-
Notifications
You must be signed in to change notification settings - Fork 82
/
NCEModule.lua
439 lines (368 loc) · 16 KB
/
NCEModule.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
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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
------------------------------------------------------------------------
--[[ Noise Contrast Estimation Module]]--
-- Ref.: A. https://www.cs.toronto.edu/~amnih/papers/ncelm.pdf
------------------------------------------------------------------------
local _ = require 'moses'
local NCEModule, parent = torch.class("nn.NCEModule", "nn.Linear")
NCEModule.version = 6 -- better bias init
-- for efficient serialization using nn.Serial
local empty = _.clone(parent.dpnn_mediumEmpty)
table.insert(empty, 'sampleidx')
table.insert(empty, 'sampleprob')
table.insert(empty, '_noiseidx')
table.insert(empty, '_noiseprob')
table.insert(empty, '_weight')
table.insert(empty, '_gradWeight')
table.insert(empty, '_gradOutput')
table.insert(empty, '_tgradOutput')
NCEModule.dpnn_mediumEmpty = empty
-- for sharedClone
local params = _.clone(parent.dpnn_parameters)
table.insert(params, 'unigrams')
table.insert(params, 'Z')
NCEModule.dpnn_parameters = params
function NCEModule:__init(inputSize, outputSize, k, unigrams, Z)
parent.__init(self, inputSize, outputSize)
assert(torch.type(k) == 'number')
assert(torch.isTensor(unigrams))
self.k = k
self.unigrams = unigrams
self.Z = torch.Tensor{Z or -1}
self.batchnoise = true
self:fastNoise()
-- output is {P_linear(target|input), P_linear(samples|input), P_noise(target), P_noise(samples)}
self.output = {torch.Tensor(), torch.Tensor(), torch.Tensor(), torch.Tensor()}
self.gradInput = {torch.Tensor(), torch.Tensor()}
end
function NCEModule:reset(stdv)
if stdv then
self.weight:uniform(-stdv, stdv)
self.bias:uniform(-stdv, stdv)
else
stdv = stdv or 1./math.sqrt(self.weight:size(2))
self.weight:uniform(-stdv, stdv)
-- this is useful for Z = 1
self.bias:fill(-math.log(self.bias:size(1)))
end
return self
end
function NCEModule:fastNoise()
-- we use alias to speedup multinomial sampling (see noiseSample method)
require 'torchx'
assert(torch.AliasMultinomial, "update torchx : luarocks install torchx")
self.unigrams:div(self.unigrams:sum())
self.aliasmultinomial = torch.AliasMultinomial(self.unigrams)
self.aliasmultinomial.dpnn_parameters = {'J', 'q'}
end
function NCEModule:updateOutput(inputTable)
local input, target = unpack(inputTable)
assert(input:dim() == 2)
assert(target:dim() == 1)
local batchsize = input:size(1)
local inputsize = self.weight:size(2)
if self.train == false and self.normalized then
self.linout = self.linout or input.new()
-- full linear + softmax
local nElement = self.linout:nElement()
self.linout:resize(batchsize, self.weight:size(1))
if self.linout:nElement() ~= nElement then
self.linout:zero()
end
self.addBuffer = self.addBuffer or input.new()
if self.addBuffer:nElement() ~= batchsize then
self.addBuffer:resize(batchsize):fill(1)
end
self.weight.addmm(self.linout, 0, self.linout, 1, input, self.weight:t())
if self.bias then self.linout:addr(1, self.addBuffer, self.bias) end
self.output = torch.type(self.output) == 'table' and input.new() or self.output
if self.logsoftmax then
input.THNN.LogSoftMax_updateOutput(
self.linout:cdata(),
self.output:cdata()
)
else
input.THNN.SoftMax_updateOutput(
self.linout:cdata(),
self.output:cdata()
)
end
elseif self.batchnoise then
self.output = (torch.type(self.output) == 'table' and #self.output == 4) and self.output
or {input.new(), input.new(), input.new(), input.new()}
assert(torch.type(target) == 'torch.CudaTensor' or torch.type(target) == 'torch.LongTensor')
self.sampleidx = self.sampleidx or target.new()
-- the last elements contain the target indices
self.sampleidx:resize(self.k + batchsize)
self.sampleidx:narrow(1,self.k+1,batchsize):copy(target)
-- sample k noise samples
self:noiseSample(self.sampleidx, 1, self.k)
self.sampleidx:resize(self.k + batchsize)
-- build (batchsize+k, inputsize) weight tensor
self._weight = self._weight or self.bias.new()
self.weight.index(self._weight, self.weight, 1, self.sampleidx)
assert(self._weight:nElement() == (self.k+batchsize)*inputsize)
self._weight:resize(self.k+batchsize, inputsize)
-- build (batchsize+k,) bias tensor
self._bias = self._bias or self.bias.new()
self._bias:index(self.bias, 1, self.sampleidx)
assert(self._bias:nElement() == (self.k+batchsize))
self._bias:resize(self.k+batchsize)
-- separate sample and target weight matrices and bias vectors
local sweight = self._weight:narrow(1, 1, self.k)
local tweight = self._weight:narrow(1, self.k+1, batchsize)
local sbias = self._bias:narrow(1, 1, self.k)
local tbias = self._bias:narrow(1, self.k+1, batchsize)
-- get model probability of targets (batchsize,)
local Pmt = self.output[1]
self._pm = self._pm or input.new()
self._pm:cmul(input, tweight)
Pmt:sum(self._pm, 2):resize(batchsize)
Pmt:add(tbias)
Pmt:exp()
-- get model probability of samples (batchsize x k) samples
local Pms = self.output[2]
Pms:resize(batchsize, self.k)
Pms:copy(sbias:view(1,self.k):expand(batchsize, self.k))
Pms:addmm(1, Pms, 1, input, sweight:t())
Pms:exp()
if self.Z[1] <= 0 then
-- approximate Z using current batch
self.Z[1] = Pms:mean()*self.weight:size(1)
print("normalization constant Z approximated to "..self.Z[1])
end
-- divide by normalization constant
Pms:div(self.Z[1])
Pmt:div(self.Z[1])
-- get noise probability (pn) for all samples
self.sampleprob = self.sampleprob or Pms.new()
self.sampleprob = self:noiseProb(self.sampleprob, self.sampleidx)
local Pnt = self.sampleprob:narrow(1,self.k+1,target:size(1))
local Pns = self.sampleprob:narrow(1,1,self.k)
Pns = Pns:resize(1, self.k):expand(batchsize, self.k)
self.output[3]:set(Pnt)
self.output[4]:set(Pns)
else
self.output = (torch.type(self.output) == 'table' and #self.output == 4) and self.output
or {input.new(), input.new(), input.new(), input.new()}
self.sampleidx = self.sampleidx or target.new()
-- the last first column will contain the target indices
self.sampleidx:resize(batchsize, self.k+1)
self.sampleidx:select(2,1):copy(target)
self._sampleidx = self._sampleidx or self.sampleidx.new()
self._sampleidx:resize(batchsize, self.k)
-- sample (batchsize x k+1) noise samples
self:noiseSample(self._sampleidx, batchsize, self.k)
self.sampleidx:narrow(2,2,self.k):copy(self._sampleidx)
-- make sure that targets are still first column of sampleidx
if not self.testedtargets then
for i=1,math.min(target:size(1),3) do
assert(self.sampleidx[{i,1}] == target[i])
end
self.testedtargets = true
end
-- build (batchsize x k+1 x inputsize) weight tensor
self._weight = self._weight or self.bias.new()
self.weight.index(self._weight, self.weight, 1, self.sampleidx:view(-1))
assert(self._weight:nElement() == batchsize*(self.k+1)*inputsize)
self._weight:resize(batchsize, self.k+1, inputsize)
-- build (batchsize x k+1) bias tensor
self._bias = self._bias or self.bias.new()
self._bias:index(self.bias, 1, self.sampleidx:view(-1))
assert(self._bias:nElement() == batchsize*(self.k+1))
self._bias:resize(batchsize, self.k+1)
-- get model probability (pm) of sample and target (batchsize x k+1) samples
self._pm = self._pm or input.new()
self._pm:resizeAs(self._bias):copy(self._bias)
self._pm:resize(batchsize, 1, self.k+1)
local _input = input:view(batchsize, 1, inputsize)
self._pm:baddbmm(1, self._pm, 1, _input, self._weight:transpose(2,3))
self._pm:resize(batchsize, self.k+1)
self._pm:exp()
if self.Z[1] <= 0 then
-- approximate Z using current batch
self.Z[1] = self._pm:mean()*self.weight:size(1)
print("normalization constant Z approximated to "..self.Z[1])
end
self._pm:div(self.Z[1]) -- divide by normalization constant
-- separate target from sample model probabilities
local Pmt = self._pm:select(2,1)
local Pms = self._pm:narrow(2,2,self.k)
self.output[1]:set(Pmt)
self.output[2]:set(Pms)
-- get noise probability (pn) for all samples
self.sampleprob = self.sampleprob or self._pm.new()
self.sampleprob = self:noiseProb(self.sampleprob, self.sampleidx)
local Pnt = self.sampleprob:select(2,1)
local Pns = self.sampleprob:narrow(2,2,self.k)
self.output[3]:set(Pnt)
self.output[4]:set(Pns)
end
return self.output
end
function NCEModule:updateGradInput(inputTable, gradOutput)
local input, target = unpack(inputTable)
assert(input:dim() == 2)
assert(target:dim() == 1)
local dPmt, dPms = gradOutput[1], gradOutput[2]
local batchsize = input:size(1)
local inputsize = self.weight:size(2)
if self.batchnoise then
local Pmt, Pms = self.output[1], self.output[2]
-- separate sample and target weight matrices
local sweight = self._weight:narrow(1, 1, self.k)
local tweight = self._weight:narrow(1, self.k+1, batchsize)
-- the rest of equation 7
-- d Pm / d linear = exp(linear)/z
self._gradOutput = self._gradOutput or dPms.new()
self._tgradOutput = self._tgradOutput or dPmt.new()
self._gradOutput:cmul(dPms, Pms)
self._tgradOutput:cmul(dPmt, Pmt)
-- gradient of linear
self.gradInput[1] = self.gradInput[1] or input.new()
self.gradInput[1]:cmul(self._tgradOutput:view(batchsize, 1):expandAs(tweight), tweight)
self.gradInput[1]:addmm(1, 1, self._gradOutput, sweight)
else
-- the rest of equation 7 (combine both sides of + sign into one tensor)
self._gradOutput = self._gradOutput or dPmt.new()
self._gradOutput:resize(batchsize, self.k+1)
self._gradOutput:select(2,1):copy(dPmt)
self._gradOutput:narrow(2,2,self.k):copy(dPms)
self._gradOutput:resize(batchsize, 1, self.k+1)
-- d Pm / d linear = exp(linear)/z
self._gradOutput:cmul(self._pm)
-- gradient of linear
self.gradInput[1] = self.gradInput[1] or input.new()
self.gradInput[1]:resize(batchsize, 1, inputsize):zero()
self.gradInput[1]:baddbmm(0, 1, self._gradOutput, self._weight)
self.gradInput[1]:resizeAs(input)
end
self.gradInput[2] = self.gradInput[2] or input.new()
if self.gradInput[2]:nElement() ~= target:nElement() then
self.gradInput[2]:resize(target:size()):zero()
end
return self.gradInput
end
function NCEModule:accGradParameters(inputTable, gradOutput, scale)
local input, target = unpack(inputTable)
assert(input:dim() == 2)
assert(target:dim() == 1)
local batchsize = input:size(1)
local inputsize = self.weight:size(2)
if self.batchnoise then
self._gradWeight = self._gradWeight or self.bias.new()
self._gradWeight:resizeAs(self._weight):zero() -- (batchsize + k) x inputsize
local sgradWeight = self._gradWeight:narrow(1, 1, self.k)
local tgradWeight = self._gradWeight:narrow(1, self.k+1, batchsize)
self._gradOutput:mul(scale)
self._tgradOutput:mul(scale)
sgradWeight:addmm(0, sgradWeight, 1, self._gradOutput:t(), input)
tgradWeight:cmul(self._tgradOutput:view(batchsize, 1):expandAs(self.gradInput[1]), input)
self.gradWeight:indexAdd(1, self.sampleidx, self._gradWeight)
self.gradBias:indexAdd(1, self.sampleidx:narrow(1,self.k+1,batchsize), self._tgradOutput)
self._tgradOutput:sum(self._gradOutput, 1) -- reuse buffer
self.gradBias:indexAdd(1, self.sampleidx:sub(1,self.k), self._tgradOutput:view(-1))
else
self._gradWeight = self._gradWeight or self.bias.new()
self._gradWeight:resizeAs(self._weight):zero() -- batchsize x k+1 x inputsize
self._gradOutput:resize(batchsize, self.k+1, 1)
self._gradOutput:mul(scale)
local _input = input:view(batchsize, 1, inputsize)
self._gradWeight:baddbmm(0, self._gradWeight, 1, self._gradOutput, _input)
local sampleidx = self.sampleidx:view(batchsize * (self.k+1))
local _gradWeight = self._gradWeight:view(batchsize * (self.k+1), inputsize)
self.gradWeight:indexAdd(1, sampleidx, _gradWeight)
local _gradOutput = self._gradOutput:view(batchsize * (self.k+1))
self.gradBias:indexAdd(1, sampleidx, _gradOutput)
end
end
function NCEModule:type(type, cache)
if type then
self.sampleidx = nil
self.sampleprob = nil
self._noiseidx = nil
self._noiseprob = nil
self._metaidx = nil
self._gradOutput = nil
self._tgradOutput = nil
self._gradWeight = nil
self._weight = nil
end
local unigrams = self.unigrams
self.unigrams = nil
local am = self.aliasmultinomial
local rtn
if type and torch.type(self.weight) == 'torch.MultiCudaTensor' then
assert(type == 'torch.CudaTensor', "Cannot convert a multicuda NCEModule to anything other than cuda")
local weight = self.weight
local gradWeight = self.gradWeight
self.weight = nil
self.gradWeight = nil
rtn = parent.type(self, type, cache)
assert(torch.type(self.aliasmultinomial.J) ~= 'torch.CudaTensor')
self.weight = weight
self.gradWeight = gradWeight
else
rtn = parent.type(self, type, cache)
end
self.unigrams = unigrams
self.aliasmultinomial = am
return rtn
end
function NCEModule:noiseProb(sampleprob, sampleidx)
assert(sampleprob)
assert(sampleidx)
self._noiseprob = self._noiseprob or self.unigrams.new()
self._noiseidx = self._noiseidx or torch.LongTensor()
self._noiseidx:resize(sampleidx:size()):copy(sampleidx)
self._noiseprob:index(self.unigrams, 1, self._noiseidx:view(-1))
sampleprob:resize(sampleidx:size()):copy(self._noiseprob)
return sampleprob
end
function NCEModule:noiseSample(sampleidx, batchsize, k)
if torch.type(sampleidx) ~= 'torch.LongTensor' then
self._noiseidx = self._noiseidx or torch.LongTensor()
self._noiseidx:resize(batchsize, k)
self.aliasmultinomial:batchdraw(self._noiseidx)
sampleidx:resize(batchsize, k):copy(self._noiseidx)
else
sampleidx:resize(batchsize, k)
self.aliasmultinomial:batchdraw(sampleidx)
end
return sampleidx
end
function NCEModule:clearState()
self.sampleidx = nil
self.sampleprob = nil
self._noiseidx = nil
self._noiseprob = nil
self._tgradOutput = nil
self._gradOutput = nil
if torch.isTensor(self.output) then
self.output:set()
else
for i,output in ipairs(self.output) do
output:set()
end
end
for i,gradInput in ipairs(self.gradInput) do
gradInput:set()
end
end
function NCEModule:multicuda(device1, device2)
assert(device1 and device2, "specify two devices as arguments")
require 'torchx'
assert(torchx.version and torchx.version >= 1, "update torchx: luarocks install torchx")
self:float()
local isize = self.weight:size(2)
local weights = {
cutorch.withDevice(device1, function() return self.weight[{{}, {1, torch.round(isize/2)}}]:cuda() end),
cutorch.withDevice(device2, function() return self.weight[{{}, {torch.round(isize/2)+1, isize}}]:cuda() end)
}
self.weight = torch.MultiCudaTensor(2, weights)
local gradWeights = {
cutorch.withDevice(device1, function() return self.gradWeight[{{}, {1, torch.round(isize/2)}}]:cuda() end),
cutorch.withDevice(device2, function() return self.gradWeight[{{}, {torch.round(isize/2)+1, isize}}]:cuda() end)
}
self.gradWeight = torch.MultiCudaTensor(2, gradWeights)
self:cuda()
end