/* ViterbiScore() unit test
*
* We can compare these scores to GViterbi() almost exactly; the only
* differences should be negligible roundoff errors. Must convert
* the optimized profile to lspace, though, rather than pspace.
*/
static void
utest_viterbi_score(ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, int M, int L, int N)
{
P7_HMM *hmm = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
ESL_DSQ *dsq = malloc(sizeof(ESL_DSQ) * (L+2));
P7_OMX *ox = p7_omx_Create(M, 0, 0);
P7_GMX *gx = p7_gmx_Create(M, L);
float sc1, sc2;
p7_oprofile_Sample(r, abc, bg, M, L, &hmm, &gm, &om);
p7_oprofile_Logify(om);
while (N--)
{
esl_rsq_xfIID(r, bg->f, abc->K, L, dsq);
p7_ViterbiScore(dsq, L, om, ox, &sc1);
p7_GViterbi (dsq, L, gm, gx, &sc2);
if (fabs(sc1-sc2) > 0.001) esl_fatal("viterbi score unit test failed: scores differ");
}
free(dsq);
p7_hmm_Destroy(hmm);
p7_omx_Destroy(ox);
p7_gmx_Destroy(gx);
p7_profile_Destroy(gm);
p7_oprofile_Destroy(om);
}
int
main(int argc, char **argv)
{
ESL_GETOPTS *go = esl_getopts_CreateDefaultApp(options, 0, argc, argv, banner, usage);
ESL_RANDOMNESS *r = esl_randomness_CreateFast(esl_opt_GetInteger(go, "-s"));
ESL_ALPHABET *abc = NULL;
P7_BG *bg = NULL;
P7_HMM *hmm = NULL;
P7_OPROFILE *om = NULL;
int M = esl_opt_GetInteger(go, "-M");
int L = esl_opt_GetInteger(go, "-L");
/* Sample a random HMM and optimized profile, in amino acid alphabet. */
if ((abc = esl_alphabet_Create(eslAMINO)) == NULL) esl_fatal("failed to create alphabet");
if ((bg = p7_bg_Create(abc)) == NULL) esl_fatal("failed to create null model");
if (( p7_oprofile_Sample(r, abc, bg, M, L, &hmm, NULL, &om)) != eslOK) esl_fatal("failed to sample HMM and profile");
/* unit test(s) */
utest_ReadWrite(hmm, om);
p7_oprofile_Destroy(om);
p7_hmm_Destroy(hmm);
p7_bg_Destroy(bg);
esl_alphabet_Destroy(abc);
esl_randomness_Destroy(r);
esl_getopts_Destroy(go);
return eslOK;
}
/*
* compare to GForward() scores.
*/
static void
utest_fwdback(ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, int M, int L, int N)
{
char *msg = "forward/backward unit test failed";
P7_HMM *hmm = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
ESL_DSQ *dsq = malloc(sizeof(ESL_DSQ) * (L+2));
P7_OMX *fwd = p7_omx_Create(M, 0, L);
P7_OMX *bck = p7_omx_Create(M, 0, L);
P7_OMX *oxf = p7_omx_Create(M, L, L);
P7_OMX *oxb = p7_omx_Create(M, L, L);
P7_GMX *gx = p7_gmx_Create(M, L);
float tolerance;
float fsc1, fsc2;
float bsc1, bsc2;
float generic_sc;
p7_FLogsumInit();
if (p7_FLogsumError(-0.4, -0.5) > 0.0001) tolerance = 1.0; /* weaker test against GForward() */
else tolerance = 0.0001; /* stronger test: FLogsum() is in slow exact mode. */
p7_oprofile_Sample(r, abc, bg, M, L, &hmm, &gm, &om);
while (N--)
{
esl_rsq_xfIID(r, bg->f, abc->K, L, dsq);
p7_Forward (dsq, L, om, oxf, &fsc1);
p7_Backward (dsq, L, om, oxf, oxb, &bsc1);
p7_ForwardParser (dsq, L, om, fwd, &fsc2);
p7_BackwardParser(dsq, L, om, fwd, bck, &bsc2);
p7_GForward (dsq, L, gm, gx, &generic_sc);
/* Forward and Backward scores should agree with high tolerance */
if (fabs(fsc1-bsc1) > 0.0001) esl_fatal(msg);
if (fabs(fsc2-bsc2) > 0.0001) esl_fatal(msg);
if (fabs(fsc1-fsc2) > 0.0001) esl_fatal(msg);
/* GForward scores should approximate Forward scores,
* with tolerance that depends on how logsum.c was compiled
*/
if (fabs(fsc1-generic_sc) > tolerance) esl_fatal(msg);
}
free(dsq);
p7_hmm_Destroy(hmm);
p7_omx_Destroy(oxb);
p7_omx_Destroy(oxf);
p7_omx_Destroy(bck);
p7_omx_Destroy(fwd);
p7_gmx_Destroy(gx);
p7_profile_Destroy(gm);
p7_oprofile_Destroy(om);
}
/* compare results to GDecoding(). */
static void
utest_null2_expectation(ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, int M, int L, int N, float tolerance)
{
char *msg = "decoding unit test failed";
P7_HMM *hmm = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
ESL_DSQ *dsq = malloc(sizeof(ESL_DSQ) * (L+2));
P7_OMX *fwd = p7_omx_Create(M, L, L);
P7_OMX *bck = p7_omx_Create(M, L, L);
P7_OMX *pp = p7_omx_Create(M, L, L);
P7_GMX *gxf = p7_gmx_Create(M, L);
P7_GMX *gxb = p7_gmx_Create(M, L);
P7_GMX *gpp = p7_gmx_Create(M, L);
float *on2 = malloc(sizeof(float) * abc->Kp);
float *gn2 = malloc(sizeof(float) * abc->Kp);
float fsc1, fsc2;
float bsc1, bsc2;
if (!gn2 || !on2) esl_fatal(msg);
if (p7_oprofile_Sample(r, abc, bg, M, L, &hmm, &gm, &om) != eslOK) esl_fatal(msg);
while (N--)
{
if (esl_rsq_xfIID(r, bg->f, abc->K, L, dsq) != eslOK) esl_fatal(msg);
if (p7_Forward (dsq, L, om, fwd, &fsc1) != eslOK) esl_fatal(msg);
if (p7_Backward (dsq, L, om, fwd, bck, &bsc1) != eslOK) esl_fatal(msg);
if (p7_Decoding(om, fwd, bck, pp) != eslOK) esl_fatal(msg);
if (p7_Null2_ByExpectation(om, pp, on2) != eslOK) esl_fatal(msg);
if (p7_GForward (dsq, L, gm, gxf, &fsc2) != eslOK) esl_fatal(msg);
if (p7_GBackward(dsq, L, gm, gxb, &bsc2) != eslOK) esl_fatal(msg);
if (p7_GDecoding(gm, gxf, gxb, gpp) != eslOK) esl_fatal(msg);
if (p7_GNull2_ByExpectation(gm, gpp, gn2) != eslOK) esl_fatal(msg);
if (esl_vec_FCompare(gn2, on2, abc->Kp, tolerance) != eslOK) esl_fatal(msg);
}
p7_gmx_Destroy(gpp);
p7_gmx_Destroy(gxf);
p7_gmx_Destroy(gxb);
p7_omx_Destroy(pp);
p7_omx_Destroy(fwd);
p7_omx_Destroy(bck);
free(on2);
free(gn2);
free(dsq);
p7_oprofile_Destroy(om);
p7_profile_Destroy(gm);
p7_hmm_Destroy(hmm);
}
/* compare results to GDecoding(). */
static void
utest_decoding(ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, int M, int L, int N, float tolerance)
{
char *msg = "decoding unit test failed";
P7_HMM *hmm = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
ESL_DSQ *dsq = malloc(sizeof(ESL_DSQ) * (L+2));
P7_OMX *fwd = p7_omx_Create(M, L, L);
P7_OMX *bck = p7_omx_Create(M, L, L);
P7_OMX *pp = p7_omx_Create(M, L, L);
P7_GMX *gxf = p7_gmx_Create(M, L);
P7_GMX *gxb = p7_gmx_Create(M, L);
P7_GMX *gxp1 = p7_gmx_Create(M, L);
P7_GMX *gxp2 = p7_gmx_Create(M, L);
float fsc1, fsc2;
float bsc1, bsc2;
if (p7_oprofile_Sample(r, abc, bg, M, L, &hmm, &gm, &om) != eslOK) esl_fatal(msg);
while (N--)
{
if (esl_rsq_xfIID(r, bg->f, abc->K, L, dsq) != eslOK) esl_fatal(msg);
if (p7_Forward (dsq, L, om, fwd, &fsc1) != eslOK) esl_fatal(msg);
if (p7_Backward (dsq, L, om, fwd, bck, &bsc1) != eslOK) esl_fatal(msg);
if (p7_Decoding(om, fwd, bck, pp) != eslOK) esl_fatal(msg);
if (p7_omx_FDeconvert(pp, gxp1) != eslOK) esl_fatal(msg);
if (p7_GForward (dsq, L, gm, gxf, &fsc2) != eslOK) esl_fatal(msg);
if (p7_GBackward(dsq, L, gm, gxb, &bsc2) != eslOK) esl_fatal(msg);
if (p7_GDecoding(gm, gxf, gxb, gxp2) != eslOK) esl_fatal(msg);
// p7_gmx_Dump(stdout, gxp1, p7_DEFAULT);
// p7_gmx_Dump(stdout, gxp2, p7_DEFAULT);
if (p7_gmx_Compare(gxp1, gxp2, tolerance) != eslOK) esl_fatal(msg);
}
p7_gmx_Destroy(gxp1);
p7_gmx_Destroy(gxp2);
p7_gmx_Destroy(gxf);
p7_gmx_Destroy(gxb);
p7_omx_Destroy(fwd);
p7_omx_Destroy(bck);
p7_omx_Destroy(pp);
free(dsq);
p7_oprofile_Destroy(om);
p7_profile_Destroy(gm);
p7_hmm_Destroy(hmm);
}
/* ViterbiFilter() unit test
*
* We can check that scores are identical (within machine error) to
* scores of generic DP with scores rounded the same way. Do this for
* a random model of length <M>, for <N> test sequences of length <L>.
*
* We assume that we don't accidentally generate a high-scoring random
* sequence that overflows ViterbiFilter()'s limited range.
*
*/
static void
utest_viterbi_filter(ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, int M, int L, int N)
{
P7_HMM *hmm = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
ESL_DSQ *dsq = malloc(sizeof(ESL_DSQ) * (L+2));
P7_OMX *ox = p7_omx_Create(M, 0, 0);
P7_GMX *gx = p7_gmx_Create(M, L);
float sc1, sc2;
p7_oprofile_Sample(r, abc, bg, M, L, &hmm, &gm, &om);
p7_profile_SameAsVF(om, gm); /* round and scale the scores in <gm> the same as in <om> */
#if 0
p7_oprofile_Dump(stdout, om); // dumps the optimized profile
p7_omx_SetDumpMode(stdout, ox, TRUE); // makes the fast DP algorithms dump their matrices
#endif
while (N--)
{
esl_rsq_xfIID(r, bg->f, abc->K, L, dsq);
p7_ViterbiFilter(dsq, L, om, ox, &sc1);
p7_GViterbi (dsq, L, gm, gx, &sc2);
#if 0
p7_gmx_Dump(stdout, gx); // dumps a generic DP matrix
#endif
sc2 /= om->scale_w;
sc2 -= 3.0;
if (fabs(sc1-sc2) > 0.001) esl_fatal("viterbi filter unit test failed: scores differ (%.2f, %.2f)", sc1, sc2);
}
free(dsq);
p7_hmm_Destroy(hmm);
p7_omx_Destroy(ox);
p7_gmx_Destroy(gx);
p7_profile_Destroy(gm);
p7_oprofile_Destroy(om);
}
/*
* 1. Compare accscore to GOptimalAccuracy().
* 2. Compare trace to GOATrace().
*
* Note: This test is subject to some expected noise and can fail
* for entirely innocent reasons. Generic Forward/Backward calculations with
* p7_GForward(), p7_GBackward() use coarse-grain table lookups to sum
* log probabilities, and sufficient roundoff error can accumulate to
* change the optimal accuracy traceback, causing this test to fail.
* So, if optacc_utest fails, before you go looking for bugs, first
* go to ../logsum.c, change the #ifdef to activate the slow/accurate
* version, recompile and rerun optacc_utest. If the failure goes away,
* you can ignore it. - SRE, Wed Dec 17 09:45:31 2008
*/
static void
utest_optacc(ESL_GETOPTS *go, ESL_RANDOMNESS *r, ESL_ALPHABET *abc, P7_BG *bg, int M, int L, int N)
{
char *msg = "optimal accuracy unit test failed";
P7_HMM *hmm = NULL;
P7_PROFILE *gm = NULL;
P7_OPROFILE *om = NULL;
ESL_SQ *sq = esl_sq_CreateDigital(abc);
P7_OMX *ox1 = p7_omx_Create(M, L, L);
P7_OMX *ox2 = p7_omx_Create(M, L, L);
P7_GMX *gx1 = p7_gmx_Create(M, L);
P7_GMX *gx2 = p7_gmx_Create(M, L);
P7_TRACE *tr = p7_trace_CreateWithPP();
P7_TRACE *trg = p7_trace_CreateWithPP();
P7_TRACE *tro = p7_trace_CreateWithPP();
float accscore_o;
float fsc, bsc, accscore;
float fsc_g, bsc_g, accscore_g, accscore_g2;
float pptol = 0.01;
float sctol = 0.001;
float gtol;
p7_FLogsumInit();
gtol = ( (p7_FLogsumError(-0.4, -0.5) > 0.0001) ? 0.1 : 0.001);
if (p7_oprofile_Sample(r, abc, bg, M, L, &hmm, &gm, &om)!= eslOK) esl_fatal(msg);
while (N--)
{
if (p7_ProfileEmit(r, hmm, gm, bg, sq, tro) != eslOK) esl_fatal(msg);
if (p7_omx_GrowTo(ox1, M, sq->n, sq->n) != eslOK) esl_fatal(msg);
if (p7_omx_GrowTo(ox2, M, sq->n, sq->n) != eslOK) esl_fatal(msg);
if (p7_gmx_GrowTo(gx1, M, sq->n) != eslOK) esl_fatal(msg);
if (p7_gmx_GrowTo(gx2, M, sq->n) != eslOK) esl_fatal(msg);
if (p7_Forward (sq->dsq, sq->n, om, ox1, &fsc) != eslOK) esl_fatal(msg);
if (p7_Backward(sq->dsq, sq->n, om, ox1, ox2, &bsc) != eslOK) esl_fatal(msg);
if (p7_Decoding(om, ox1, ox2, ox2) != eslOK) esl_fatal(msg);
if (p7_OptimalAccuracy(om, ox2, ox1, &accscore) != eslOK) esl_fatal(msg);
#if 0
p7_omx_FDeconvert(ox1, gx1);
p7_gmx_Dump(stdout, gx1, p7_DEFAULT);
p7_omx_FDeconvert(ox2, gx1);
p7_gmx_Dump(stdout, gx1, p7_DEFAULT);
#endif
if (p7_OATrace(om, ox2, ox1, tr) != eslOK) esl_fatal(msg);
if (p7_GForward (sq->dsq, sq->n, gm, gx1, &fsc_g) != eslOK) esl_fatal(msg);
if (p7_GBackward(sq->dsq, sq->n, gm, gx2, &bsc_g) != eslOK) esl_fatal(msg);
#if 0
p7_gmx_Dump(stdout, gx1, p7_DEFAULT); /* fwd */
p7_gmx_Dump(stdout, gx2, p7_DEFAULT); /* bck */
#endif
if (p7_GDecoding(gm, gx1, gx2, gx2) != eslOK) esl_fatal(msg);
if (p7_GOptimalAccuracy(gm, gx2, gx1, &accscore_g) != eslOK) esl_fatal(msg);
#if 0
p7_gmx_Dump(stdout, gx1, p7_DEFAULT); /* oa */
p7_gmx_Dump(stdout, gx2, p7_DEFAULT); /* pp */
#endif
if (p7_GOATrace(gm, gx2, gx1, trg) != eslOK) esl_fatal(msg);
if (p7_trace_SetPP(tro, gx2) != eslOK) esl_fatal(msg);
if (esl_opt_GetBoolean(go, "--traces"))
{
p7_trace_Dump(stdout, tro, gm, sq->dsq);
p7_trace_Dump(stdout, tr, gm, sq->dsq);
p7_trace_Dump(stdout, trg, gm, sq->dsq);
}
if (p7_trace_Validate(tr, abc, sq->dsq, NULL) != eslOK) esl_fatal(msg);
if (p7_trace_Validate(trg, abc, sq->dsq, NULL) != eslOK) esl_fatal(msg);
if (p7_trace_Compare(tr, trg, pptol) != eslOK) esl_fatal(msg);
accscore_o = p7_trace_GetExpectedAccuracy(tro); /* according to gx2; see p7_trace_SetPP() call above */
accscore_g2 = p7_trace_GetExpectedAccuracy(trg);
#if 0
printf("%f %f %f %f\n", accscore, accscore_g, accscore_g2, accscore_o);
#endif
if (esl_FCompare(fsc, bsc, sctol) != eslOK) esl_fatal(msg);
if (esl_FCompare(fsc_g, bsc_g, gtol) != eslOK) esl_fatal(msg);
if (esl_FCompare(fsc, fsc_g, gtol) != eslOK) esl_fatal(msg);
if (esl_FCompare(accscore, accscore_g, gtol) != eslOK) esl_fatal(msg);
if (esl_FCompare(accscore_g, accscore_g2, gtol) != eslOK) esl_fatal(msg);
if (accscore_g2 < accscore_o) esl_fatal(msg);
/* the above deserves explanation:
* - accscore_o is the accuracy of the originally emitted trace, according
* to the generic posterior decoding matrix <gx2>. This is a lower bound
* on the expected # of accurately aligned residues found by a DP
* optimization.
* - accscore is the accuracy found by the fast (vector) code DP implementation.
* - accscore_g is the accuracy found by the generic DP implementation.
* accscore and accscore_g should be nearly identical,
* within tolerance of roundoff error accumulation and
* the imprecision of Logsum() tables.
* - accscore_g2 is the accuracy of the traceback identified by the generic
* DP implementation. It should be identical (within order-of-evaluation
* roundoff error) to accscore_g.
*
* the "accscore_g2 < accscore_o" test is carefully contrived.
* accscore_o is a theoretical lower bound but because of fp error,
* accscore and (much more rarely) even accscore_g can exceed accscore_o.
* accscore_g2, however, is calculated with identical order of evaluation
* as accscore_o if the optimal trace does turn out to be identical to
* the originally emitted trace. It should be extremely unlikely (though
* not impossible) for accscore_o to exceed accscore_g2. (The DP algorithm
* would have to identify a trace that was different than the original trace,
* which the DP algorithm, by order-of-evaluation, assigned higher accuracy,
* but order-of-evaluation in traceback dependent code assigned lower accuracy.
* [xref J5/29]
*/
esl_sq_Reuse(sq);
p7_trace_Reuse(tr);
p7_trace_Reuse(trg);
p7_trace_Reuse(tro);
}
p7_trace_Destroy(tro);
p7_trace_Destroy(trg);
p7_trace_Destroy(tr);
p7_gmx_Destroy(gx2);
p7_gmx_Destroy(gx1);
p7_omx_Destroy(ox2);
p7_omx_Destroy(ox1);
esl_sq_Destroy(sq);
p7_oprofile_Destroy(om);
p7_profile_Destroy(gm);
p7_hmm_Destroy(hmm);
}