forked from nagadomi/kaggle-ndsb2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcreate_model.lua
216 lines (195 loc) · 7.44 KB
/
create_model.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
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
require 'pl'
local __FILE__ = (function() return string.gsub(debug.getinfo(2, 'S').source, "^@", "") end)()
package.path = path.join(path.dirname(__FILE__), "..", "lib", "?.lua;") .. package.path
require 'cunn'
local USE_CUDNN = false
local ALPHA = 0.1
local TAG_FEAT_N = 20
if USE_CUDNN then
require 'cudnn'
cudnn.benchmark = true
function cudnn.SpatialConvolution:reset(stdv)
stdv = math.sqrt(2 / ((1.0 + ALPHA * ALPHA) * self.kW * self.kH * self.nOutputPlane))
self.weight:normal(0, stdv)
self.bias:zero()
end
function cudnn.VolumetricConvolution:reset(stdv)
stdv = math.sqrt(2 / ((1.0 + ALPHA * ALPHA) * self.kT * self.kW * self.kH * self.nOutputPlane))
self.weight:normal(0, stdv)
self.bias:zero()
end
end
function nn.SpatialConvolution:reset(stdv)
stdv = math.sqrt(2 / ((1.0 + ALPHA * ALPHA) * self.kW * self.kH * self.nOutputPlane))
self.weight:normal(0, stdv)
self.bias:zero()
end
function nn.VolumetricConvolution:reset(stdv)
stdv = math.sqrt(2 / ((1.0 + ALPHA * ALPHA) * self.kT * self.kW * self.kH * self.nOutputPlane))
self.weight:normal(0, stdv)
self.bias:zero()
end
function nn.Linear:reset(stdv)
stdv = math.sqrt(2 / ((1.0 + ALPHA * ALPHA) * self.weight:size(2)))
self.weight:normal(0, stdv)
self.bias:zero()
end
function VolumetricConvolution(nInputPlane, nOutputPlane,
kT, kW, kH, dT, dW, dH, padT, padW, padH)
if USE_CUDNN then
return cudnn.VolumetricConvolution(nInputPlane, nOutputPlane,
kT, kW, kH, dT, dW, dH, padT, padW, padH)
else
return nn.VolumetricConvolution(nInputPlane, nOutputPlane,
kT, kW, kH, dT, dW, dH, padT, padW, padH)
end
end
function VolumetricMaxPooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
if USE_CUDNN then
return cudnn.VolumetricMaxPooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
else
return nn.VolumetricMaxPooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
end
end
function VolumetricAveragePooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
if USE_CUDNN then
return cudnn.VolumetricAveragePooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
else
return nn.VolumetricAveragePooling(kT, kW, kH, dT, dW, dH, padT, padW, padH)
end
end
local function create_model_1max()
local model = nn.Sequential()
local pt = nn.ParallelTable()
local cnn = nn.Sequential()
-- input: Bx16x3x64x64
cnn:add(VolumetricConvolution(16, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricMaxPooling(1, 2, 2))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 3, 3, 3, 1, 1, 1, 1, 1, 1)) -- conv 3d
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricAveragePooling(3, 32, 32))
cnn:add(nn.View(64))
cnn:add(nn.Dropout(0.5))
cnn:add(nn.Linear(64, 256))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(nn.Dropout(0.5))
cnn:add(nn.Linear(256, 16))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(nn.Linear(16, 1))
local tnet = nn.Sequential()
tnet:add(nn.Linear(TAG_FEAT_N, 16))
tnet:add(nn.ReLU())
tnet:add(nn.Linear(16, 16))
tnet:add(nn.ReLU())
tnet:add(nn.Linear(16, 1))
pt:add(cnn)
pt:add(tnet)
model:add(pt)
model:add(nn.CMulTable())
model:add(nn.View(1))
return model
end
local function create_model_2max()
local model = nn.Sequential()
local pt = nn.ParallelTable()
local cnn = nn.Sequential()
-- input: Bx16x3x64x64
cnn:add(VolumetricConvolution(16, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(64, 64, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricMaxPooling(1, 2, 2))
cnn:add(VolumetricConvolution(64, 96, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(96, 96, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricMaxPooling(1, 2, 2))
cnn:add(VolumetricConvolution(96, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 128, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(128, 256, 1, 3, 3, 1, 1, 1, 0, 1, 1))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricConvolution(256, 256, 3, 3, 3, 1, 1, 1, 1, 1, 1)) -- conv 3d
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(VolumetricAveragePooling(3, 16, 16))
cnn:add(nn.View(256))
cnn:add(nn.Dropout(0.5))
cnn:add(nn.Linear(256, 512))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(nn.Dropout(0.5))
cnn:add(nn.Linear(512, 16))
cnn:add(nn.LeakyReLU(ALPHA))
cnn:add(nn.Linear(16, 1))
-- input:
local tnet = nn.Sequential()
tnet:add(nn.Linear(TAG_FEAT_N, 16))
tnet:add(nn.ReLU())
tnet:add(nn.Linear(16, 16))
tnet:add(nn.ReLU())
tnet:add(nn.Linear(16, 1))
pt:add(cnn)
pt:add(tnet)
model:add(pt)
model:add(nn.CMulTable())
model:add(nn.View(1))
return model
end
local function create_model(model_id)
if model_id == 1 then
return create_model_1max():cuda()
else
return create_model_2max():cuda()
end
end
--model = create_model_1max()
--model:cuda()
--print(model:forward({torch.Tensor(32, 16, 3, 64, 64):uniform():cuda(), torch.Tensor(32, TAG_FEAT_N):cuda()}):size())
return create_model