X-Git-Url: https://git.donarmstrong.com/?p=rsem.git;a=blobdiff_plain;f=EM.cpp;h=7016b8944def26003c755697627019fe0dd0d6ea;hp=b2493b2fa21982125b0cba196325ca86487848c6;hb=refs%2Fheads%2Fupstream;hpb=635ca2939cfb1f519f19e9dec072ddd05e9fb450 diff --git a/EM.cpp b/EM.cpp index b2493b2..7016b89 100644 --- a/EM.cpp +++ b/EM.cpp @@ -7,6 +7,8 @@ #include #include #include +#include +#include #include #include "utils.h" @@ -42,6 +44,8 @@ #include "HitWrapper.h" #include "BamWriter.h" +#include "WriteResults.h" + using namespace std; const double STOP_CRITERIA = 0.001; @@ -55,7 +59,7 @@ struct Params { int read_type; int m, M; // m genes, M isoforms -int N0, N1, N2, N_tot; +READ_INT_TYPE N0, N1, N2, N_tot; int nThreads; @@ -66,8 +70,8 @@ bool genGibbsOut; // generate file for Gibbs sampler char refName[STRLEN], outName[STRLEN]; char imdName[STRLEN], statName[STRLEN]; -char refF[STRLEN], groupF[STRLEN], cntF[STRLEN], tiF[STRLEN]; -char mparamsF[STRLEN], bmparamsF[STRLEN]; +char refF[STRLEN], cntF[STRLEN], tiF[STRLEN]; +char mparamsF[STRLEN]; char modelF[STRLEN], thetaF[STRLEN]; char inpSamType; @@ -81,15 +85,21 @@ vector theta, eel; // eel : expected effective length double *probv, **countvs; Refs refs; -GroupInfo gi; Transcripts transcripts; ModelParams mparams; +bool hasSeed; +seedType seed; + template void init(ReadReader **&readers, HitContainer **&hitvs, double **&ncpvs, ModelType **&mhps) { - int nReads, nHits, rt; - int nrLeft, nhT, curnr; // nrLeft : number of reads left, nhT : hit threshold per thread, curnr: current number of reads + READ_INT_TYPE nReads; + HIT_INT_TYPE nHits; + int rt; // read type + + READ_INT_TYPE nrLeft, curnr; // nrLeft : number of reads left, curnr: current number of reads + HIT_INT_TYPE nhT; // nhT : hit threshold per thread char datF[STRLEN]; int s; @@ -127,7 +137,7 @@ void init(ReadReader **&readers, HitContainer **&hitvs, doubl ncpvs = new double*[nThreads]; for (int i = 0; i < nThreads; i++) { - int ntLeft = nThreads - i - 1; // # of threads left + HIT_INT_TYPE ntLeft = nThreads - i - 1; // # of threads left general_assert(readers[i]->locate(curnr), "Read indices files do not match!"); @@ -135,13 +145,13 @@ void init(ReadReader **&readers, HitContainer **&hitvs, doubl general_assert(hitvs[i]->read(fin), "Cannot read alignments from .dat file!"); --nrLeft; - if (verbose && nrLeft % 1000000 == 0) { printf("DAT %d reads left!\n", nrLeft); } + if (verbose && nrLeft % 1000000 == 0) { cout<< "DAT "<< nrLeft << " reads left"<< endl; } } ncpvs[i] = new double[hitvs[i]->getN()]; memset(ncpvs[i], 0, sizeof(double) * hitvs[i]->getN()); curnr += hitvs[i]->getN(); - if (verbose) { printf("Thread %d : N = %d, NHit = %d\n", i, hitvs[i]->getN(), hitvs[i]->getNHits()); } + if (verbose) { cout<<"Thread "<< i<< " : N = "<< hitvs[i]->getN()<< ", NHit = "<< hitvs[i]->getNHits()<< endl; } } fin.close(); @@ -175,16 +185,16 @@ void* E_STEP(void* arg) { ReadType read; - int N = hitv->getN(); + READ_INT_TYPE N = hitv->getN(); double sum; vector fracs; //to remove this, do calculation twice - int fr, to, id; + HIT_INT_TYPE fr, to, id; if (needCalcConPrb || updateModel) { reader->reset(); } if (updateModel) { mhp->init(); } memset(countv, 0, sizeof(double) * (M + 1)); - for (int i = 0; i < N; i++) { + for (READ_INT_TYPE i = 0; i < N; i++) { if (needCalcConPrb || updateModel) { general_assert(reader->next(read), "Can not load a read!"); } @@ -199,7 +209,7 @@ void* E_STEP(void* arg) { fracs[0] = probv[0] * ncpv[i]; if (fracs[0] < EPSILON) fracs[0] = 0.0; sum += fracs[0]; - for (int j = fr; j < to; j++) { + for (HIT_INT_TYPE j = fr; j < to; j++) { HitType &hit = hitv->getHitAt(j); if (needCalcConPrb) { hit.setConPrb(model->getConPrb(read, hit)); } id = j - fr + 1; @@ -213,7 +223,7 @@ void* E_STEP(void* arg) { countv[0] += fracs[0]; if (updateModel) { mhp->updateNoise(read, fracs[0]); } if (calcExpectedWeights) { ncpv[i] = fracs[0]; } - for (int j = fr; j < to; j++) { + for (HIT_INT_TYPE j = fr; j < to; j++) { HitType &hit = hitv->getHitAt(j); id = j - fr + 1; fracs[id] /= sum; @@ -224,7 +234,7 @@ void* E_STEP(void* arg) { } else if (calcExpectedWeights) { ncpv[i] = 0.0; - for (int j = fr; j < to; j++) { + for (HIT_INT_TYPE j = fr; j < to; j++) { HitType &hit = hitv->getHitAt(j); hit.setConPrb(0.0); } @@ -243,20 +253,20 @@ void* calcConProbs(void* arg) { double *ncpv = (double*)(params->ncpv); ReadType read; - int N = hitv->getN(); - int fr, to; + READ_INT_TYPE N = hitv->getN(); + HIT_INT_TYPE fr, to; assert(model->getNeedCalcConPrb()); reader->reset(); - for (int i = 0; i < N; i++) { + for (READ_INT_TYPE i = 0; i < N; i++) { general_assert(reader->next(read), "Can not load a read!"); fr = hitv->getSAt(i); to = hitv->getSAt(i + 1); ncpv[i] = model->getNoiseConPrb(read); - for (int j = fr; j < to; j++) { + for (HIT_INT_TYPE j = fr; j < to; j++) { HitType &hit = hitv->getHitAt(j); hit.setConPrb(model->getConPrb(read, hit)); } @@ -265,111 +275,11 @@ void* calcConProbs(void* arg) { return NULL; } -template -void calcExpectedEffectiveLengths(ModelType& model) { - int lb, ub, span; - double *pdf = NULL, *cdf = NULL, *clen = NULL; // clen[i] = sigma_{j=1}^{i}pdf[i]*(lb+i) - - model.getGLD().copyTo(pdf, cdf, lb, ub, span); - clen = new double[span + 1]; - clen[0] = 0.0; - for (int i = 1; i <= span; i++) { - clen[i] = clen[i - 1] + pdf[i] * (lb + i); - } - - eel.clear(); - eel.resize(M + 1, 0.0); - for (int i = 1; i <= M; i++) { - int totLen = refs.getRef(i).getTotLen(); - int fullLen = refs.getRef(i).getFullLen(); - int pos1 = max(min(totLen - fullLen + 1, ub) - lb, 0); - int pos2 = max(min(totLen, ub) - lb, 0); - - if (pos2 == 0) { eel[i] = 0.0; continue; } - - eel[i] = fullLen * cdf[pos1] + ((cdf[pos2] - cdf[pos1]) * (totLen + 1) - (clen[pos2] - clen[pos1])); - assert(eel[i] >= 0); - if (eel[i] < MINEEL) { eel[i] = 0.0; } - } - - delete[] pdf; - delete[] cdf; - delete[] clen; -} - template void writeResults(ModelType& model, double* counts) { - double denom; - char outF[STRLEN]; - FILE *fo; - - sprintf(modelF, "%s.model", statName); - model.write(modelF); - - //calculate tau values - double *tau = new double[M + 1]; - memset(tau, 0, sizeof(double) * (M + 1)); - - denom = 0.0; - for (int i = 1; i <= M; i++) - if (eel[i] >= EPSILON) { - tau[i] = theta[i] / eel[i]; - denom += tau[i]; - } - - general_assert(denom > 0, "No alignable reads?!"); - - for (int i = 1; i <= M; i++) { - tau[i] /= denom; - } - - //isoform level results - sprintf(outF, "%s.iso_res", imdName); - fo = fopen(outF, "w"); - for (int i = 1; i <= M; i++) { - const Transcript& transcript = transcripts.getTranscriptAt(i); - fprintf(fo, "%s%c", transcript.getTranscriptID().c_str(), (i < M ? '\t' : '\n')); - } - for (int i = 1; i <= M; i++) - fprintf(fo, "%.2f%c", counts[i], (i < M ? '\t' : '\n')); - for (int i = 1; i <= M; i++) - fprintf(fo, "%.15g%c", tau[i], (i < M ? '\t' : '\n')); - for (int i = 1; i <= M; i++) { - const Transcript& transcript = transcripts.getTranscriptAt(i); - fprintf(fo, "%s%c", transcript.getGeneID().c_str(), (i < M ? '\t' : '\n')); - } - fclose(fo); - - //gene level results - sprintf(outF, "%s.gene_res", imdName); - fo = fopen(outF, "w"); - for (int i = 0; i < m; i++) { - const string& gene_id = transcripts.getTranscriptAt(gi.spAt(i)).getGeneID(); - fprintf(fo, "%s%c", gene_id.c_str(), (i < m - 1 ? '\t' : '\n')); - } - for (int i = 0; i < m; i++) { - double sumC = 0.0; // sum of counts - int b = gi.spAt(i), e = gi.spAt(i + 1); - for (int j = b; j < e; j++) sumC += counts[j]; - fprintf(fo, "%.2f%c", sumC, (i < m - 1 ? '\t' : '\n')); - } - for (int i = 0; i < m; i++) { - double sumT = 0.0; // sum of tau values - int b = gi.spAt(i), e = gi.spAt(i + 1); - for (int j = b; j < e; j++) sumT += tau[j]; - fprintf(fo, "%.15g%c", sumT, (i < m - 1 ? '\t' : '\n')); - } - for (int i = 0; i < m; i++) { - int b = gi.spAt(i), e = gi.spAt(i + 1); - for (int j = b; j < e; j++) { - fprintf(fo, "%s%c", transcripts.getTranscriptAt(j).getTranscriptID().c_str(), (j < e - 1 ? ',' : (i < m - 1 ? '\t' :'\n'))); - } - } - fclose(fo); - - delete[] tau; - - if (verbose) { printf("Expression Results are written!\n"); } + sprintf(modelF, "%s.model", statName); + model.write(modelF); + writeResultsEM(M, refName, imdName, transcripts, theta, eel, countvs[0]); } template @@ -417,7 +327,6 @@ void EM() { Params fparams[nThreads]; pthread_t threads[nThreads]; pthread_attr_t attr; - void *status; int rc; @@ -465,7 +374,7 @@ void EM() { } for (int i = 0; i < nThreads; i++) { - rc = pthread_join(threads[i], &status); + rc = pthread_join(threads[i], NULL); pthread_assert(rc, "pthread_join", "Cannot join thread " + itos(i) + " (numbered from 0) at ROUND " + itos(ROUND) + "!"); } @@ -501,11 +410,11 @@ void EM() { if (bChange < change) bChange = change; } - if (verbose) printf("ROUND = %d, SUM = %.15g, bChange = %f, totNum = %d\n", ROUND, sum, bChange, totNum); + if (verbose) { cout<< "ROUND = "<< ROUND<< ", SUM = "<< setprecision(15)<< sum<< ", bChange = " << setprecision(6)<< bChange<< ", totNum = " << totNum<< endl; } } while (ROUND < MIN_ROUND || (totNum > 0 && ROUND < MAX_ROUND)); - //while (ROUND < MAX_ROUND); +// } while (ROUND < 1); - if (totNum > 0) fprintf(stderr, "Warning: RSEM reaches %d iterations before meeting the convergence criteria.\n", MAX_ROUND); + if (totNum > 0) { cout<< "Warning: RSEM reaches "<< MAX_ROUND<< " iterations before meeting the convergence criteria."<< endl; } //generate output file used by Gibbs sampler if (genGibbsOut) { @@ -515,59 +424,39 @@ void EM() { pthread_assert(rc, "pthread_create", "Cannot create thread " + itos(i) + " (numbered from 0) when generating files for Gibbs sampler!"); } for (int i = 0; i < nThreads; i++) { - rc = pthread_join(threads[i], &status); + rc = pthread_join(threads[i], NULL); pthread_assert(rc, "pthread_join", "Cannot join thread " + itos(i) + " (numbered from 0) when generating files for Gibbs sampler!"); } } model.setNeedCalcConPrb(false); sprintf(out_for_gibbs_F, "%s.ofg", imdName); - fo = fopen(out_for_gibbs_F, "w"); - fprintf(fo, "%d %d\n", M, N0); + ofstream fout(out_for_gibbs_F); + fout<< M<< " "<< N0<< endl; for (int i = 0; i < nThreads; i++) { - int numN = hitvs[i]->getN(); - for (int j = 0; j < numN; j++) { - int fr = hitvs[i]->getSAt(j); - int to = hitvs[i]->getSAt(j + 1); - int totNum = 0; - - if (ncpvs[i][j] >= EPSILON) { ++totNum; fprintf(fo, "%d %.15g ", 0, ncpvs[i][j]); } - for (int k = fr; k < to; k++) { + READ_INT_TYPE numN = hitvs[i]->getN(); + for (READ_INT_TYPE j = 0; j < numN; j++) { + HIT_INT_TYPE fr = hitvs[i]->getSAt(j); + HIT_INT_TYPE to = hitvs[i]->getSAt(j + 1); + HIT_INT_TYPE totNum = 0; + + if (ncpvs[i][j] >= EPSILON) { ++totNum; fout<< "0 "<< setprecision(15)<< ncpvs[i][j]<< " "; } + for (HIT_INT_TYPE k = fr; k < to; k++) { HitType &hit = hitvs[i]->getHitAt(k); if (hit.getConPrb() >= EPSILON) { ++totNum; - fprintf(fo, "%d %.15g ", hit.getSid(), hit.getConPrb()); + fout<< hit.getSid()<< " "<< setprecision(15)<< hit.getConPrb()<< " "; } } - if (totNum > 0) { fprintf(fo, "\n"); } + if (totNum > 0) { fout<< endl; } } } - fclose(fo); + fout.close(); } - sprintf(thetaF, "%s.theta", statName); - fo = fopen(thetaF, "w"); - fprintf(fo, "%d\n", M + 1); - - // output theta' - for (int i = 0; i < M; i++) fprintf(fo, "%.15g ", theta[i]); - fprintf(fo, "%.15g\n", theta[M]); - - //calculate expected effective lengths for each isoform - calcExpectedEffectiveLengths(model); - - //correct theta vector - sum = theta[0]; - for (int i = 1; i <= M; i++) - if (eel[i] < EPSILON) { theta[i] = 0.0; } - else sum += theta[i]; - - general_assert(sum >= EPSILON, "No Expected Effective Length is no less than" + ftos(MINEEL, 6) + "?!"); - - for (int i = 0; i <= M; i++) theta[i] /= sum; - //calculate expected weights and counts using learned parameters + //just use the raw theta learned from the data, do not correct for eel or mw updateModel = false; calcExpectedWeights = true; for (int i = 0; i <= M; i++) probv[i] = theta[i]; for (int i = 0; i < nThreads; i++) { @@ -575,7 +464,7 @@ void EM() { pthread_assert(rc, "pthread_create", "Cannot create thread " + itos(i) + " (numbered from 0) when calculating expected weights!"); } for (int i = 0; i < nThreads; i++) { - rc = pthread_join(threads[i], &status); + rc = pthread_join(threads[i], NULL); pthread_assert(rc, "pthread_join", "Cannot join thread " + itos(i) + " (numbered from 0) when calculating expected weights!"); } model.setNeedCalcConPrb(false); @@ -589,15 +478,18 @@ void EM() { /* destroy attribute */ pthread_attr_destroy(&attr); - //convert theta' to theta - double *mw = model.getMW(); - sum = 0.0; - for (int i = 0; i <= M; i++) { - theta[i] = (mw[i] < EPSILON ? 0.0 : theta[i] / mw[i]); - sum += theta[i]; - } - assert(sum >= EPSILON); - for (int i = 0; i <= M; i++) theta[i] /= sum; + + sprintf(thetaF, "%s.theta", statName); + fo = fopen(thetaF, "w"); + fprintf(fo, "%d\n", M + 1); + + // output theta' + for (int i = 0; i < M; i++) fprintf(fo, "%.15g ", theta[i]); + fprintf(fo, "%.15g\n", theta[M]); + + //calculate expected effective lengths for each isoform + calcExpectedEffectiveLengths(M, refs, model, eel); + polishTheta(M, theta, eel, model.getMW()); // output theta for (int i = 0; i < M; i++) fprintf(fo, "%.15g ", theta[i]); @@ -611,27 +503,29 @@ void EM() { sprintf(outBamF, "%s.transcript.bam", outName); if (bamSampling) { - int local_N; - int fr, to, len, id; + READ_INT_TYPE local_N; + HIT_INT_TYPE fr, to, len, id; vector arr; - uniform01 rg(engine_type(time(NULL))); + engine_type engine(hasSeed ? seed : time(NULL)); + uniform_01_dist uniform_01; + uniform_01_generator rg(engine, uniform_01); - if (verbose) printf("Begin to sample reads from their posteriors.\n"); + if (verbose) cout<< "Begin to sample reads from their posteriors."<< endl; for (int i = 0; i < nThreads; i++) { local_N = hitvs[i]->getN(); - for (int j = 0; j < local_N; j++) { + for (READ_INT_TYPE j = 0; j < local_N; j++) { fr = hitvs[i]->getSAt(j); to = hitvs[i]->getSAt(j + 1); len = to - fr + 1; arr.assign(len, 0); arr[0] = ncpvs[i][j]; - for (int k = fr; k < to; k++) arr[k - fr + 1] = arr[k - fr] + hitvs[i]->getHitAt(k).getConPrb(); + for (HIT_INT_TYPE k = fr; k < to; k++) arr[k - fr + 1] = arr[k - fr] + hitvs[i]->getHitAt(k).getConPrb(); id = (arr[len - 1] < EPSILON ? -1 : sample(rg, arr, len)); // if all entries in arr are 0, let id be -1 - for (int k = fr; k < to; k++) hitvs[i]->getHitAt(k).setConPrb(k - fr + 1 == id ? 1.0 : 0.0); + for (HIT_INT_TYPE k = fr; k < to; k++) hitvs[i]->getHitAt(k).setConPrb(k - fr + 1 == id ? 1.0 : 0.0); } } - if (verbose) printf("Sampling is finished.\n"); + if (verbose) cout<< "Sampling is finished."<< endl; } BamWriter writer(inpSamType, inpSamF, pt_fn_list, outBamF, transcripts); @@ -646,8 +540,8 @@ int main(int argc, char* argv[]) { ifstream fin; bool quiet = false; - if (argc < 5) { - printf("Usage : rsem-run-em refName read_type sampleName sampleToken [-p #Threads] [-b samInpType samInpF has_fn_list_? [fn_list]] [-q] [--gibbs-out] [--sampling]\n\n"); + if (argc < 6) { + printf("Usage : rsem-run-em refName read_type sampleName imdName statName [-p #Threads] [-b samInpType samInpF has_fn_list_? [fn_list]] [-q] [--gibbs-out] [--sampling] [--seed seed]\n\n"); printf(" refName: reference name\n"); printf(" read_type: 0 single read without quality score; 1 single read with quality score; 2 paired-end read without quality score; 3 paired-end read with quality score.\n"); printf(" sampleName: sample's name, including the path\n"); @@ -657,6 +551,7 @@ int main(int argc, char* argv[]) { printf(" -q: set it quiet\n"); printf(" --gibbs-out: generate output file used by Gibbs sampler. (default: off)\n"); printf(" --sampling: sample each read from its posterior distribution when bam file is generated. (default: off)\n"); + printf(" --seed uint32: the seed used for the BAM sampling. (default: off)\n"); printf("// model parameters should be in imdName.mparams.\n"); exit(-1); } @@ -666,8 +561,8 @@ int main(int argc, char* argv[]) { strcpy(refName, argv[1]); read_type = atoi(argv[2]); strcpy(outName, argv[3]); - sprintf(imdName, "%s.temp/%s", argv[3], argv[4]); - sprintf(statName, "%s.stat/%s", argv[3], argv[4]); + strcpy(imdName, argv[4]); + strcpy(statName, argv[5]); nThreads = 1; @@ -675,8 +570,9 @@ int main(int argc, char* argv[]) { bamSampling = false; genGibbsOut = false; pt_fn_list = pt_chr_list = NULL; + hasSeed = false; - for (int i = 5; i < argc; i++) { + for (int i = 6; i < argc; i++) { if (!strcmp(argv[i], "-p")) { nThreads = atoi(argv[i + 1]); } if (!strcmp(argv[i], "-b")) { genBamF = true; @@ -690,6 +586,12 @@ int main(int argc, char* argv[]) { if (!strcmp(argv[i], "-q")) { quiet = true; } if (!strcmp(argv[i], "--gibbs-out")) { genGibbsOut = true; } if (!strcmp(argv[i], "--sampling")) { bamSampling = true; } + if (!strcmp(argv[i], "--seed")) { + hasSeed = true; + int len = strlen(argv[i + 1]); + seed = 0; + for (int k = 0; k < len; k++) seed = seed * 10 + (argv[i + 1][k] - '0'); + } } general_assert(nThreads > 0, "Number of threads should be bigger than 0!"); @@ -700,9 +602,6 @@ int main(int argc, char* argv[]) { sprintf(refF, "%s.seq", refName); refs.loadRefs(refF); M = refs.getM(); - sprintf(groupF, "%s.grp", refName); - gi.load(groupF); - m = gi.getm(); sprintf(tiF, "%s.ti", refName); transcripts.readFrom(tiF); @@ -717,7 +616,7 @@ int main(int argc, char* argv[]) { general_assert(N1 > 0, "There are no alignable reads!"); - if (nThreads > N1) nThreads = N1; + if ((READ_INT_TYPE)nThreads > N1) nThreads = N1; //set model parameters mparams.M = M; @@ -748,7 +647,7 @@ int main(int argc, char* argv[]) { time_t b = time(NULL); - printTimeUsed(a, b); + printTimeUsed(a, b, "EM.cpp"); return 0; }