/
model_big.lua
181 lines (174 loc) · 6.87 KB
/
model_big.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
require 'cunn'
require 'cudnn'
require 'nngraph'
require 'nnx'
require 'Reparametrize'
require 'model'
function getEncoder(nFeatures, sSize, zSize, zSample, sSample,
nChannels, hInput, wInput)
-- Returns an encoder. It expects inputs of size nChannels*hInput*wInput,
-- and produces a code z of size (2,nFeatures*hInput*wInput/8), where
-- the first coordinate is the mean and the second is the variance,
-- and a code s of size sSize.
local input = nn.Identity()()
local x = input
assert((hInput <= 32) and (wInput <= 32), 'input size must be <= 32x32')
if (hInput < 32) or (wInput < 32) then
local padh, padw = 32 - hInput, 32 - wInput
x = nn.SpatialZeroPadding(math.floor(padw/2), math.ceil(padw/2),
math.floor(padh/2), math.ceil(padh/2))(x)
hInput, wInput = 32, 32
end
local x = convblock(nChannels , nFeatures , 3, 1)(x) -- /1
local x = convblock(nFeatures , nFeatures , 3, 1)(x) -- /1
local x = convblock(nFeatures , nFeatures*2, 2, 2)(x) -- /2
local x = convblock(nFeatures*2, nFeatures*2, 3, 1)(x) -- /2
local x = convblock(nFeatures*2, nFeatures*4, 2, 2)(x) -- /4
local x = convblock(nFeatures*4, nFeatures*4, 3, 1)(x) -- /4
local x = convblock(nFeatures*4, nFeatures*8, 2, 2)(x) -- /8
local x = convblock(nFeatures*8, nFeatures*8, 3, 1)(x) -- /8
local npix = nFeatures*8*hInput*wInput/8/8
local x = nn.View(npix):setNumInputDims(3)(x)
local s, z
if zSample then
z = nn.Linear(npix, zSize*2):cuda()(x)
z = nn.View(2, zSize):setNumInputDims(1)(z)
else
z = nn.Linear(npix, zSize):cuda()(x)
end
if sSample then
s = nn.Linear(npix, sSize*2):cuda()(x)
s = nn.View(2, sSize):setNumInputDims(1)(s)
else
s = nn.Linear(npix, sSize):cuda()(x)
end
return nn.gModule({input}, {z, s}):cuda()
end
function getDecoder(nFeatures, zSize, sSize, zSample, sSample,
nChannels, hInput, wInput)
local inputz, inputs = nn.Identity()(), nn.Identity()()
local z, s = inputz, inputs
if zSample then
z = nn.SplitTable(2)(z)
z = nn.Reparametrize(zSize)(z)
end
if sSample then
s = nn.SplitTable(2)(s)
s = nn.Reparametrize(sSize)(s)
end
local npix = nFeatures*8*4*4
z = nn.Linear(zSize, npix):cuda()(z)
s = nn.Linear(sSize, npix):cuda()(s)
local x = nn.CAddTable(){z, s}
local x = nn.View(nFeatures*8, 4, 4)(x)
local x = convblock(nFeatures*8, nFeatures*8, 3, 1)(x)
local x = convblock(nFeatures*8, nFeatures*4, 2, 1/2)(x)
local x = convblock(nFeatures*4, nFeatures*4, 3, 1)(x)
local x = convblock(nFeatures*4, nFeatures*2, 2, 1/2)(x)
local x = convblock(nFeatures*2, nFeatures*2, 3, 1)(x)
local x = convblock(nFeatures*2, nFeatures , 2, 1/2)(x)
local x = convblock(nFeatures , nFeatures , 3, 1)(x)
local x = conv (nFeatures , nChannels , 3, 1)(x)
assert((hInput <= 32) and (wInput <= 32), 'input size must be <= 32x32')
if (hInput < 32) or (wInput < 32) then
local padh, padw = 32 - hInput, 32 - wInput
x = nn.SpatialZeroPadding(-math.floor(padw/2), -math.ceil(padw/2),
-math.floor(padh/2), -math.ceil(padh/2))(x)
end
local x = nn.Tanh()(x)
return nn.gModule({inputz, inputs}, {x}):cuda()
end
function getDiscriminator(nFeatures, nSid, nChannels, hInput, wInput)
if type(nSid) == 'table' then
assert(nFeatures*hInput*wInput/8/8 % #nSid == 0,
"nFeatures must be a multiple of " .. #nSid .. " (multi-class)")
end
local inputpix, inputsid = nn.Identity()(), nn.Identity()()
local x = inputpix
local sid = nn.FunctionWrapper(
function(self) self.gradInput = torch.CudaTensor() end,
function(self, input) return input end,
function(self, input, gradOutput)
self.gradInput:resizeAs(input):zero()
return self.gradInput
end)(inputsid) --because lookuptable produce no gradient
assert((hInput <= 32) and (wInput <= 32), 'input size must be <= 32x32')
if (hInput < 32) or (wInput < 32) then
local padh, padw = 32 - hInput, 32 - wInput
x = nn.SpatialZeroPadding(math.floor(padw/2), math.ceil(padw/2),
math.floor(padh/2), math.ceil(padh/2))(x)
hInput, wInput = 32, 32
end
local function getLookup(n, upsampling)
local out = nn.Sequential()
if type(nSid) == 'table' then
out:add(getMultiLookupTable(nSid, nFeatures*n*hInput*wInput/8/8/#nSid))
else
out:add(nn.LookupTable(nSid, nFeatures*n*hInput*wInput/8/8))
end
out:add(nn.View(nFeatures*n, hInput/8, wInput/8):setNumInputDims(1))
if upsampling ~= 1 then
out:add(nn.SpatialUpSamplingNearest(upsampling))
end
return out
end
local x = conv(nChannels, nFeatures, 3, 1)(x)
local x = nn.LeakyReLU(0.2)(x)
local x = conv(nFeatures, nFeatures, 2, 2)(x)
local lut1 = getLookup(1, 4)(sid)
local x = nn.CAddTable(){x, lut1}
local x = nn.SpatialBatchNormalization(nFeatures):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local x = conv(nFeatures, nFeatures*2, 2, 2)(x)
local x = nn.SpatialBatchNormalization(nFeatures*2):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local x = conv(nFeatures*2, nFeatures*2, 3, 1)(x)
local x = nn.SpatialBatchNormalization(nFeatures*2):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local lut2 = getLookup(2, 2)(sid)
local x = nn.CAddTable(){x, lut2}
local x = nn.SpatialBatchNormalization(nFeatures*2):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local x = conv(nFeatures*2, nFeatures*4, 2, 2)(x)
local x = nn.SpatialBatchNormalization(nFeatures*4):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local x = conv(nFeatures*4, nFeatures*4, 3, 1)(x)
local lut3 = getLookup(4, 1)(sid)
local x = nn.CAddTable(){x, lut3}
local x = nn.SpatialBatchNormalization(nFeatures*4):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local x = conv(nFeatures*4, nFeatures*4, 3, 1)(x)
local x = nn.SpatialBatchNormalization(nFeatures*4):cuda()(x)
local x = nn.LeakyReLU(0.2)(x)
local npix = nFeatures*4*hInput*wInput/8/8
local x = nn.View(npix):setNumInputDims(3)(x)
local x = nn.Dropout()(x)
local x = nn.Linear(npix, 1):cuda()(x)
local x = nn.Sigmoid()(x)
return nn.gModule({inputpix, inputsid}, {x}):cuda()
end
function getClassifierFromS(sSize, nClasses, sSample, nHidden, dropout)
local classifier = nn.Sequential()
if sSample then
classifier:add(nn.SplitTable(2))
classifier:add(nn.Reparametrize(sSize))
end
classifier:add(nn.Linear(sSize, nHidden):cuda())
--classifier:add(nn.BatchNormalization(nHidden, nil, nil, false))
classifier:add(nn.ReLU())
classifier:add(nn.Dropout())
if type(nClasses) == 'table' then
local ct = nn.ConcatTable()
classifier:add(ct)
for i = 1, #nClasses do
local seq = nn.Sequential()
ct:add(seq)
seq:add(nn.Linear(nHidden, nClasses[i]):cuda())
seq:add(nn.LogSoftMax())
end
else
classifier:add(nn.Linear(nHidden, nClasses):cuda())
classifier:add(nn.LogSoftMax())
end
return classifier:cuda()
end