-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
184 lines (150 loc) · 6.87 KB
/
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import torchvision.models as models
import torch_geometric as pyg
import torch_geometric.nn as pygnn
import torch_geometric.utils as pyg_utils
class RAN(nn.Module):
def __init__(self, parts, map_threshold, att_dim,
drop_out=0., map_size=14, cov_channel=2048, is_hidden=False):
super(RAN, self).__init__()
self.map_threshold = map_threshold
self.parts = parts
self.map_size = map_size
self.cov_channel = cov_channel
self.pool = nn.MaxPool2d(self.map_size, self.map_size)
self.cov = nn.Conv2d(self.cov_channel, self.parts, 1)
self.p_linear = nn.Linear(self.cov_channel*self.parts, att_dim, False)
self.dropout2 = nn.Dropout(drop_out)
def forward(self, features):
# assert features
w = features.size()
weights = torch.sigmoid(self.cov(features))
# threshold the weights
batch, parts, width, height = weights.size()
weights_layout = weights.view(batch, -1)
threshold_value, _ = weights_layout.max(dim=1)
local_max, _ = weights.view(batch, parts, -1).max(dim=2)
threshold_value = self.map_threshold*threshold_value.view(batch, 1) \
.expand(batch, parts)
weights = weights*local_max.ge(threshold_value).view(batch, parts, 1, 1). \
float().expand(batch, parts, width, height)
blocks = []
for k in range(self.parts):
Y = features*weights[:, k, :, :].unsqueeze(dim=1).expand(
w[0], self.cov_channel, w[2], w[3])
blocks.append(self.pool(Y).squeeze().view(-1, self.cov_channel))
p_output = self.dropout2(self.p_linear(torch.cat(blocks, dim=1)))
parts = torch.cat([b.unsqueeze(dim=1) for b in blocks], dim=1)
return p_output, parts.permute(0, 2, 1)
class GAT(nn.Module):
def __init__(self, num_node, input_dim, embed_dim, res=False,
dropout=0., concat_embed=True, layer_num=3, heads=1):
super(GAT, self).__init__()
self.num_node = num_node
self.input_dim = input_dim
self.embed_dim = embed_dim
self.dropout = dropout
self.res = res
assert embed_dim % heads == 0
self.conv_list = nn.ModuleList()
self.conv_list.append(pygnn.GATConv(input_dim, embed_dim // heads,
heads=heads, add_self_loops=False))
for _ in range(layer_num - 1):
conv = pygnn.GATConv(embed_dim, embed_dim // heads,
heads=heads, add_self_loops=False)
self.conv_list.append(conv)
self.concat_embed = concat_embed
conv_concat_dim = self.input_dim + len(self.conv_list) * embed_dim
self.mlp = nn.Sequential(
nn.Linear(conv_concat_dim, input_dim),
nn.ReLU())
def forward(self, data):
x, edge_index = data.x, data.edge_index
x_concat = [x]
for conv in self.conv_list:
x = conv(x, edge_index=edge_index)
x = x + F.relu(x) if self.res else F.relu(x)
x_concat.append(x)
return self.mlp(torch.cat(x_concat, dim=1))
class GNDAN(nn.Module):
def __init__(self, config):
super(GNDAN, self).__init__()
self.config = config
self.dim_f = config.dim_f
self.dim_v = config.dim_v
self.nclass = config.num_class
self.att = nn.Parameter(torch.empty(
self.nclass, config.num_attribute), requires_grad=False)
self.bias = nn.Parameter(torch.tensor(1), requires_grad=False)
self.mask_bias = nn.Parameter(torch.empty(
1, self.nclass), requires_grad=False)
self.V = nn.Parameter(torch.empty(
config.num_attribute, self.dim_v), requires_grad=True)
self.W_1 = nn.Parameter(torch.empty(
self.dim_v, self.dim_f), requires_grad=True)
self.W_2 = nn.Parameter(torch.empty(
self.dim_v, self.dim_f), requires_grad=True)
# graph model
num_node = config.num_attribute
self.graph_model = GAT(num_node=num_node,
input_dim=self.dim_f,
embed_dim=config.GAT_embed_dim,
res=config.GAT_res,
heads=config.GAT_heads,
concat_embed=config.GAT_concat_embed,
layer_num=config.GAT_layer_num)
self.aren = RAN(parts=config.AREN_parts,
map_threshold=config.AREN_map_threshold,
att_dim=config.num_attribute,
map_size=int(np.sqrt(config.resnet_region)),
drop_out=config.AREN_dropout,
is_hidden=False)
# bakcbone
resnet101 = models.resnet101(pretrained=True)
self.resnet101 = nn.Sequential(*list(resnet101.children())[:-2])
def forward(self, imgs):
Fs = self.resnet101(imgs)
ran_embed, _ = self.aren(Fs)
rgat_embed = self.forward_feature_graph(Fs)
package = {}
package['rgat_embed'] = self.forward_attribute(rgat_embed)
package['ran_embed'] = self.forward_attribute(ran_embed)
coef_ran = self.config.coef_ran
coef_rgat = self.config.coef_rgat
package['embed'] = (coef_rgat * package['rgat_embed'] +
coef_ran * package['ran_embed']) / (coef_rgat + coef_ran)
return package
def forward_attribute(self, embed):
embed = torch.einsum('ki,bi->bk', self.att, embed)
self.vec_bias = self.mask_bias*self.bias
embed = embed + self.vec_bias
return embed
def forward_feature_graph(self, Fs):
if len(Fs.shape) == 4:
shape = Fs.shape
Fs = Fs.reshape(shape[0], shape[1], shape[2]
* shape[3]) # batch x 2048 x 49
V_n = F.normalize(self.V) if self.config.normalize_V else self.V
Fs = F.normalize(Fs, dim=1)
A = torch.einsum('iv,vf,bfr->bir', V_n, self.W_2, Fs)
A = F.softmax(A, dim=-1)
F_p = torch.einsum('bir,bfr->bif', A, Fs)
F_p = F_p.permute(0, 2, 1)
F_pn = F.normalize(F_p, dim=1)
dense_adj_batch = F_pn.permute(0, 2, 1) @ F_pn
feature_batch = F_p.permute(0, 2, 1)
graph_list = []
for node_feature, dense_adj in zip(feature_batch, dense_adj_batch):
edge_index, edge_attr = pyg_utils.dense_to_sparse(dense_adj)
graph_list.append(pyg.data.Data(
node_feature, edge_index, edge_attr))
graph = pyg.data.Batch.from_data_list(graph_list)
outputs = self.graph_model(graph)
F_p = pyg_utils.to_dense_batch(x=outputs, batch=graph.batch)[0]
embed = torch.einsum('iv,vf,bif->bi', V_n, self.W_1, F_p)
return embed
if __name__ == '__main__':
pass