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
|
require 'torch'
require 'nn'
if opt.type == 'cuda' then
require 'cunn'
end
----------------------------------------------------------------------
print(sys.COLORS.red .. '==> construct CNN')
-- This gets around ????% on test starting around epoc ??
-- ????% on training
--
-- th run.lua
local function ConvBatchDropReLU(CNN, input, output, dropFrac )
CNN:add(nn.SpatialConvolution(input, output, 3,3, 1,1, 1,1))
CNN:add(nn.SpatialBatchNormalization(output))
CNN:add(nn.SpatialDropout(dropFrac))
CNN:add(nn.ReLU())
end
local CNN = nn.Sequential()
ConvBatchDropReLU(CNN,3,64,0.2)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 16
ConvBatchDropReLU(CNN,64,128,0.4)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 8
ConvBatchDropReLU(CNN,128,256,0.4)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 4
ConvBatchDropReLU(CNN,256,512,0.4)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 2
ConvBatchDropReLU(CNN,512,512,0.4)
CNN:add(nn.SpatialMaxPooling(2,2,2,2)) -- 1
CNN:add(nn.View(512))
CNN:add(nn.Dropout(0.5))
local classifier = nn.Sequential()
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))
classifier:add(nn.LogSoftMax())
local model = nn.Sequential()
model:add(CNN)
model:add(classifier)
-- Loss: NLL
local loss = nn.ClassNLLCriterion()
----------------------------------------------------------------------
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,
}
|
