/
STWDNN.py
180 lines (148 loc) · 6.22 KB
/
STWDNN.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
""" Vision Transformer (ViT) in PyTorch
Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
import torch.nn as nn
from spconv_utils import replace_feature, spconv
from functools import partial
class SparseBasicBlock(spconv.SparseModule):
expansion = 1
def __init__(self, inplanes, planes, planes2, stride=1, norm_fn=None, downsample=None, indice_key=None):
super(SparseBasicBlock, self).__init__()
assert norm_fn is not None
bias = norm_fn is not None
self.conv1 = spconv.SubMConv2d(
inplanes, planes, kernel_size=3, padding= 1, stride=stride, bias=bias, indice_key=indice_key
)
self.bn1 = norm_fn(planes)
self.relu = nn.ReLU()
self.conv2 = spconv.SubMConv2d(
planes, planes2, kernel_size=3, padding= 1, stride=stride, bias=bias, indice_key=indice_key
)
self.bn2 = norm_fn(planes2)
self.stride = stride
self.bn0 = norm_fn(inplanes)
def forward(self, x, bnf):
if bnf :
x = replace_feature(x, self.bn0(x.features))
out = self.conv1(x)
out = replace_feature(out, self.bn1(out.features))
out = replace_feature(out, self.relu(out.features))
out = self.conv2(out)
out = replace_feature(out, self.bn2(out.features))
out = replace_feature(out, self.relu(out.features))
return out
class STWDNN(nn.Module):
def __init__(self, embed_dim=256):
super().__init__()
self.embed_dim = embed_dim
self.pfe = nn.Sequential(
nn.Conv1d(14, 64, kernel_size=1 ),
nn.BatchNorm1d(64),
nn.ReLU(True),
nn.Conv1d(64, embed_dim, kernel_size=1 ),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
nn.Conv1d(embed_dim, embed_dim, kernel_size=1 ),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
nn.Conv1d(embed_dim, embed_dim, kernel_size=1 ),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
)
self.sfe = nn.Sequential(
nn.Conv1d(15, 64, kernel_size=1 ),
nn.BatchNorm1d(64),
nn.ReLU(True),
nn.Conv1d(64, embed_dim, kernel_size=1 ),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
nn.Conv1d(embed_dim, embed_dim, kernel_size=1 ),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
nn.Conv1d(embed_dim, embed_dim, kernel_size=1 ),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
)
self.ofe = nn.Sequential(
nn.Linear(embed_dim, embed_dim),
nn.BatchNorm1d(embed_dim),
nn.ReLU(True),
nn.Linear(embed_dim, 1),
nn.Sigmoid(),
)
self.cfe = nn.Sequential(
nn.Linear(4, embed_dim)
)
self.maxpool = nn.MaxPool2d(kernel_size = 8, stride = 4)
self.avgpool1 = nn.AvgPool2d(kernel_size = 8, stride = 4)
self.avgpool2 = nn.AvgPool2d(kernel_size=3, stride=3)
self.bnraw1 = nn.BatchNorm1d(14)
self.bnraw2 = nn.BatchNorm1d(15)
self.bnraw3 = nn.BatchNorm2d(11)
norm_fn = partial(nn.BatchNorm1d, eps=1e-3, momentum=0.01)
self.rfe1 = SparseBasicBlock(11, 64, 256, norm_fn=norm_fn, indice_key='res1')
self.rfe2 = SparseBasicBlock(256, 256, 256, norm_fn=norm_fn, indice_key='res2')
# size of the inputs of forward [200, 14], [200, 39, 5, 14], [200, 39, 5, 1], [200, 39, 5, 1], [200, 16, 16, 11], [200, 16, 16, 1]
def forward(self, pointfeature, stationfeature, stationpm, stationmask, regionfeature, regionmask):
pointfeature = pointfeature.unsqueeze(-1)
regionfeature = regionfeature.permute(0, 3, 1, 2)
regionmask = regionmask.permute(0, 3, 1, 2)
stationfeature = stationfeature.reshape(200, -1, 14).transpose(1,2)
stationpm = stationpm.reshape(200, -1, 1).transpose(1,2)
stationmask = stationmask.reshape(200, -1, 1).transpose(1,2)
pointfeature = self.bnraw1(pointfeature)
pf1 = self.pfe(pointfeature)
dstf = stationfeature - pointfeature
dstf = torch.cat([dstf, stationpm], 1)
dstf = self.bnraw2(dstf)
dstf = self.sfe(dstf)
dstf = dstf * stationmask
dstf = torch.sum(dstf, 2)
ssm = torch.sum(stationmask, 2)
dstf = dstf/(ssm + 0.0000001)
# with the help of the spconv lib now using the 3*3 conv is better than 1*1 conv
regionfeature = regionfeature - regionfeature[:,:,8:9,8:9]
dicuv = torch.where(regionmask.squeeze(1) == 1)
fearegion = regionfeature[dicuv[0], :, dicuv[1], dicuv[2]]
dicuv = torch.cat([dicuv[0].unsqueeze(1),dicuv[1].unsqueeze(1),dicuv[2].unsqueeze(1)], 1)
regionfeaturesp = spconv.SparseConvTensor(
features=fearegion,
indices=dicuv.int(),
spatial_shape=[16,16],
batch_size=200
)
rf1 = self.rfe1(regionfeaturesp, True)
rf1 = rf1.dense()
rf1 = rf1 * regionmask
avgrf1 = self.avgpool1(rf1)
avgrmask1 = self.avgpool1(regionmask)
maxmask = self.maxpool(regionmask)
avgrf1 = avgrf1/(avgrmask1 + 0.0000001)
avgrf1 = avgrf1 * maxmask
dicuv2 = torch.where(maxmask.squeeze(1) == 1)
fearegion2 = avgrf1[dicuv2[0], :, dicuv2[1], dicuv2[2]]
dicuv2 = torch.cat([dicuv2[0].unsqueeze(1), dicuv2[1].unsqueeze(1), dicuv2[2].unsqueeze(1)], 1)
regionfeaturesp2 = spconv.SparseConvTensor(
features=fearegion2,
indices=dicuv2.int(),
spatial_shape=[3,3],
batch_size=200
)
rf2 = self.rfe2(regionfeaturesp2, False)
rf2 = rf2.dense()
rf2 = rf2 * maxmask
avgrf2 = self.avgpool2(rf2)
avgrmask2 = self.avgpool2(maxmask)
avgrf2 = avgrf2 / (avgrmask2 + 0.0000001)
avgrf2 = avgrf2.squeeze()
fetavg = torch.mean(pf1, 1)
fetmax, _ = torch.max(pf1, 1)
fetmin, _ = torch.min(pf1, 1)
fetstd = torch.std(pf1, 1)
at2 = torch.cat([fetavg, fetmax, fetmin, fetstd], 1)
wei1 = self.cfe(at2)
pf1 = pf1.squeeze()
allf = pf1 + wei1 * dstf + wei1 * avgrf2
allf = self.ofe(allf)
return allf