group_size)
g4 = fuse_hsp(cls_modulated_vector, newp[3],self.group_size)
spa_1 = F.interpolate(spa_mask, size=[g1.size(2), g1.size(3)], mode='bilinear')
spa_1 = spa_1.expand_as(g1)
spa_2 = F.interpolate(spa_mask, size=[g2.size(2), g2.size(3)], mode='bilinear')
spa_2 = spa_2.expand_as(g2)
spa_3 = F.interpolate(spa_mask, size=[g3.size(2), g3.size(3)], mode='bilinear')
spa_3 = spa_3.expand_as(g3)
spa_4 = F.interpolate(spa_mask, size=[g4.size(2), g4.size(3)], mode='bilinear')
spa_4 = spa_4.expand_as(g4)
y4 = newp_T[3] * g4 + spa_4
for k in range(len(self.concat4)):
y4 = self.concat4[k](y4)
y3 = newp_T[2] * g3 + spa_3
for k in range(len(self.concat3)):
y3 = self.concat3[k](y3)
if k == 1:
y3 = y3 + y4
y2 = newp[1] * g2 + spa_2
for k in range(len(self.concat2)):
y2 = self.concat2[k](y2)
if k == 1:
y2 = y2 + y3
y1 = newp[0] * g1 + spa_1
for k in range(len(self.concat1)):
y1 = self.concat1[k](y1)
if k == 1:
y1 = y1 + y2
y = y1
if self.demo_mode:
tmp=F.interpolate(y1, size=[14,14], mode='bilinear')
tmp=tmp.permute(0,2,3,1).contiguous().reshape(tmp.shape[0]*tmp.shape[2]*tmp.shape[3],tmp.shape[1])
tmp=tmp/torch.norm(tmp,p=2,dim=1).unsqueeze(1)
feat2=(tmp@tmp.t())
feat=F.interpolate(y, size=[14,14], mode='bilinear')
# decoder
for k in range(len(self.mask)):
y = self.mask[k](y)
mask_pred = y[:, 0, :, :]
if self.demo_mode:
return cls_pred, mask_pred,feat,feat2
else:
return cls_pred, mask_pred
# build the whole network
def build_model(device,demo_mode=False):
return Model(device,
vgg(base['vgg']),
incr_channel(),
incr_channel2(),
hsp(512, 64),
hsp(64**2, 32),
cls_modulation_branch(32**2, 512),
cls_branch(512, 78),
concat_r(),
concat_1(),
mask_branch(),
intra(),demo_mode)
# weight init
def xavier(param):
init.xavier_uniform_(param)
def weights_init(m):
if isinstance(m, nn.Conv2d):
xavier(m.weight.data)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)