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
|
require 'torch'
require 'nn'
if opt.type == 'cuda' then
require 'cunn'
end
----------------------------------------------------------------------
print(sys.COLORS.red .. '==> construct CNN')
-- This gets around 89.66% on test starting around epoc 140
-- 98.75% on training
--
-- th run.lua -r 0.29106 -sgdLearningRateDecay 0.0 --sgdMomentum 0.00152 -sgdWeightDecay 0.0 -s full --model version005 -p cuda -t 2
local function ConvBatchReLUDrop(CNN, input, output, dropFrac )
CNN:add(nn.SpatialConvolution(input, output, 3,3, 1,1, 1,1))
CNN:add(nn.SpatialBatchNormalization(output))
CNN:add(nn.ReLU(true))
if dropFrac > 0 then
CNN:add(nn.SpatialDropout(dropFrac))
end
end
-- http://torch.ch/blog/2015/07/30/cifar.html
local CNN = nn.Sequential()
ConvBatchReLUDrop(CNN,3,64,0.3)
ConvBatchReLUDrop(CNN,64,64,0.0)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 16
ConvBatchReLUDrop(CNN,64,128,0.4)
ConvBatchReLUDrop(CNN,128,128,0.0)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 8
ConvBatchReLUDrop(CNN,128,256,0.4)
ConvBatchReLUDrop(CNN,256,256,0.4)
ConvBatchReLUDrop(CNN,256,256,0.0)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 4
ConvBatchReLUDrop(CNN,256,512,0.4)
ConvBatchReLUDrop(CNN,512,512,0.4)
ConvBatchReLUDrop(CNN,512,512,0.0)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 2
ConvBatchReLUDrop(CNN,512,512,0.4)
ConvBatchReLUDrop(CNN,512,512,0.4)
ConvBatchReLUDrop(CNN,512,512,0.0)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 1
CNN:add(nn.View(512))
local classifier = nn.Sequential()
classifier:add(nn.Dropout(0.5))
classifier:add(nn.Linear(512, 512))
classifier:add(nn.BatchNormalization(512))
classifier:add(nn.ReLU())
classifier:add(nn.Dropout(0.5))
classifier:add(nn.Linear(512, 10))
local model = nn.Sequential()
model:add(CNN)
model:add(classifier)
-- Loss: NLL
local loss = nn.CrossEntropyCriterion()
-- initialization from MSR
local function MSRinit(net)
local function init(name)
for k,v in pairs(net:findModules(name)) do
local n = v.kW*v.kH*v.nOutputPlane
v.weight:normal(0,math.sqrt(2/n))
v.bias:zero()
end
end
init'nn.SpatialConvolution'
end
MSRinit(CNN)
----------------------------------------------------------------------
print(sys.COLORS.red .. '==> here is the CNN:')
print(model)
if opt.type == 'cuda' then
model:cuda()
loss:cuda()
end
-- return package:
return {
model = model,
loss = loss,
}
|
