-
Notifications
You must be signed in to change notification settings - Fork 8
/
BlockMixture.lua
148 lines (130 loc) · 4.61 KB
/
BlockMixture.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
local BlockMixture, parent = torch.class('nn.BlockMixture', 'nn.Module')
------------------------------------------------------------------------
--[[ BlockMixture ]]--
-- n > 1 BlockSparse Modules gated by a one gater with
-- n - 1 output spaces, one for each of the hidden layers
------------------------------------------------------------------------
function BlockMixture:__init(experts, gater, expertScale, gaterScale)
parent.__init(self)
self.gater = gater
assert(#experts > 1, "need at least 2 experts")
self.experts = experts
self.expertScale = expertScale or 1
self.gaterScale = gaterScale or 1
self.modules = {gater}
for i, expert in ipairs(self.experts) do
table.insert(self.modules, expert)
end
self.batchSize = 0
end
function BlockMixture:updateOutput(input)
if #self.experts == 2 then
self.gaterOutputs = {self.gater:updateOutput(input)}
else
self.gaterOutputs = self.gater:updateOutput(input)
end
self.expertInputs = {{input, self.gaterOutputs[1]}}
for i=1,#self.experts - 1 do
self.expertInputs[i+1] = {self.experts[i]:updateOutput(self.expertInputs[i]), self.gaterOutputs[i+1]}
end
self.output = self.experts[#self.experts]:updateOutput(self.expertInputs[#self.expertInputs][1])
return self.output
end
function BlockMixture:updateGradInput(input, gradOutput)
self.expertGradInputs = {}
self.expertGradInputs[#self.experts] = self.experts[#self.experts]:updateGradInput(self.expertInputs[#self.expertInputs][1], gradOutput)
for i=#self.experts-1,1,-1 do
self.expertGradInputs[i] = self.experts[i]:updateGradInput(self.expertInputs[i], self.expertGradInputs[i+1][1])
end
self.gaterGradOutputs = {}
for i=1,#self.gaterOutputs do
self.gaterGradOutputs[i] = self.expertGradInputs[i][2]
end
if #self.experts == 2 then
self._gradInput = self.gater:updateGradInput(input, self.gaterGradOutputs[1])
else
self._gradInput = self.gater:updateGradInput(input, self.gaterGradOutputs)
end
self.gradInput:resizeAs(input)
self.gradInput:copy(self.expertGradInputs[1][1])
self.gradInput:add(self._gradInput)
return self.gradInput
end
function BlockMixture:accGradParameters(input, gradOutput, scale)
scale = scale or 1
if self.expertScale > 0 then
self.experts[#self.experts]:accGradParameters(self.expertInputs[#self.expertInputs][1], gradOutput, scale)
for i=#self.experts-1,1,-1 do
self.experts[i]:accGradParameters(self.expertInputs[i], self.expertGradInputs[i+1][1], scale)
end
end
if self.gaterScale > 0 then
if #self.experts == 2 then
self.gater:accGradParameters(input, self.gaterGradOutputs[1], scale)
else
self.gater:accGradParameters(input, self.gaterGradOutputs, scale)
end
end
end
function BlockMixture:accUpdateGradParameters(input, gradOutput, lr)
if self.expertScale > 0 then
self.experts[#self.experts]:accUpdateGradParameters(self.expertInputs[#self.expertInputs][1], gradOutput, lr)
for i=#self.experts-1,1,-1 do
self.experts[i]:accUpdateGradParameters(self.expertInputs[i], self.expertGradInputs[i+1][1], lr*self.expertScale)
end
end
if self.gaterScale > 0 then
if #self.experts == 2 then
self.gater:accUpdateGradParameters(input, self.gaterGradOutputs[1], lr*self.gaterScale)
else
self.gater:accUpdateGradParameters(input, self.gaterGradOutputs, lr*self.gaterScale)
end
end
end
function BlockMixture:zeroGradParameters()
for i,module in ipairs(self.modules) do
module:zeroGradParameters()
end
end
function BlockMixture:updateParameters(learningRate)
if self.expertScale > 0 then
for i,expert in ipairs(self.experts) do
expert:updateParameters(learningRate*self.expertScale)
end
end
if self.gaterScale > 0 then
self.gater:updateParameters(learningRate*self.gaterScale)
end
end
function BlockMixture:share(mlp,...)
for i=1,#self.modules do
self.modules[i]:share(mlp.modules[i],...);
end
end
function BlockMixture:parameters()
local function tinsert(to, from)
if type(from) == 'table' then
for i=1,#from do
tinsert(to,from[i])
end
else
table.insert(to,from)
end
end
local w = {}
local gw = {}
for i=1,#self.modules do
local mw,mgw = self.modules[i]:parameters()
if mw then
tinsert(w,mw)
tinsert(gw,mgw)
end
end
return w,gw
end
function BlockMixture:type(type)
for i,module in ipairs(self.modules) do
module:type(type)
end
self.gradInput = self.gradInput:type(type)
end