/
FSPNet_model.py
249 lines (203 loc) · 11.2 KB
/
FSPNet_model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import vit
class GCN(nn.Module):
def __init__(self, num_state, num_node, bias=False): # num_state=384 num_node=16
super(GCN, self).__init__()
self.conv1 = nn.Conv1d(num_node, num_node, kernel_size=1)
self.relu = nn.ReLU(inplace=True)
self.conv2 = nn.Conv1d(num_state, num_state, kernel_size=1, bias=bias)
def forward(self, x): # x [16,384,16]
h = self.conv1(x.permute(0, 2, 1)).permute(0, 2, 1)
h = h - x
h = self.relu(self.conv2(h))
return h
class Converter(nn.Module):
def __init__(self, dim_in, dim_temp=384, img_size=384, mids=4):
super(Converter, self).__init__()
self.img_size = img_size
self.dim_in = dim_in
self.dim_temp = dim_temp
self.num_n = mids * mids
self.conv_fc = nn.Conv2d(self.dim_in * 2, self.dim_temp, kernel_size=1)
# f1
self.norm_layer_f1 = nn.LayerNorm(dim_in)
self.conv_f1_Q = nn.Conv2d(self.dim_in, self.dim_temp, kernel_size=1)
self.conv_f1_K = nn.Conv2d(self.dim_in, self.dim_temp, kernel_size=1)
self.ap_f1 = nn.AdaptiveAvgPool2d(output_size=(mids + 2, mids + 2))
self.gcn_f1 = GCN(num_state=self.dim_temp, num_node=self.num_n)
self.conv_f1_extend = nn.Conv2d(self.dim_temp, self.dim_in, kernel_size=1, bias=False)
# f2
self.norm_layer_f2 = nn.LayerNorm(dim_in)
self.conv_f2_Q = nn.Conv2d(self.dim_in, self.dim_temp, kernel_size=1)
self.conv_f2_K = nn.Conv2d(self.dim_in, self.dim_temp, kernel_size=1)
self.ap_f2 = nn.AdaptiveAvgPool2d(output_size=(mids + 2, mids + 2))
self.gcn_f2 = GCN(num_state=self.dim_temp, num_node=self.num_n)
self.conv_f2_extend = nn.Conv2d(self.dim_temp, self.dim_in, kernel_size=1, bias=False)
def forward(self, token_pair):
# tokens list 12x[8,578,768]
bs, num_token, chs = token_pair[0].shape
tokens_ls = []
for index in range(len(token_pair) // 2):
f1_ = self.norm_layer_f1(token_pair[index * 2][:, 2:, :]) # [8,576,768]
f2_ = self.norm_layer_f2(token_pair[index * 2 + 1][:, 2:, :]) # [8,576,768]
f1_ = f1_.permute(0, 2, 1).view(bs, chs, int(self.img_size // 16), int(self.img_size // 16)).contiguous()
# [8,768,24,24]
f2_ = f2_.permute(0, 2, 1).view(bs, chs, int(self.img_size // 16), int(self.img_size // 16)).contiguous()
# [8,768,24,24]
f1, f2 = f1_, f2_ # [8,768,24,24] / [8,768,24,24]
fc = self.conv_fc(torch.cat((f1, f2), dim=1)) # [8,384,24,24]
fc_att = torch.nn.functional.softmax(fc, dim=1)[:, 1, :, :].unsqueeze(1) # [8,1,24,24]
# f1 pass
f1_Q = self.conv_f1_Q(f1).view(bs, self.dim_temp, -1).contiguous() # [8,384,576] [bs,chs,24*24]
f1_K = self.conv_f1_K(f1) # [8,384,24,24]
f1_masked = f1_K * fc_att # [8,384,24,24]
f1_V = self.ap_f1(f1_masked)[:, :, 1:-1, 1:-1].reshape(bs, self.dim_temp, -1) # [8,384,16]
f1_proj_reshaped = torch.matmul(f1_V.permute(0, 2, 1), f1_K.reshape(bs, self.dim_temp, -1)) # [8,16,576]
f1_proj_reshaped = torch.nn.functional.softmax(f1_proj_reshaped, dim=1) # [8,16,576] Tv
f1_rproj_reshaped = f1_proj_reshaped # [8,16,576]
f1_n_state = torch.matmul(f1_Q, f1_proj_reshaped.permute(0, 2, 1)) # [16,384,16] Ta
f1_n_rel = self.gcn_f1(f1_n_state) # [16,384,16]
f1_state_reshaped = torch.matmul(f1_n_rel, f1_rproj_reshaped) # [16,384,576]
f1_state = f1_state_reshaped.view(bs, self.dim_temp, *f1.size()[2:]) # [16,384,24,24]
f1_out = f1_ + (self.conv_f1_extend(f1_state)) # [16,768,24,24]
# f2 pass
f2_Q = self.conv_f2_Q(f2).view(bs, self.dim_temp, -1).contiguous() # [8,384,576] [bs,chs,24*24]
f2_K = self.conv_f2_K(f2) # [8,384,24,24]
f2_masked = f2_K * fc_att # [8,384,24,24]
f2_V = self.ap_f2(f2_masked)[:, :, 1:-1, 1:-1].reshape(bs, self.dim_temp, -1) # [8,384,16]
f2_proj_reshaped = torch.matmul(f2_V.permute(0, 2, 1), f2_K.reshape(bs, self.dim_temp, -1)) # [8,16,576]
f2_proj_reshaped = torch.nn.functional.softmax(f2_proj_reshaped, dim=1) # [8,16,576]
f2_rproj_reshaped = f2_proj_reshaped # [8,16,576]
f2_n_state = torch.matmul(f2_Q, f2_proj_reshaped.permute(0, 2, 1)) # [16,384,16]
f2_n_rel = self.gcn_f2(f2_n_state) # [16,384,16]
f2_state_reshaped = torch.matmul(f2_n_rel, f2_rproj_reshaped) # [16,384,576]
f2_state = f2_state_reshaped.view(bs, self.dim_temp, *f2.size()[2:]) # [16,384,24,24]
f2_out = f2_ + (self.conv_f2_extend(f2_state)) # [16,768,24,24]
tokens_ls.extend([f1_out, f2_out])
return tokens_ls
class UpSampling2x(nn.Module):
def __init__(self, in_chs, out_chs):
super(UpSampling2x, self).__init__()
temp_chs = out_chs * 4
self.up_module = nn.Sequential(
nn.Conv2d(in_chs, temp_chs, 1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True),
nn.PixelShuffle(2)
)
def forward(self, features):
return self.up_module(features)
class GroupFusion(nn.Module):
def __init__(self, in_chs, out_chs, start=False): # 768, 384
super(GroupFusion, self).__init__()
temp_chs = in_chs
if start:
in_chs = in_chs
else:
in_chs *= 2
self.gf1 = nn.Sequential(nn.Conv2d(in_chs, temp_chs, 1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True),
nn.Conv2d(temp_chs, temp_chs, 3, padding=1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True),
nn.Conv2d(temp_chs, temp_chs, 3, padding=1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True))
self.gf2 = nn.Sequential(nn.Conv2d((temp_chs + temp_chs), temp_chs, 1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True),
nn.Conv2d(temp_chs, temp_chs, 3, padding=1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True),
nn.Conv2d(temp_chs, temp_chs, 3, padding=1, bias=False),
nn.BatchNorm2d(temp_chs),
nn.ReLU(inplace=True))
self.up2x = UpSampling2x(temp_chs, out_chs)
def forward(self, f_r, f_l):
f_r = self.gf1(f_r) # chs 768
f12 = self.gf2(torch.cat((f_r, f_l), dim=1)) # chs 768
return f12, self.up2x(f12)
class OutPut(nn.Module):
def __init__(self, in_chs, scale=1):
super(OutPut, self).__init__()
self.out = nn.Sequential(nn.Conv2d(in_chs, in_chs, 1, bias=False),
nn.BatchNorm2d(in_chs),
nn.ReLU(inplace=True),
nn.UpsamplingBilinear2d(scale_factor=scale),
nn.Conv2d(in_chs, 1, 1),
nn.Sigmoid())
def forward(self, feat):
return self.out(feat)
class Model(nn.Module):
def __init__(self, ckpt, img_size=384):
super(Model, self).__init__()
self.encoder = vit.deit_base_distilled_patch16_384()
if ckpt is not None:
ckpt = torch.load(ckpt, map_location='cpu')
msg = self.encoder.load_state_dict(ckpt["model"], strict=False)
print("====================================")
print(msg)
self.img_size = img_size
self.vit_chs = 768
self.group_converter_0 = Converter(dim_in=self.vit_chs, img_size=self.img_size)
self.group_converter_1 = Converter(dim_in=self.vit_chs, img_size=self.img_size)
self.group_converter_2 = Converter(dim_in=self.vit_chs, img_size=self.img_size)
self.group_converter_3 = Converter(dim_in=self.vit_chs, img_size=self.img_size)
self.group_converter_4 = Converter(dim_in=self.vit_chs, img_size=self.img_size)
self.group_converter_5 = Converter(dim_in=self.vit_chs, img_size=self.img_size)
self.gf1_1 = GroupFusion(768, 384)
self.gf1_2 = GroupFusion(768, 384)
self.gf1_3 = GroupFusion(768, 384)
self.gf1_4 = GroupFusion(768, 384)
self.gf1_5 = GroupFusion(768, 384)
self.gf1_6 = GroupFusion(768, 384, start=True)
self.gf2_1 = GroupFusion(384, 192)
self.gf2_2 = GroupFusion(384, 192)
self.gf2_3 = GroupFusion(384, 192, start=True)
self.gf3_1 = GroupFusion(192, 192)
self.gf3_2 = GroupFusion(192, 192, start=True)
self.gf4_1 = GroupFusion(192, 192, start=True)
self.out1 = OutPut(in_chs=768, scale=16)
self.out2 = OutPut(in_chs=384, scale=8)
self.out3 = OutPut(in_chs=192, scale=4)
self.out4 = OutPut(in_chs=192)
def group_converter_fn(self, tokens):
group_converter_ls = [self.group_converter_0, self.group_converter_1, self.group_converter_2,
self.group_converter_3, self.group_converter_4, self.group_converter_5]
tokens_ls = []
for index in range(len(tokens) // 2):
token_pair = [tokens[index * 2], tokens[index * 2 + 1]]
token_pair_out = group_converter_ls[index](token_pair)
tokens_ls.extend(token_pair_out)
return tokens_ls
def group_pyramid_decode(self, feature):
# list 12x[8,768,24,24]
# layer1 out
f1_6_l, f2_6 = self.gf1_6(feature[-1], feature[-2])
f1_5_l, f2_5 = self.gf1_5(torch.cat((feature[-3], f1_6_l), dim=1), feature[-4])
f1_4_l, f2_4 = self.gf1_4(torch.cat((feature[-5], f1_5_l), dim=1), feature[-6])
f1_3_l, f2_3 = self.gf1_3(torch.cat((feature[-7], f1_4_l), dim=1), feature[-8])
f1_2_l, f2_2 = self.gf1_2(torch.cat((feature[-9], f1_3_l), dim=1), feature[-10])
f1_1_o, f2_1 = self.gf1_1(torch.cat((feature[-11], f1_2_l), dim=1), feature[-12]) # f1_1_l [bs,768,24,24]
# layer2 out
f2_3_l, f3_3 = self.gf2_3(f2_6, f2_5)
f2_2_l, f3_2 = self.gf2_2(torch.cat((f2_4, f2_3_l), dim=1), f2_3)
f2_1_o, f3_1 = self.gf2_1(torch.cat((f2_2, f2_2_l), dim=1), f2_1) # f2_1_l [bs,384,48,48]
# layer3 out
f3_2_l, f4_2 = self.gf3_2(f3_3, f3_2)
f3_1_o, f4_1 = self.gf3_1(torch.cat((f3_2, f3_2_l), dim=1), f3_1) # f3_1_l [bs,192,96,96]
# layer4 out
_, f5_1 = self.gf4_1(f4_2, f4_1)
return f1_1_o, f2_1_o, f3_1_o, f5_1
def pred_out(self, gpd_outs):
return self.out1(gpd_outs[0]), self.out2(gpd_outs[1]), self.out3(gpd_outs[2]), self.out4(gpd_outs[3])
def forward(self, img):
# B Seq
B, C, H, W = img.size()
x = self.encoder(img) # list 12x[8,576,768]
feature = self.group_converter_fn(x)
gpd_outs = self.group_pyramid_decode(feature)
return self.pred_out(gpd_outs)