-
Notifications
You must be signed in to change notification settings - Fork 88
/
decoder.lua
148 lines (120 loc) · 4.32 KB
/
decoder.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
----------------------------------------------------------------------
-- Create model and calulate loss to optimize for decoder.
--
-- Adam Paszke,
-- May 2016.
----------------------------------------------------------------------
require 'torch' -- torch
require 'image' -- to visualize the dataset
torch.setdefaulttensortype('torch.FloatTensor')
----------------------------------------------------------------------
print '==> define parameters'
local histClasses = opt.datahistClasses
local classes = opt.dataClasses
----------------------------------------------------------------------
print '==> construct model'
-- encoder CNN:
nn.DataParallelTable.deserializeNGPUs = 1
model = torch.load(opt.CNNEncoder)
if torch.typename(model) == 'nn.DataParallelTable' then model = model:get(1) end
model:remove(#model.modules) -- remove the classifier
-- SpatialMaxUnpooling requires nn modules...
model:apply(function(module)
if module.modules then
for i,submodule in ipairs(module.modules) do
if torch.typename(submodule):match('cudnn.SpatialMaxPooling') then
module.modules[i] = nn.SpatialMaxPooling(2, 2, 2, 2) -- TODO: make more flexible
end
end
end
end)
-- find pooling modules
local pooling_modules = {}
model:apply(function(module)
if torch.typename(module):match('nn.SpatialMaxPooling') then
table.insert(pooling_modules, module)
end
end)
assert(#pooling_modules == 3, 'There should be 3 pooling modules')
-- kill gradient
-- local grad_killer = nn.Identity()
-- function grad_killer:updateGradInput(input, gradOutput)
-- return self.gradInput:resizeAs(gradOutput):zero()
-- end
-- model:add(grad_killer)
-- decoder:
print(pooling_modules)
function bottleneck(input, output, upsample, reverse_module)
local internal = output / 4
local input_stride = upsample and 2 or 1
local module = nn.Sequential()
local sum = nn.ConcatTable()
local main = nn.Sequential()
local other = nn.Sequential()
sum:add(main):add(other)
main:add(cudnn.SpatialConvolution(input, internal, 1, 1, 1, 1, 0, 0):noBias())
main:add(nn.SpatialBatchNormalization(internal, 1e-3))
main:add(cudnn.ReLU(true))
if not upsample then
main:add(cudnn.SpatialConvolution(internal, internal, 3, 3, 1, 1, 1, 1))
else
main:add(nn.SpatialFullConvolution(internal, internal, 3, 3, 2, 2, 1, 1, 1, 1))
end
main:add(nn.SpatialBatchNormalization(internal, 1e-3))
main:add(cudnn.ReLU(true))
main:add(cudnn.SpatialConvolution(internal, output, 1, 1, 1, 1, 0, 0):noBias())
main:add(nn.SpatialBatchNormalization(output, 1e-3))
other:add(nn.Identity())
if input ~= output or upsample then
other:add(cudnn.SpatialConvolution(input, output, 1, 1, 1, 1, 0, 0):noBias())
other:add(nn.SpatialBatchNormalization(output, 1e-3))
if upsample and reverse_module then
other:add(nn.SpatialMaxUnpooling(reverse_module))
end
end
if upsample and not reverse_module then
main:remove(#main.modules) -- remove BN
return main
end
return module:add(sum):add(nn.CAddTable()):add(cudnn.ReLU(true))
end
--model:add(bottleneck(128, 128))
model:add(bottleneck(128, 64, true, pooling_modules[3])) -- 32x64
model:add(bottleneck(64, 64))
model:add(bottleneck(64, 64))
model:add(bottleneck(64, 16, true, pooling_modules[2])) -- 64x128
model:add(bottleneck(16, 16))
model:add(nn.SpatialFullConvolution(16, #classes, 2, 2, 2, 2))
if cutorch.getDeviceCount() > 1 then
local gpu_list = {}
for i = 1,cutorch.getDeviceCount() do gpu_list[i] = i end
model = nn.DataParallelTable(1):add(model:cuda(), gpu_list)
print(opt.nGPU .. " GPUs being used")
end
-- Loss: NLL
print('defining loss function:')
local normHist = histClasses / histClasses:sum()
local classWeights = torch.Tensor(#classes):fill(1)
for i = 1, #classes do
-- Ignore unlabeled and egoVehicle
if i == 1 then
classWeights[i] = 0
end
if histClasses[i] < 1 then
print("Class " .. tostring(i) .. " not found")
classWeights[i] = 0
else
classWeights[i] = 1 / (torch.log(1.02 + normHist[i]))
end
end
loss = cudnn.SpatialCrossEntropyCriterion(classWeights)
model:cuda()
loss:cuda()
----------------------------------------------------------------------
print '==> here is the model:'
print(model)
-- return package:
return {
model = model,
loss = loss,
}