-
Notifications
You must be signed in to change notification settings - Fork 8
/
BlockSparse.lua
214 lines (188 loc) · 7.51 KB
/
BlockSparse.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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
local BlockSparse, parent = torch.class('nn.BlockSparse', 'nn.Module')
------------------------------------------------------------------------
--[[ BlockSparse ]]--
-- Use for Distributed Conditional Computation
-- Inputs and outputs are sparse
-- Weights are organized as a matrix of blocks.
------------------------------------------------------------------------
-- 1. Dense input, sparse output:
-- Input : {activation, {outputIndices, outputScales}}
-- Output : {activation, {outputIndices, outputScales}}
-- 2. Sparse input, sparse output:
-- Input : {{activation, {inputIndices, inputScales}}, {outputIndices, outputScales}}
-- Output : {activation, {inputIndices, inputScales}}
-- 3. Sparse input, dense output:
-- Input : {activation, {inputIndice, inputScale}}
-- Output : tensor of activations.
function BlockSparse:__init(nInputBlock, inputSize, nOutputBlock, outputSize, accUpdate)
parent.__init(self)
self.nInputBlock = nInputBlock
self.nOutputBlock = nOutputBlock
self.inputSize = inputSize
self.outputSize = outputSize
self.accUpdate = accUpdate or false
self.weight = torch.Tensor(nOutputBlock, nInputBlock, outputSize, inputSize)
self.bias = torch.Tensor(nOutputBlock, outputSize)
if not self.accUpdate then
self.gradWeight = torch.Tensor(nOutputBlock, nInputBlock, outputSize, inputSize):zero()
self.gradBias = torch.Tensor(nOutputBlock, outputSize):zero()
end
-- sqrt(inputWindowSize*outputWindowSize) smaller than this use
-- cublasSgemmBatched.
self.batchedGemmMax = 200
-- for dense inputs or outputs
self.inputIndice = torch.Tensor()
self.outputIndice = torch.Tensor()
self.inputScale = torch.Tensor()
self.outputScale = torch.Tensor()
-- for cuda
self.inputIndiceHost = torch.LongTensor()
self.outputIndiceHost = torch.LongTensor()
self.inputHost = torch.CharTensor()
self.weightHost = torch.CharTensor()
self.outputHost = torch.CharTensor()
-- etc
self._output = torch.Tensor()
self.gradOutputScale = torch.Tensor()
self._gradInput = torch.Tensor()
self.gradInput = {}
self.batchSize = 0
self:reset()
end
function BlockSparse:reset(stdv)
if stdv then
stdv = stdv * math.sqrt(3)
else
stdv = 1/math.sqrt(self.nInputBlock*0.1*self.inputSize)
end
self.weight:uniform(-stdv, stdv)
self.bias:uniform(-stdv, stdv)
end
function BlockSparse:updateOutput(inputTable)
local input, inputIndice, outputIndice, inputScale, outputScale = self:unpackInput(inputTable)
if batchSize ~= input:size(1) then
self.inputIndice:resize(input:size(1),1):fill(1)
self.outputIndice:resize(input:size(1),1):fill(1)
self.inputScale:resize(input:size(1),1):fill(1)
self.outputScale:resize(input:size(1),1):fill(1)
self.batchSize = input:size(1)
self.inputWindowSize = inputIndice:size(2)
self.outputWindowSize = outputIndice:size(2)
end
local output = input.nn.BlockSparse_updateOutput(
self, input, inputIndice, outputIndice, inputScale, outputScale
)
self.output = self:packOutput(output, outputIndice, outputScale)
return self.output
end
function BlockSparse:updateGradInput(inputTable, gradOutputTable)
local input, inputIndice, outputIndice, inputScale, outputScale = self:unpackInput(inputTable)
local gradOutput = self:unpackGradOutput(gradOutputTable)
local gradInput, gradOutputScale = input.nn.BlockSparse_updateGradInput(
self, input, inputIndice, outputIndice, inputScale, outputScale, gradOutput
)
self:packGradInput(outputIndice, gradInput, gradOutputScale)
return self.gradInput
end
function BlockSparse:accGradParameters(inputTable, gradOutputTable, scale)
local input, inputIndice, outputIndice, inputScale, outputScale = self:unpackInput(inputTable)
local gradOutput = self:unpackGradOutput(gradOutputTable)
scale = scale or 1
input.nn.BlockSparse_accGradParameters(
self, input, inputIndice, outputIndice, inputScale, outputScale, gradOutput, scale
)
end
function BlockSparse:getBlockParameters(inputIdx, outputIdx)
local weight = self.weight[outputIdx][inputIdx]
local bias = self.bias[outputIdx]
local gradWeight = self.gradWeight[outputIdx][inputIdx]
local gradBias = self.gradBias[outputIdx]
return {weight, bias}, {gradWeight, gradBias}
end
function BlockSparse:type(type)
if type and (type == 'torch.FloatTensor' or type == 'torch.DoubleTensor' or type == 'torch.CudaTensor') then
self.weight = self.weight:type(type)
self.bias = self.bias:type(type)
if not self.accUpdate then
self.gradWeight = self.gradWeight:type(type)
self.gradBias = self.gradBias:type(type)
end
self._output = self._output:type(type)
self._gradInput = self._gradInput:type(type)
self.inputIndice = self.inputIndice:type(type)
self.outputIndice = self.outputIndice:type(type)
self.inputScale = self.inputScale:type(type)
self.outputScale = self.outputScale:type(type)
self.gradOutputScale = self.gradOutputScale:type(type)
if type == 'torch.CudaTensor' then
self.inputCuda = torch.CudaTensor()
self.weightCuda = torch.CudaTensor()
self.outputCuda = torch.CudaTensor()
self.outputBatched = torch.CudaTensor()
self.gradInputBatched = torch.CudaTensor()
self._gradOutput = torch.CudaTensor()
end
end
return self
end
-- generate a Clone that shares parameters and metadata
-- without wasting memory
function BlockSparse:sharedClone()
error"NotImplemented"
end
-- we do not need to accumulate parameters when sharing
BlockSparse.sharedAccUpdateGradParameters = BlockSparse.accUpdateGradParameters
function BlockSparse:unpackInput(inputTable)
local input, inputIndice, outputIndice, inputScale, outputScale, innerTable
if self.nInputBlock == 1 then
input, innerTable = unpack(inputTable)
outputIndice, outputScale = unpack(innerTable)
inputIndice = self.inputIndice
inputScale = self.inputScale
elseif self.nOutputBlock == 1 then
input, innerTable = unpack(inputTable)
inputIndice, inputScale = unpack(innerTable)
outputIndice = self.outputIndice
outputScale = self.outputScale
else
input, innerTable = unpack(inputTable[1])
inputIndice, inputScale = unpack(innerTable)
outputIndice, outputScale = unpack(inputTable[2])
end
return input, inputIndice, outputIndice, inputScale, outputScale
end
function BlockSparse:unpackGradOutput(gradOutputTable)
local gradOutput
if self.nInputBlock == 1 then
gradOutput = gradOutputTable[1]
elseif self.nOutputBlock == 1 then
gradOutput = gradOutputTable
else
gradOutput = gradOutputTable[1]
end
return gradOutput
end
function BlockSparse:packGradInput(outputIndice, gradInput, gradOutputScale)
local gradInputTable = self.gradInput
if self.nInputBlock == 1 then
gradInputTable[1] = gradInput
gradInputTable[2] = {outputIndice, gradOutputScale}
elseif self.nOutputBlock == 1 then
gradInputTable[1] = gradInput
gradInputTable[2] = {outputIndice, gradOutputScale}
else
gradInputTable[1] = {gradInput}
gradInputTable[2] = {outputIndice, gradOutputScale}
end
end
function BlockSparse:packOutput(output, outputIndice, outputScale)
local outputTable
if self.nInputBlock == 1 then
outputTable = {output, {outputIndice, outputScale}}
elseif self.nOutputBlock == 1 then
outputTable = output
else
outputTable = {output, {outputIndice, outputScale}}
end
return outputTable
end