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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
|
--[[
-- Tests the implementation of RNN binding using the cudnn v5 library. Cross-check the checksums with cudnn reference
-- sample checksums.
-- ]]
require 'cudnn'
require 'cunn'
local ffi = require 'ffi'
local errcheck = cudnn.errcheck
local cudnntest = torch.TestSuite()
local mytester
local tolerance = 300
function cudnntest.testRNNRELU()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 2
local rnn = cudnn.RNNReLU(hiddenSize, hiddenSize, numberOfLayers)
rnn.mode = 'CUDNN_RNN_RELU'
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 1.315793E+06, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 1.315212E+05, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 6.676003E+01, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 6.425067E+01, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 1.453750E+09, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testRNNBatchFirst()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 2
local batchFirst = true
local rnn = cudnn.RNNReLU(hiddenSize, hiddenSize, numberOfLayers, batchFirst)
rnn.mode = 'CUDNN_RNN_RELU'
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn, batchFirst)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 1.315793E+06, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 1.315212E+05, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 6.676003E+01, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 6.425067E+01, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 1.453750E+09, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testRNNTANH()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 2
local rnn = cudnn.RNNTanh(hiddenSize, hiddenSize, numberOfLayers)
rnn.mode = 'CUDNN_RNN_TANH'
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 6.319591E+05, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 6.319605E+04, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 4.501830E+00, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 4.489546E+00, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 5.012598E+07, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testRNNLSTM()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 8
local rnn = cudnn.LSTM(hiddenSize, hiddenSize, numberOfLayers)
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 5.749536E+05, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumc, 4.365091E+05, tolerance, 'checkSum with reference for localSumc failed')
mytester:assertalmosteq(checkSums.localSumh, 5.774818E+04, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 3.842206E+02, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdc, 9.323785E+03, tolerance, 'checkSum with reference for localSumdc failed')
mytester:assertalmosteq(checkSums.localSumdh, 1.182566E+01, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 4.313461E+08, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testRNNGRU()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 6
local rnn = cudnn.GRU(hiddenSize, hiddenSize, numberOfLayers)
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 6.358978E+05, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 6.281680E+04, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 6.296622E+00, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 2.289960E+05, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 5.397419E+07, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testBiDirectionalRELURNN()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 2
local nbDirections = 2
local batchFirst = false
local rnn = cudnn.RNN(hiddenSize, hiddenSize, numberOfLayers)
rnn.bidirectional = 'CUDNN_BIDIRECTIONAL'
rnn.mode = 'CUDNN_RNN_RELU'
rnn.numDirections = 2
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn, batchFirst, nbDirections)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 1.388634E+01, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 1.288997E+01, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 1.288729E+01, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 1.279004E+01, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 7.061081E+07, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testBiDirectionalTANHRNN()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 2
local nbDirections = 2
local batchFirst = false
local rnn = cudnn.RNN(hiddenSize, hiddenSize, numberOfLayers)
rnn.bidirectional = 'CUDNN_BIDIRECTIONAL'
rnn.mode = 'CUDNN_RNN_TANH'
rnn.numDirections = 2
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn, batchFirst, nbDirections)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 1.388634E+01, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 1.288997E+01, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 1.288729E+01, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 1.279004E+01, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 7.061081E+07, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testBiDirectionalLSTMRNN()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 8
local nbDirections = 2
local batchFirst = false
local rnn = cudnn.BLSTM(hiddenSize, hiddenSize, numberOfLayers)
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn, batchFirst, nbDirections)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 3.134097E+04, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumc, 3.845626E+00, tolerance, 'checkSum with reference for localSumc failed')
mytester:assertalmosteq(checkSums.localSumh, 1.922855E+00, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 4.794993E+00, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdc, 2.870925E+04, tolerance, 'checkSum with reference for localSumdc failed')
mytester:assertalmosteq(checkSums.localSumdh, 2.468645E+00, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 1.121568E+08, tolerance, 'checkSum with reference for localSumdw failed')
end
function cudnntest.testBiDirectionalGRURNN()
local miniBatch = 64
local seqLength = 20
local hiddenSize = 512
local numberOfLayers = 2
local numberOfLinearLayers = 6
local nbDirections = 2
local batchFirst = false
local rnn = cudnn.RNN(hiddenSize, hiddenSize, numberOfLayers)
rnn.bidirectional = 'CUDNN_BIDIRECTIONAL'
rnn.mode = 'CUDNN_GRU'
rnn.numDirections = 2
local checkSums = getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn, batchFirst, nbDirections)
-- Checksums to check against are retrieved from cudnn RNN sample.
mytester:assertalmosteq(checkSums.localSumi, 6.555183E+04, tolerance, 'checkSum with reference for localsumi failed')
mytester:assertalmosteq(checkSums.localSumh, 5.830924E+00, tolerance, 'checkSum with reference for localSumh failed')
mytester:assertalmosteq(checkSums.localSumdi, 4.271801E+00, tolerance, 'checkSum with reference for localSumdi failed')
mytester:assertalmosteq(checkSums.localSumdh, 6.555744E+04, tolerance, 'checkSum with reference for localSumdh failed')
mytester:assertalmosteq(checkSums.localSumdw, 1.701796E+08, tolerance, 'checkSum with reference for localSumdw failed')
end
--[[
-- Method gets Checksums of RNN to compare with ref Checksums in cudnn RNN C sample.
-- ]]
function getRNNCheckSums(miniBatch, seqLength, hiddenSize, numberOfLayers, numberOfLinearLayers, rnn, batchFirst, nbDirections)
local biDirectionalScale = nbDirections or 1
-- Reset the rnn and weight descriptor (since we are manually setting values for matrix/bias.
rnn:reset()
rnn:resetWeightDescriptor()
local input
if (batchFirst) then
input = torch.CudaTensor(miniBatch, seqLength, hiddenSize):fill(1)
else
input = torch.CudaTensor(seqLength, miniBatch, hiddenSize):fill(1) -- Input initialised to 1s.
end
local weights = rnn:weights()
local biases = rnn:biases()
-- Matrices are initialised to 1 / matrixSize, biases to 1 unless bi-directional.
for layer = 1, numberOfLayers do
for layerId = 1, numberOfLinearLayers do
if (biDirectionalScale == 2) then
rnn.weight:fill(1 / rnn.weight:size(1))
else
local weightTensor = weights[layer][layerId]
weightTensor:fill(1.0 / weightTensor:size(1))
local biasTensor = biases[layer][layerId]
biasTensor:fill(1)
end
end
end
-- Set hx/cx/dhy/dcy data to 1s.
rnn.hiddenInput = torch.CudaTensor(numberOfLayers * biDirectionalScale, miniBatch, hiddenSize):fill(1)
rnn.cellInput = torch.CudaTensor(numberOfLayers * biDirectionalScale, miniBatch, hiddenSize):fill(1)
rnn.gradHiddenOutput = torch.CudaTensor(numberOfLayers * biDirectionalScale, miniBatch, hiddenSize):fill(1)
rnn.gradCellOutput = torch.CudaTensor(numberOfLayers * biDirectionalScale, miniBatch, hiddenSize):fill(1)
local testOutputi = rnn:forward(input)
-- gradInput set to 1s.
local gradInput
if(batchFirst) then
gradInput = torch.CudaTensor(miniBatch, seqLength, hiddenSize * biDirectionalScale):fill(1)
else
gradInput = torch.CudaTensor(seqLength, miniBatch, hiddenSize * biDirectionalScale):fill(1)
end
rnn:backward(input, gradInput)
-- Sum up all values for each.
local localSumi = torch.sum(testOutputi)
local localSumh = torch.sum(rnn.hiddenOutput)
local localSumc = torch.sum(rnn.cellOutput)
local localSumdi = torch.sum(rnn.gradInput)
local localSumdh = torch.sum(rnn.gradHiddenInput)
local localSumdc = torch.sum(rnn.gradCellInput)
local localSumdw = torch.sum(rnn.gradWeight)
local checkSums = {
localSumi = localSumi,
localSumh = localSumh,
localSumc = localSumc,
localSumdi = localSumdi,
localSumdh = localSumdh,
localSumdc = localSumdc,
localSumdw = localSumdw
}
return checkSums
end
mytester = torch.Tester()
mytester:add(cudnntest)
mytester:run()
|
