-
Notifications
You must be signed in to change notification settings - Fork 3
/
GBN_Online_CommonMode_CongPerp.m
446 lines (412 loc) · 20.6 KB
/
GBN_Online_CommonMode_CongPerp.m
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
440
441
442
443
444
445
446
function [ParaGlobal,PerpCun,AccCun,ParaLocal] = GBN_Online_CommonMode_CongPerp(X_all,prepar,K,T,trial,...
DataType,dataname,SuPara,Settings)
% Yulai Cong
% 2017 03 18
%% Settings
eta = SuPara.eta ;
ac = SuPara.ac ; bc = SuPara.bc ;
a0pj = SuPara.a0pj ; b0pj = SuPara.b0pj ;
e0cj = SuPara.e0cj ; f0cj = SuPara.f0cj ;
e0c0 = SuPara.e0c0 ; f0c0 = SuPara.f0c0 ;
a0gamma = SuPara.a0gamma ; b0gamma = SuPara.b0gamma ;
IsDisplay = Settings.IsDisplay ;
FigureGap = Settings.FigureGap ;
IsTestPointPerp = Settings.IsTestPointPerp ;
PointTestLocation = Settings.PointTestLocation ;
IsTestPerpZhou = Settings.IsTestPerpZhou ;
PerpZhouBurnin = Settings.PerpZhouBurnin ;
PerpZhouStep = Settings.PerpZhouStep ;
IsTestAcc = Settings.IsTestAcc ;
if ~isfield(prepar, 'Y')
Settings.IsTestAcc = 0 ;
IsTestAcc = Settings.IsTestAcc ;
end
IsSaveData = Settings.IsSaveData ;
SaveDataLocation = Settings.SaveDataLocation ;
IterAll = Settings.IterAll ;
MiniBatchSZ = Settings.MiniBatchSZ ;
BurninMB = Settings.BurninMB ;
CollectionMB = Settings.CollectionMB ;
AlgoChosen = Settings.AlgoChosen ;
FurCollapse = Settings.FurCollapse ;
epsi0FR = Settings.epsi0FR ;
tao0FR = Settings.tao0FR ; kappa0FR = Settings.kappa0FR ;
epsipiet = (tao0FR + (1:IterAll)) .^ (-kappa0FR) ;
epsipiet = epsi0FR * epsipiet / epsipiet(1) ;
epsi0 = Settings.epsi0 ;
tao0 = Settings.tao0 ; kappa0 = Settings.kappa0 ;
epsit = (tao0 + (1:IterAll)) .^ (-kappa0) ;
epsit = epsi0 * epsit / epsit(1) ;
etaStr = num2str(eta(1)) ; etaStr(etaStr=='.') = [] ;
MBSizeBurnCol = [num2str(MiniBatchSZ),'_',num2str(BurninMB),'_',num2str(CollectionMB)] ;
StepSet = [num2str(epsi0),'_',num2str(tao0),'_',num2str(kappa0),'_',num2str(epsi0FR),'_',num2str(tao0FR),'_',num2str(kappa0FR)] ;
StepSet(StepSet=='.') = [] ;
%% Initialization
if ~strcmp(dataname,'Wiki')
XOnline = X_all(:,prepar.trindx);
[V,NOnline] = size(XOnline);
else
OriFromChangyouChen = Settings.OriFromChangyouChen ;
if OriFromChangyouChen
fid_train = fopen('wiki_train.txt', 'r');
V = 7702 ;
NOnline = 1000000 ;
else
XOnline = X_all;
[V,NOnline] = size(XOnline);
end
end
%%============================ Global parameter initialization ============================
Phi = cell(T,1); Eta = cell(T,1);
for t = 1:T
Eta{t} = eta(t) ;
if t == 1
Phi{t} = 0.05 + 0.95 * rand(V,K(1));
else
Phi{t} = 0.05 + 0.95 * rand(K(t-1),K(t));
end
Phi{t} = bsxfun(@rdivide, Phi{t}, max(realmin,sum(Phi{t},1)));
end
r_k = 1/K(T)*ones(K(T),1); gamma0 = 1 ; c0 = 1 ;
PhiBar = cell(T,1); PhiHat = cell(T,1);
for t = 1:T
PhiBar{t} = Phi{t}(1:end-1,:) ;
PhiHat{t} = Phi{t} ;
end
%%============================ Local parameter initialization ============================
Theta = cell(T,1); c_j = cell(T+1,1);
Xt_to_t1 = cell(T,1); WSZS = cell(T,1);
%%============================ Register initialization ============================
Ml = cell(T+2,1); % Global
EWSZS = cell(T,1); ESumLnPC = cell(T,1); % Local
EXr = [] ; ESumLnPT1 = [] ;
%%============================ Others ============================
PerpCun = [] ; AccCun = [] ;
TimeAll = 0 ;
RiemannCoefCun = cell(1,T+2) ;
PhiUsageCun = cell(1,T) ;
for t = 1:T
PhiUsageCun{t} = 0 ;
end
%%============================ PerpZhou ============================
if IsTestPerpZhou & strcmp(DataType, 'Count')
NTest = size(prepar.XHOtest,2) ;
FlagHOTrain = prepar.XHOtrain > 0 ;
FlagHOTest = prepar.XHOtest > 0 ;
Xt_to_t1Test = cell(T,1); ThetaTest = cell(T,1); c_jTest = cell(T+1,1);
PhiThetaTest = 0 ; CountNum = 0 ;
for t=T:-1:1
ThetaTest{t} = ones(K(t),NTest)/K(t);
end
for t=1:(T+1)
c_jTest{t}=ones(1,NTest);
end
p_jTest = Calculate_pj(c_jTest,T);
end
%% Online Iterations
for epochi = 1:IterAll
if epochi == 1
idxall = 1 : NOnline ;
else
idxall = randperm(NOnline) ;
end
MBratio = NOnline / MiniBatchSZ ;
N = MiniBatchSZ ;
for MBt = 1:floor(MBratio)
MBObserved = (epochi-1)*floor(MBratio)+MBt ;
tic
%% Get minibatch
if ~strcmp(dataname,'Wiki')
X = XOnline(:,idxall( (MBt-1)*N + (1:N) )) ;
else
if OriFromChangyouChen
X = GetBatch(fid_train, N, 'wiki_train.txt') ;
else
X = XOnline(:,idxall( (MBt-1)*N + (1:N) )) ;
end
end
%% Initialize ParaLocal
if strcmp(DataType, 'Positive') | strcmp(DataType, 'Binary')
[ii,jj] = find(X>eps); %Treat values smaller than eps as 0
iijj = find(X>eps);
end
if strcmp(DataType, 'Positive')
a_j = ones(1,N);
end
for t=1:(T+1)
c_j{t}=ones(1,N);
end
p_j = Calculate_pj(c_j,T);
for t=T:-1:1
Theta{t} = ones(K(t),N) / K(t);
EWSZS{t} = 0 ;
ESumLnPC{t} = 0 ;
end
EXr = 0 ; ESumLnPT1 = 0 ;
%% Sample and Collect local information
for iter = 1 : (BurninMB + CollectionMB)
%%============= Upward Pass ===================
for t = 1:T
if t == 1
switch DataType
case 'Positive'
Rate = Phi{1}*Theta{1};
Rate = 2*sqrt(a_j(jj)'.*X(iijj).*Rate(iijj));
M = Truncated_bessel_rnd( Rate );
a_j = randg(full(sparse(1,jj,M,1,N))+ac) ./ (bc+sum(X,1));
Xt = sparse(ii,jj,M,V,N);
case 'Binary'
Rate = Phi{1}*Theta{1};
M = truncated_Poisson_rnd(Rate(iijj));
Xt = sparse(ii,jj,M,V,N);
case 'Count'
Xt = sparse(X);
end
[Xt_to_t1{t},WSZS{t}] = Multrnd_Matrix_mex_fast(Xt,Phi{t},Theta{t});
else
[Xt_to_t1{t},WSZS{t}] = CRT_Multrnd_Matrix(sparse(Xt_to_t1{t-1}),Phi{t},Theta{t});
end
end
Xr = CRT_sum_mex_matrix_v1(sparse(Xt_to_t1{T}'),r_k')';
%%============= Downward Pass ======================
if iter >= min(BurninMB/2,10)
if T > 1
p_j{2} = betarnd( sum(Xt_to_t1{1},1)+a0pj , sum(Theta{2},1)+b0pj );
else
p_j{2} = betarnd( sum(Xt_to_t1{1},1)+a0pj , sum(r_k,1)+b0pj );
end
p_j{2} = min( max(p_j{2},realmin) , 1-realmin);
c_j{2} = (1-p_j{2})./p_j{2};
for t = 3:(T+1)
if t == T+1
c_j{t} = randg(sum(r_k)*ones(1,N)+e0cj) ./ (sum(Theta{t-1},1)+f0cj);
else
c_j{t} = randg(sum(Theta{t},1)+e0cj) ./ (sum(Theta{t-1},1)+f0cj);
end
end
p_j_temp = Calculate_pj(c_j,T);
p_j(3:end)=p_j_temp(3:end);
end
for t = T:-1:1
if t == T
shape = r_k;
else
shape = Phi{t+1}*Theta{t+1};
end
Theta{t} = bsxfun(@times, randg(bsxfun(@plus,shape,Xt_to_t1{t})), 1 ./ max(realmin, c_j{t+1}-logmax(1-p_j{t})) );
if nnz(isnan(Theta{t})) | nnz(isinf(Theta{t})) | nnz(sum(Theta{t})==0)
warning(['Theta Nan',num2str(nnz(isnan(Theta{t}))),'_Inf',num2str(nnz(isinf(Theta{t}))),'_ORsum=0]']);
Theta{t}(isnan(Theta{t})) = 0 ;
end
end
%%============= Collection ======================
if iter > BurninMB
for t = T:-1:1
if ~FurCollapse
EWSZS{t} = EWSZS{t} + WSZS{t} ;
else
PhiTheta = Phi{t} * Theta{t} ;
PhiTheta(PhiTheta<=eps) = eps ;
if t == 1
EWSZS{t} = EWSZS{t} + (X./PhiTheta)*(Theta{t}.') ;
else
EWSZS{t} = EWSZS{t} + (psi(Xt_to_t1{t-1}+PhiTheta)-psi(PhiTheta))*(Theta{t}.') ;
end
end
if nnz(isnan(EWSZS{t})) | nnz(isinf(EWSZS{t}))
warning(['EWSZS Nan',num2str(nnz(isnan(EWSZS{t}))),'_Inf',num2str(nnz(isinf(EWSZS{t})))]);
EWSZS{t}(isnan(EWSZS{t})) = 0 ;
end
end
EXr = EXr + Xr ;
ESumLnPT1 = ESumLnPT1 + sum(logmax(1-p_j{T+1})) ;
end
end
%% Update global parameters Phi
%%================= Update Mk ===============
for t = 1:T
PhiUsageCun{t} = PhiUsageCun{t} + sum(EWSZS{t},1) ;
if ~FurCollapse
EWSZS{t} = MBratio * EWSZS{t}/CollectionMB ;
else
EWSZS{t} = MBratio * (Phi{t} .* EWSZS{t}/CollectionMB) ;
end
switch AlgoChosen
case 1 % SGRLD
Ml{t} = 1 ;
RiemannCoefCun{t}(:,MBObserved) = Ml{t}(:) ;
case 3 % TLFSGR
if t == 1
if (MBObserved == 1)
Ml{1} = mean( sum(EWSZS{1},1) ) ;
else
Ml{1} = (1 - epsipiet(MBObserved)) * Ml{1} ...
+ epsipiet(MBObserved) * mean( sum(EWSZS{1},1) ) ;
end
else
Ml{t} = mean(Ml{1}) ;
end
RiemannCoefCun{t}(:,MBObserved) = Ml{t}(:) ;
case 4 % TLASGR
if (MBObserved == 1)
Ml{t} = sum(EWSZS{t},1) ;
else
Ml{t} = (1 - epsipiet(MBObserved)) * Ml{t} ...
+ epsipiet(MBObserved) * sum(EWSZS{t},1) ;
end
RiemannCoefCun{t}(:,MBObserved) = Ml{t}(:) ;
otherwise
error('Wrong AlgoChosen!!!') ;
end
%%================= Update Global parameters Phi ===============
for iii = 1:1
switch AlgoChosen
case 1 % SGRLD
tmp = (EWSZS{t}+Eta{t}) - bsxfun(@times, sum(EWSZS{t}+PhiHat{t},1), Phi{t}) ;
tmp1 = 2 * PhiHat{t} ;
PhiHat{t} = abs( PhiHat{t} + epsit(MBObserved)*tmp ...
+ sqrt(epsit(MBObserved)*tmp1) .* randn(size(PhiHat{t})) ) ;
Phi{t} = bsxfun(@rdivide, PhiHat{t}, sum(PhiHat{t},1) ) ; % figure(25),DispDictionaryImagesc(Phi{1});drawnow;
case 3 % TLFSGR
tmp = (EWSZS{t}+Eta{t}) ;
tmp = bsxfun(@times, 1./Ml{t}, tmp - bsxfun(@times, sum(tmp,1), Phi{t})) ;
tmp1 = bsxfun(@times, 2./Ml{t}, Phi{t}) ;
tmp = Phi{t} + epsit(MBObserved)*tmp + sqrt(epsit(MBObserved)*tmp1) .* randn(size(Phi{t})) ;
Phi{t} = ProjSimplexSpecial(tmp , Phi{t}, eps) ; % figure(25),DispDictionaryImagesc(Phi{1});drawnow;
case 4 % TLASGR
tmp = (EWSZS{t}+Eta{t}) ;
tmp = bsxfun(@times, 1./Ml{t}, tmp - bsxfun(@times, sum(tmp,1), Phi{t})) ;
tmp1 = bsxfun(@times, 2./Ml{t}, Phi{t}) ;
tmp = Phi{t} + epsit(MBObserved)*tmp + sqrt(epsit(MBObserved)*tmp1) .* randn(size(Phi{t})) ;
Phi{t} = ProjSimplexSpecial(tmp , Phi{t},eps) ; % figure(25),DispDictionaryImagesc(Phi{1});drawnow;
otherwise
error('Wrong AlgoChosen!!!') ;
end
end
if nnz(imag(Phi{t})~=0) | nnz(isnan(Phi{t})) | nnz(isinf(Phi{t}))
warning(['Phi Nan',num2str(nnz(isnan(Phi{t}))),'_Inf',num2str(nnz(isinf(Phi{t})))]);
Phi{t}(isnan(Phi{t})) = 0 ;
end
end
%% Update Global parameters r gamma0 c0
EXr = MBratio * EXr/CollectionMB ;
ESumLnPT1 = MBratio * ESumLnPT1/CollectionMB ;
if (MBObserved == 1)
Ml{T+1} = -ESumLnPT1 ;
else
Ml{T+1} = (1 - epsipiet(MBObserved)) * Ml{T+1} ...
+ epsipiet(MBObserved) * (-ESumLnPT1) ;
end
RiemannCoefCun{T+1}(:,MBObserved) = Ml{T+1}(:) ;
tmp = 1/Ml{T+1} * ( (EXr+gamma0/K(T)) - r_k*(c0 - ESumLnPT1) ) ;
tmp1 = 2/Ml{T+1} * r_k ;
r_k = abs( r_k + epsit(MBObserved)*tmp + sqrt(epsit(MBObserved)*tmp1) .* randn(size(r_k)) ) ;
%% ==================== Figure ====================
TimeOneIter = toc ;
TimeAll = TimeAll + TimeOneIter ;
if (mod(MBObserved,5) == 0) | (MBObserved < 5)
fprintf('Layer: %d, Sweepi: %d, MBObserved: %d, Time: %d, r_k: %d, gamma0: %d, c0: %d\n',T,epochi,MBObserved,TimeOneIter,mean(r_k),gamma0,c0) ;
end
if IsDisplay && (mod(MBObserved,FigureGap)==0)
tmp = 1 ;
for t = 1:T
tmp = tmp * Phi{t} ;
if strcmp(DataType,'Binary')
figure(90+t),DispDictionaryImshow(1-exp(-tmp));drawnow;
else
figure(90+t),DispDictionaryImshow(tmp);drawnow;
end
end
end
%% ==================== Quantify Heldout Perplexity Zhou ====================
if IsTestPerpZhou & strcmp(DataType, 'Count') & isfield(prepar,'XHOtrain') & isfield(prepar,'XHOtest') %& (MBObserved <= (PerpZhouBurnin + 3000))
for t = 1:T
if t == 1
Xt_to_t1Test{t} = Multrnd_Matrix_mex_fast_v1(prepar.XHOtrain,Phi{t},ThetaTest{t});
else
Xt_to_t1Test{t} = CRT_Multrnd_Matrix(sparse(Xt_to_t1Test{t-1}),Phi{t},ThetaTest{t});
end
end
if T > 1
p_jTest{2} = betarnd( sum(Xt_to_t1Test{1},1)+a0pj , sum(ThetaTest{2},1)+b0pj );
else
p_jTest{2} = betarnd( sum(Xt_to_t1Test{1},1)+a0pj , sum(r_k,1)+b0pj );
end
p_jTest{2} = min( max(p_jTest{2},realmin) , 1-realmin); c_jTest{2} = (1-p_jTest{2})./p_jTest{2};
for t = 3:(T+1)
if t == T+1
c_jTest{t} = randg(sum(r_k)*ones(1,NTest)+e0cj) ./ (sum(ThetaTest{t-1},1)+f0cj);
else
c_jTest{t} = randg(sum(ThetaTest{t},1)+e0cj) ./ (sum(ThetaTest{t-1},1)+f0cj);
end
end
p_j_temp = Calculate_pj(c_jTest,T); p_jTest(3:end)=p_j_temp(3:end);
for t = T:-1:1
if t == T
shape = r_k;
else
shape = Phi{t+1}*ThetaTest{t+1};
end
ThetaTest{t} = bsxfun(@times, randg(bsxfun(@plus,shape,Xt_to_t1Test{t})), 1 ./ max(realmin, c_jTest{t+1}-log(max(1-p_jTest{t},realmin)) ) );
if nnz(isnan(ThetaTest{t})) | nnz(isinf(ThetaTest{t})) | nnz(sum(ThetaTest{t})==0)
warning(['PerpZhou: Theta Nan',num2str(nnz(isnan(ThetaTest{t}))),'_Inf',num2str(nnz(isinf(ThetaTest{t}))),'_ORsum=0]']);
ThetaTest{t}(isnan(ThetaTest{t})) = 0 ;
end
end
tmp = Phi{1}*ThetaTest{1};
if nnz(isnan(tmp)) | nnz(isinf(tmp))
tmp = Phi{1}*ThetaTest{1};
end
temp = bsxfun(@rdivide, tmp, max(realmin,sum(tmp,1)) );
PerpCun.TmpPerpZhouTest(MBObserved) = ...
exp(- sum(prepar.XHOtest(FlagHOTest).*log(max(realmin,temp(FlagHOTest)))) / sum(prepar.XHOtest(:)) );
if (MBObserved > PerpZhouBurnin && mod(MBObserved,PerpZhouStep) == 0)
PhiThetaTest = PhiThetaTest + tmp;
CountNum = CountNum + 1 ;
temp = PhiThetaTest / CountNum ;
temp = bsxfun(@rdivide, temp, max(realmin,sum(temp,1)) );
PerpCun.PerpZhouTrain(CountNum) = exp(- sum(prepar.XHOtrain(FlagHOTrain).*log(max(realmin,temp(FlagHOTrain)))) / sum(prepar.XHOtrain(:)) );
PerpCun.PerpZhouTest(CountNum) = exp(- sum(prepar.XHOtest(FlagHOTest).*log(max(realmin,temp(FlagHOTest)))) / sum(prepar.XHOtest(:)) );
PerpCun.PerpZhouMBSeen(CountNum) = MBObserved ;
PerpCun.PerpZhouDataSeen(CountNum) = MBObserved * MiniBatchSZ ;
PerpCun.PerpZhouTimeUsed(CountNum) = TimeAll ;
if (mod(MBObserved,5) == 0)
fprintf('PerpZhouVal: MBObserved: %d, HeldoutTrain: %d, HeldoutTest: %d, DataSeen: %d, TimeUsed: %d\n',...
MBObserved,full(PerpCun.PerpZhouTrain(end)),full(PerpCun.PerpZhouTest(end)),full(PerpCun.PerpZhouDataSeen(end)),full(PerpCun.PerpZhouTimeUsed(end))) ;
end
end
end
%% ==================== Quantification Point Perp ====================
if IsTestPointPerp
if strcmp(DataType, 'Positive')
ParaGlobal.a_jmean = median(a_j) ;
end
ParaGlobal.Phi = Phi ; ParaGlobal.Eta = Eta ;
ParaGlobal.r_k = r_k ; ParaGlobal.gamma0 = gamma0 ; ParaGlobal.c0 = c0 ;
ParaGlobal.c_jmean = zeros(1,T+1);
for t = 1:(T+1)
ParaGlobal.c_jmean(t) = median(c_j{t}) ;
end
if IsTestPointPerp & strcmp(DataType, 'Count') & isfield(prepar,'XHOtrain') & isfield(prepar,'XHOtest')
tmp = find(PointTestLocation == MBObserved) ;
if tmp
PerpCun.PointPerpMBSeen(tmp) = MBObserved ;
PerpCun.PointPerpDataSeen(tmp) = MBObserved * MiniBatchSZ ;
PerpCun.PointPerpTimeUsed(tmp) = TimeAll ;
[PerpCun.PointPerpVal(tmp),PerpCun.PointPerpValTrain(tmp)] = ...
QuantGBN_PerpWord(prepar.XHOtrain,prepar.XHOtest,ParaGlobal,K,T,DataType,SuPara,Settings) ;
end
end
end
%% others
if (MBObserved >= IterAll)
break ;
end
end
if (MBObserved >= IterAll)
break ;
end
end
if strcmp(dataname,'Wiki') & OriFromChangyouChen
fclose(fid_train) ;
end