-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_regressor.lua
161 lines (146 loc) · 7.76 KB
/
model_regressor.lua
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
require 'nn'
require 'cutorch'
require 'cunn'
require 'cudnn'
torch.setdefaulttensortype('torch.FloatTensor')
local function tool_regress(input_channels, first_layer_channels)
local m = nn.Sequential()
m:add(nn.SpatialConvolution(input_channels, first_layer_channels, 9, 9, 1, 1, 4, 4))
-- :add(nn.SpatialBatchNormalization(first_layer_channels))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels, first_layer_channels*2, 13, 13, 1, 1, 6, 6))
-- :add(nn.SpatialBatchNormalization(first_layer_channels*2))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*2, first_layer_channels*4, 15,15,1,1,7,7))
-- :add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*4, first_layer_channels*4, 1,1,1,1))
-- :add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*4, first_layer_channels*4, 1, 1, 1,1))
-- :add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*4, input_channels, 1, 1, 1, 1))
-- :add(nn.SpatialBatchNormalization(input_channels))
:add(nn.ReLU())
-- m:add(nn.Sigmoid())
return m
end
local function tool_regressFull(joint_num, compo_num, first_layer_channels)
local function Dual(in_c, dual_c1, dual_c2, kernel_size, stride, pad)
local dual = nn.ConcatTable()
dual:add(nn.Sequential()
:add(nn.SpatialConvolution(in_c, dual_c1, kernel_size, kernel_size, stride, stride, pad, pad))
:add(nn.SpatialBatchNormalization(dual_c1))
:add(nn.ReLU())
)
dual:add(nn.Sequential()
:add(nn.SpatialConvolution(in_c, dual_c2, kernel_size, kernel_size, stride, stride, pad, pad))
:add(nn.SpatialBatchNormalization(dual_c2))
:add(nn.ReLU())
)
dual = nn.Sequential():add(dual):add(nn.JoinTable(2))
return dual
end
local function DualNoR(in_c, dual_c1, dual_c2, kernel_size, stride, pad)
local dual = nn.ConcatTable()
dual:add(nn.Sequential()
:add(nn.SpatialConvolution(in_c, dual_c1, kernel_size, kernel_size, stride, stride, pad, pad))
:add(nn.SpatialBatchNormalization(dual_c1))
)
dual:add(nn.Sequential()
:add(nn.SpatialConvolution(in_c, dual_c2, kernel_size, kernel_size, stride, stride, pad, pad))
:add(nn.SpatialBatchNormalization(dual_c2))
)
dual = nn.Sequential():add(dual):add(nn.JoinTable(2))
return dual
end
local input_channels = joint_num+compo_num+3
local m = nn.Sequential()
-- local dual1 = Dual(input_channels, first_layer_channels, first_layer_channels, 3, 1, 1)
-- local dual2 = Dual(first_layer_channels*2, first_layer_channels*2, first_layer_channels*2, 5, 1, 2)
-- local dual3 = Dual(first_layer_channels*4, first_layer_channels*4, first_layer_channels*4, 5, 1, 2)
local dual1 = Dual(input_channels, first_layer_channels, first_layer_channels, 3, 1, 1)
local cbr1 = nn.Sequential()
:add(nn.SpatialConvolution(first_layer_channels*2, first_layer_channels*2, 1, 1, 1, 1, 0, 0))
:add(nn.SpatialBatchNormalization(first_layer_channels*2))
:add(nn.ReLU())
local dual2 = Dual(first_layer_channels*2, first_layer_channels*2, first_layer_channels*2, 3, 1, 1)
local cbr2 = nn.Sequential()
:add(nn.SpatialConvolution(first_layer_channels*4, first_layer_channels*4, 1, 1, 1, 1, 0, 0))
:add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
local dual3 = Dual(first_layer_channels*4, first_layer_channels*4, first_layer_channels*4, 3, 1, 1)
local cbr3 = nn.Sequential()
:add(nn.SpatialConvolution(first_layer_channels*8, first_layer_channels*8, 1, 1, 1, 1, 0, 0))
:add(nn.SpatialBatchNormalization(first_layer_channels*8))
:add(nn.ReLU())
local dual4 = Dual(first_layer_channels*8, first_layer_channels*4, first_layer_channels*4, 3, 1, 1)
local cbr4 = nn.Sequential()
:add(nn.SpatialConvolution(first_layer_channels*8, first_layer_channels*8, 1, 1, 1, 1, 0, 0))
:add(nn.SpatialBatchNormalization(first_layer_channels*8))
:add(nn.ReLU())
local dual5 = Dual(first_layer_channels*8, first_layer_channels*4, first_layer_channels*4, 1, 1, 0)
local cbr5 = nn.Sequential()
:add(nn.SpatialConvolution(first_layer_channels*8, first_layer_channels*8, 1, 1, 1, 1, 0, 0))
:add(nn.SpatialBatchNormalization(first_layer_channels*8))
:add(nn.ReLU())
local dual6 = DualNoR(first_layer_channels*8, joint_num, compo_num, 1, 1, 0)
m:add(dual1)--:add(cbr1)
:add(dual2)--:add(cbr2)
:add(dual3)--:add(cbr3)
:add(dual4)--:add(cbr4)
:add(dual5)--:add(cbr5)
:add(dual6)
return m
end
local function tool_regressFullNoConcat(joint_num, compo_num, first_layer_channels)
local input_channels = 3 + joint_num + compo_num
local output_channels = joint_num + compo_num
local m = nn.Sequential()
m:add(nn.SpatialConvolution(input_channels, first_layer_channels, 3, 3, 1, 1, 1, 1))
:add(nn.SpatialBatchNormalization(first_layer_channels))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels, first_layer_channels*2, 3, 3, 1, 1, 1, 1))
:add(nn.SpatialBatchNormalization(first_layer_channels*2))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*2, first_layer_channels*4, 3, 3, 1, 1, 1, 1))
:add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*4, first_layer_channels*4, 3, 3, 1, 1, 1, 1))
:add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*4, first_layer_channels*4, 1, 1, 1, 1))
:add(nn.SpatialBatchNormalization(first_layer_channels*4))
:add(nn.ReLU())
m:add(nn.SpatialConvolution(first_layer_channels*4, output_channels, 1, 1, 1, 1))
:add(nn.SpatialBatchNormalization(output_channels))
-- :add(nn.ReLU())
-- m:add(nn.Sigmoid())
return m
end
local joint_num = 5
local compo_num = 4
--local regress_net = tool_regress(joint_num+compo_num, 64)
local regress_net = tool_regressFullNoConcat(joint_num, compo_num, 64)
--local regress_net = tool_regressFull(joint_num, compo_num, 32)
print(regress_net)
--regress_net:cuda()
--local inp = torch.CudaTensor(1, 3+joint_num+compo_num, 384, 480):fill(1)
--print(inp:size())
--for i=1, 1000 do
-- local outp = regress_net:forward(inp)
--end
local saveDir = '/home/xiaofei/workspace/toolPose/models'
--local modelConf = {type='toolPoseRegressFull', v='256*320_ftblr_head_noConcat' }
local modelConf = {type='toolPoseRegressFull', v='256*320_ftblr_random_head_noConcat'}
--
local saveID = modelConf.type .. '_v' .. modelConf.v
local initModelPath = paths.concat(saveDir, 'model.' .. saveID .. '.init.t7')
torch.save(paths.concat(saveDir, initModelPath), regress_net)
print('saved model ' .. saveID .. ' to ' .. paths.concat(saveDir, initModelPath))
-- toolPoseRegress v=1: inputsize = [384, 480], input only detMap(down_sample=4)
-- toolPoseRegressFull v=1: inputsize = [384, 480], input frame+detMap(down_sample=1)
-- toolPoseRegressFull v=2: inputsize = [256, 320], input frame+detMap(down_sample=1)
-- toolPoseRegressFull v=4: inputsize = [256, 320], input frame+detMap(round)(down_sample=1), without concat
-- toolPoseRegressFull