forked from BitSeq/BitSeq
-
Notifications
You must be signed in to change notification settings - Fork 0
/
estimateHyperPar.cpp
369 lines (357 loc) · 15.2 KB
/
estimateHyperPar.cpp
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
/*
* Hyperparameter model in estimate[*]HyperPar.cpp always depends on the model used in
* relevant estimate[*]DE.cpp
*/
// DECLARATIONS: {{{
#include <algorithm>
#include <cmath>
#include <fstream>
#include <sstream>
#include "boost/random/mersenne_twister.hpp"
#include "boost/random/normal_distribution.hpp"
#include "boost/random/uniform_01.hpp"
using namespace std;
#include "ArgumentParser.h"
#include "lowess.h"
#include "MyTimer.h"
#include "misc.h"
#include "PosteriorSamples.h"
#include "TranscriptExpression.h"
#include "common.h"
using ns_params::paramT;
//}}}
// Defaults: {{{
#define ALPHA_PROP 0.1
#define BETA_PROP 0.08
#define subM_MIN 10
#define subM_MAX 5000
#define SAMPLES_N 2
#define MAX_ITER 1000
#define MAX_RETRIES 10
#define MAX_PARAM 5000
//}}}
extern "C" int estimateHyperPar(int *argc,char* argv[]){
string programDescription =
"Estimate expression dependent hyperparameters from the dataset.\n\
[sample Files] should contain transposed MCMC samples from replicates.\n\
To distinguish conditions use C between them e.g.:\n\
samplesC1-R1.rpkm samplesC1-R2.rpkm C samplesC2-R1.rpkm samplesC2-R2.rpkm";
// Intro: {{{
// Set options {{{
ArgumentParser args(programDescription,"[sampleFiles]",1);
args.addOptionB("V","veryVerbose","veryVerbose",0,"More verbose output.");
args.addOptionS("o","outFile","outFileName",1,"Name of the output file.");
args.addOptionS("p","paramsAllFile","paramsAllFileName",0,"Name of the file to which to store all parameter values generated prior to lowess smoothing.");
args.addOptionS("","meanFile","meanFileName",0,"Name of the file containing joint mean and variance.");
args.addOptionL("g","groupsNumber","groupsN",0,"Number of groups of transcript of similar size.",200);
args.addOptionL("s","samplesNumber","samplesN",0,"Number of samples generated for each group.",SAMPLES_N);
args.addOptionD("l","lambda0","lambda0",0,"Precision scaling parameter lambda0.",2.0);
args.addOptionD("","exThreshold","exT",0,"Threshold of lowest expression for which the estimation is done.",-5);
args.addOptionB("S","smoothOnly","smoothOnly",0,"Input file contains previously sampled hyperparameters which should smoothed only.");
args.addOptionD("","lowess-f","lowess-f",0,"Parameter F for lowess smoothing specifying amount of smoothing.",0.2);
args.addOptionL("","lowess-steps","lowess-steps",0,"Parameter Nsteps for lowess smoothing specifying number of iterations.",5);
args.addOptionB("","noforce","noforce",0,"Do not force smoothing of the parameters.",false);
args.addOptionS("","norm","normalization",0,"Normalization constants for each input file provided as comma separated list of doubles (e.g. 1.0017,1.0,0.9999 ).");
args.addOptionL("","seed","seed",0,"Random initialization seed.");
if(!args.parse(*argc,argv))return 0;
if(args.verbose)buildTime(argv[0],__DATE__,__TIME__);
// }}}
MyTimer timer;
timer.start(1);
long i,M=0,N,RTN,C;
bool storeAll=args.isSet("paramsAllFileName");
vector<paramT> params;
paramT param;
TranscriptExpression trExp;
ofstream outF;
if(! args.flag("smoothOnly")){
if(! args.isSet("meanFileName")){
error("Main: Please provide mean file name (--meanFile).\n");
return 1;
}
trExp.readExpression(args.getS("meanFileName"), MEAN_VARIANCE);
// Force user to use logged mean and samples.
if(!trExp.isLogged()){
error("Main: Please compute the expression mean from logged samples (getVariance --log ...).\n");
return 1;
}
M = trExp.getM();
if(args.verbose)message("Transcripts in expression file: %ld\n",M);
trExp.doSort(true);
}
if(!ns_misc::openOutput(args, &outF)) return 1;
///}}}
if(args.flag("smoothOnly")){
// Reading previously sampled parameters. (header is copie into outF)
readParams(args.args()[0], ¶ms, &outF);
}else{
// Sampling parameters based on data
// Read conditions {{{
Conditions cond;
if(!ns_misc::readConditions(args, &C, &M, &N, &cond)) return 1;
RTN = cond.getRN();
if(args.verbose)message("Number of all replicates: %ld\n",RTN);
// Prepare file for storing all sampled parameters.
ofstream paramsF;
if(storeAll){
if(!ns_misc::openOutput(args.getS("paramsAllFileName"), ¶msF)) return 1;
paramsF<<"# lambda0 "<<args.getD("lambda0")<<endl;
}
// }}}
// Declarations {{{
vector<long double> mu0(subM_MAX,0);
vector<vector<vector<double> > > tr(subM_MAX,vector<vector<double> >(RTN));
vector<vector<long double> > bAdd(subM_MAX,vector<long double> (C,0));
boost::random::mt11213b rng_mt(ns_misc::getSeed(args));
boost::random::uniform_01<long double> uniformDistribution;
boost::random::normal_distribution<long double> normalDistributionA,normalDistributionB;
typedef boost::random::normal_distribution<long double>::param_type nDP;
long double alpha,beta,alphaP,betaP,prob,probAll,probC,mean,old_mult,proposalMultiplier,acceptR,sum,sumS,lambda0,exDelta,exLast;
long samp,samplesN,samplesREDO,maxIter,r,c,m,curM,Rc,subM;
bool breaked=false,good=false;
//}}}
// Initial values {{{
alpha=uniformDistribution(rng_mt)*10.0;
beta=uniformDistribution(rng_mt)*5.0;
old_mult=0;
proposalMultiplier=2.0;
prob = 0;
lambda0 = args.getD("lambda0");
samplesN = args.getL("samplesN");
curM=0;
exDelta = (trExp.exp(0)-trExp.exp(M-1))/args.getL("groupsN");
exLast = trExp.exp(0);
if(args.verbose)message("Expression step: %Lg\n",exDelta);
// }}}
timer.split();
if(args.verbose)message("Running sampler.\n");
while(curM<M){
// Reading next group of transcripts {{{
mean=0;
m = 0;
while((curM<M)&&(m<subM_MAX)){
if(trExp.exp(curM)<args.getD("exT")){
if(args.verbose)message("skipping expression: %lg\n",trExp.exp(curM));
break;
}
for(r=0;r<RTN;r++){
good = cond.getTranscript(r, trExp.id(curM), tr[m][r],samplesN+MAX_RETRIES);
if(!good)break;
// If sampels were not logged, log them now.
if(!cond.logged())
for(samp=0;samp<samplesN+MAX_RETRIES;samp++){
tr[m][r][samp] = (tr[m][r][samp] == 0)? ns_misc::LOG_ZERO:log(tr[m][r][samp]);
}
}
if(good){
mu0[m]=trExp.exp(curM);
mean+=mu0[m];
m++;
}
curM++;
if(args.flag("veryVerbose"))if(progressLog(curM,M,10,' '))timer.split(0,'m');
if((m>=subM_MIN)&&(exDelta<exLast-trExp.exp(curM-1)))break;
}
exLast = trExp.exp(curM-1);
if(m<subM_MIN)break;
subM = m;
mean/=subM;
if(args.flag("veryVerbose"))message("# mean: %Lg subM: %ld\n",mean,subM);
if(storeAll)paramsF<<"# mean: "<<mean<<" subM: "<<subM<<endl;
samplesREDO = 0;
//}}}
for(samp=0;samp<samplesN+samplesREDO;samp++){
// Computing Badd_gc and initializing {{{
for(m=0;m<subM;m++){
i=0; // counter over all replicates;
for(c=0;c<C;c++){
sum = 0;
sumS = 0;
Rc=cond.getRC(c);
for(r=0;r<Rc;r++){
sum += tr[m][i][samp];
sumS += tr[m][i][samp]*tr[m][i][samp];
i++;
}
bAdd[m][c]=0.5*(sumS + mu0[m]*mu0[m]*lambda0 -
(sum+mu0[m]*lambda0)*(sum+mu0[m]*lambda0)/(lambda0+Rc));
}
}
acceptR=0;
old_mult=0;
proposalMultiplier=proposalMultiplier*2.0;
normalDistributionA.param(nDP(0,ALPHA_PROP*proposalMultiplier));
normalDistributionB.param(nDP(0,BETA_PROP*proposalMultiplier));
maxIter=0;
breaked = false;
R_INTERUPT;
//}}}
while((acceptR<0.25)||(acceptR>0.5)||(old_mult!=proposalMultiplier)){
// Convergence control based on acceptance ratio. {{{
maxIter++;
if(maxIter>MAX_ITER){
if(args.flag("veryVerbose"))
message("(BREAKED acceptR %Lg mult %Lg)\n",acceptR,proposalMultiplier);
if(storeAll)
paramsF<<"#(BREAKED acceptR "<<acceptR<<" mult "<<proposalMultiplier<<")"<<endl;
breaked=true;
break;
}
if((alpha>MAX_PARAM)||(beta>MAX_PARAM)){
if(args.flag("veryVerbose"))
message("(OVERFLOW acceptR %Lg mult %Lg)\n",acceptR,proposalMultiplier);
if(storeAll)
paramsF<<"#(OVERFLOW acceptR "<<acceptR<<" mult "<<proposalMultiplier<<")"<<endl;
breaked=true;
break;
}
old_mult=proposalMultiplier;
if(acceptR<0.25)proposalMultiplier/=1.02;
if(acceptR>0.5)proposalMultiplier*=1.02;
if(old_mult!=proposalMultiplier){
normalDistributionA.param(nDP(0,ALPHA_PROP*proposalMultiplier));
normalDistributionB.param(nDP(0,BETA_PROP*proposalMultiplier));
}
//}}}
acceptR=0;
R_INTERUPT;
for(i=0;i<1000;i++){ // Sampling 1000 samples {{{
alphaP = alpha + normalDistributionA(rng_mt);
if(alphaP<0)alphaP = -alphaP;
betaP= beta + normalDistributionB(rng_mt);
if(betaP<0)betaP = -betaP;
if((alphaP==0)||(betaP==0)){
prob=0;
}else{
prob = 1.0;
probAll = pow(betaP,alphaP) / pow(beta,alpha);
for(c=0;c<C;c++){
probC = lgamma(alphaP + cond.getRC(c)/2.0)+
lgamma(alpha) -
lgamma(alpha + cond.getRC(c)/2.0) -
lgamma(alphaP);
probC = probAll * exp(probC);
for(m=0;m<subM;m++){
// message(" (var_g %lg) (pow %lg %lg %lg) ",bAdd[g]/2.0,pow(beta+bAdd[g]/2, alpha),pow(betaP+bAdd[g]/2, alphaP),pow((beta+bAdd[g]/2)/(betaP+bAdd[g]/2),SUB_N/2));
prob *= probC;
prob *= pow(beta+bAdd[m][c], alpha) /
pow(betaP+bAdd[m][c], alphaP);
prob *= pow( (beta+bAdd[m][c])/(betaP+bAdd[m][c]), (long double)(cond.getRC(c)/2.0));
}
}
if((prob>1.0)||(uniformDistribution(rng_mt)< prob)){
alpha=alphaP;
beta=betaP;
acceptR++;
}
}
} //}}}
acceptR/=i;
}
// Save generated parameters {{{
if(storeAll)
paramsF<<"#(acceptR "<<acceptR<<" mult "<<proposalMultiplier<<" iter "<<maxIter<<")"<<endl;
if(!breaked){
if(args.flag("veryVerbose")) message("%Lg %Lg\n",alpha,beta);
if(storeAll) paramsF<<alpha<<" "<<beta<<" "<<mean<<endl;
param.expr=mean;
param.alpha=alpha;
param.beta=beta;
params.push_back(param);
}else{
if(args.flag("veryVerbose")) message("# %Lg %Lg %Lg\n",alpha,beta,mean);
if(storeAll) paramsF<<"# "<<alpha<<" "<<beta<<endl;
proposalMultiplier=2;
normalDistributionA.param(nDP(0,ALPHA_PROP*proposalMultiplier));
normalDistributionB.param(nDP(0,BETA_PROP*proposalMultiplier));
alpha=uniformDistribution(rng_mt)*10.0;
beta=uniformDistribution(rng_mt)*5.0;
if(samplesREDO<MAX_RETRIES){
samplesREDO++;
}
}
//}}}
}
if((args.verbose)&&(!args.flag("veryVerbose"))){
messageF(".");
}
}
cond.close();
if(storeAll)paramsF.close();
outF<<"# lambda0 "<<args.getD("lambda0")<<endl;
if(args.verbose)message("\nSampling done.\n");
}
sort(params.begin(),params.end());
long pAll=(long)params.size(), pDistinct;
if(args.verbose)message("Have %ld parameters to smooth.\n",pAll);
vector<double> exp(pAll),alp(pAll),bet(pAll),alpS,betS;
for(i=0;i<pAll;i++){
exp[i]=params[i].expr;
alp[i]=params[i].alpha;
bet[i]=params[i].beta;
}
double f = args.getD("lowess-f");
long iter = args.getL("lowess-steps"),iterAdd;
bool redoSmooth;
for(iterAdd=0;iterAdd<6;iterAdd++){ // Increase iteration if anything is <=0
redoSmooth = false;
lowess(exp,alp,f,iter+iterAdd,alpS);
for(i=0;i<pAll;i++)
if(alpS[i]<=0){
redoSmooth = true;
if(args.flag("veryVerbose"))message(" negative alpha: %lg exp: %lg\n",alpS[i],exp[i]);
}
if(!redoSmooth)break;
if(args.verbose)message("Re-Smoothing alpha.\n");
}
outF<<"# alphaSmooth f: "<<f<<" nSteps: "<<iter+iterAdd<<endl;
if(args.verbose)message("# alphaSmooth f: %lg nSteps: %ld\n",f,iter+iterAdd);
if((iterAdd==6)&&(args.flag("noforce"))){
error("Main: Unable to produce smooth alpha >0.\nTry adjusting the parameter lowess-f.\n");
outF.close();
remove(args.getS("outFileName").c_str());
return 0;
}
for(iterAdd=0;iterAdd<6;iterAdd++){ // Increase iteration if anything is <=0
redoSmooth = false;
lowess(exp,bet,f,iter+iterAdd,betS);
for(i=0;i<pAll;i++)
if(betS[i]<=0){
redoSmooth = true;
if(args.flag("veryVerbose"))message(" negative beta: %lg exp: %lg\n",betS[i],exp[i]);
}
if(!redoSmooth)break;
if(args.verbose)message("Re-Smoothing beta.\n");
}
outF<<"# betaSmooth f: "<<f<<" nSteps: "<<iter+iterAdd<<endl;
if(args.verbose)message("# betaSmooth f: %lg nSteps: %ld\n",f,iter+iterAdd);
if((iterAdd==6)&&(args.flag("noforce"))){
error("Main: Unable to produce smooth beta >0.\nTry adjusting the parameter lowess-f.\n");
outF.close();
remove(args.getS("outFileName").c_str());
return 0;
}
if(!args.flag("noforce")){
for(i=0;i<pAll;i++)
while((i<pAll)&&((alpS[i]<=0)||(betS[i]<=0))){
message("Removing: %lg %lg %lg\n",alpS[i],betS[i],exp[i]);
alpS.erase(alpS.begin()+i); betS.erase(betS.begin()+i); exp.erase(exp.begin()+i);
pAll = alpS.size();
}
}
pDistinct = 1;
for(i=1;i<pAll;i++)if(exp[i]!=exp[i-1])pDistinct++;
outF<<"# PN "<<pDistinct<<" hyperparameters"<<endl;
outF<<"# columns: alpha beta expression "<<endl;
outF<<alpS[0]<<" "<<betS[0]<<" "<<exp[0]<<endl;
for(i=1;i<pAll;i++)
if(exp[i]!=exp[i-1])outF<<alpS[i]<<" "<<betS[i]<<" "<<exp[i]<<endl;
outF.close();
if(args.verbose){message("DONE.\n");timer.stop(1,'m');}
return 0;
}
#ifndef BIOC_BUILD
int main(int argc,char* argv[]){
return estimateHyperPar(&argc,argv);
}
#endif