def main(dictAlg, sortedAlgs, targets, outputdir='.', verbose=True, function_targets_line=True): # [1, 13, 101] """Generate one table per func with results of multiple algorithms.""" """Difference with the first version: * numbers aligned using the decimal separator * premices for dispersion measure * significance test against best algorithm * table width... """ # TODO: method is long, terrible to read, split if possible if not bestalg.bestalgentries2009: bestalg.loadBBOB2009() # Sort data per dimension and function dictData = {} dsListperAlg = list(dictAlg[i] for i in sortedAlgs) for n, entries in enumerate(dsListperAlg): tmpdictdim = entries.dictByDim() for d in tmpdictdim: tmpdictfun = tmpdictdim[d].dictByFunc() for f in tmpdictfun: dictData.setdefault((d, f), {})[n] = tmpdictfun[f] nbtests = len(dictData) for df in dictData: # Generate one table per df # best 2009 refalgentry = bestalg.bestalgentries2009[df] refalgert = refalgentry.detERT(targets) refalgevals = (refalgentry.detEvals((targetf, ))[0][0]) refalgnbruns = len(refalgevals) refalgnbsucc = numpy.sum(numpy.isnan(refalgevals) == False) # Process the data # The following variables will be lists of elements each corresponding # to an algorithm algnames = [] #algdata = [] algerts = [] algevals = [] algdisp = [] algnbsucc = [] algnbruns = [] algmedmaxevals = [] algmedfinalfunvals = [] algtestres = [] algentries = [] for n in sorted(dictData[df].keys()): entries = dictData[df][n] # the number of datasets for a given dimension and function (df) # should be strictly 1. TODO: find a way to warn # TODO: do this checking before... why wasn't it triggered by ppperprof? if len(entries) > 1: txt = ("There is more than a single entry associated with " "folder %s on %d-D f%d." % (sortedAlgs[n], df[0], df[1])) raise Exception(txt) entry = entries[0] algentries.append(entry) algnames.append(sortedAlgs[n]) evals = entry.detEvals(targets) #tmpdata = [] tmpdisp = [] tmpert = [] for i, e in enumerate(evals): succ = (numpy.isnan(e) == False) e[succ == False] = entry.maxevals[succ == False] ert = toolsstats.sp(e, issuccessful=succ)[0] #tmpdata.append(ert/refalgert[i]) if succ.any(): tmp = toolsstats.drawSP(e[succ], entry.maxevals[succ == False], [10, 50, 90], samplesize=samplesize)[0] tmpdisp.append((tmp[-1] - tmp[0]) / 2.) else: tmpdisp.append(numpy.nan) tmpert.append(ert) algerts.append(tmpert) algevals.append(evals) #algdata.append(tmpdata) algdisp.append(tmpdisp) algmedmaxevals.append(numpy.median(entry.maxevals)) algmedfinalfunvals.append(numpy.median(entry.finalfunvals)) #algmedmaxevals.append(numpy.median(entry.maxevals)/df[0]) #algmedfinalfunvals.append(numpy.median(entry.finalfunvals)) algtestres.append(significancetest(refalgentry, entry, targets)) # determine success probability for Df = 1e-8 e = entry.detEvals((targetf, ))[0] algnbsucc.append(numpy.sum(numpy.isnan(e) == False)) algnbruns.append(len(e)) # Process over all data # find best values... nalgs = len(dictData[df]) maxRank = 1 + numpy.floor( 0.14 * nalgs) # number of algs to be displayed in bold isBoldArray = [] # Point out the best values algfinaldata = [ ] # Store median function values/median number of function evaluations tmptop = getTopIndicesOfColumns(algerts, maxRank=maxRank) for i, erts in enumerate(algerts): tmp = [] for j, ert in enumerate(erts): # algi targetj tmp.append(i in tmptop[j] or (nalgs > 7 and algerts[i][j] <= 3. * refalgert[j])) isBoldArray.append(tmp) algfinaldata.append((algmedfinalfunvals[i], algmedmaxevals[i])) # significance test of best given algorithm against all others best_alg_idx = numpy.array(algerts).argsort(0)[ 0, :] # indexed by target index significance_versus_others = significance_all_best_vs_other( algentries, targets, best_alg_idx)[0] # Create the table table = [] spec = r'@{}c@{}|*{%d}{@{\,}r@{}X@{\,}}|@{}r@{}@{}l@{}' % ( len(targets) ) # in case StrLeft not working: replaced c@{} with l@{ } spec = r'@{}c@{}|*{%d}{@{}r@{}X@{}}|@{}r@{}@{}l@{}' % ( len(targets) ) # in case StrLeft not working: replaced c@{} with l@{ } extraeol = [] # Generate header lines if with_table_heading: header = funInfos[df[1]] if funInfos else 'f%d' % df[1] table.append([ r'\multicolumn{%d}{@{\,}c@{\,}}{{\textbf{%s}}}' % (2 * len(targets) + 2, header) ]) extraeol.append('') if function_targets_line is True or (function_targets_line and df[1] in function_targets_line): curline = [r'$\Delta f_\mathrm{opt}$'] for t in targets[0:-1]: curline.append(r'\multicolumn{2}{@{\,}X@{\,}}{%s}' % writeFEvals2(t, precision=1, isscientific=True)) curline.append( r'\multicolumn{2}{@{\,}X@{}|}{%s}' % writeFEvals2(targets[-1], precision=1, isscientific=True)) curline.append(r'\multicolumn{2}{@{}l@{}}{\#succ}') table.append(curline) extraeol.append(r'\hline') # extraeol.append(r'\hline\arrayrulecolor{tableShade}') curline = [r'ERT$_{\text{best}}$' ] if with_table_heading else [r'\textbf{f%d}' % df[1]] for i in refalgert[0:-1]: curline.append(r'\multicolumn{2}{@{\,}X@{\,}}{%s}' % writeFEvalsMaxPrec(float(i), 2)) curline.append(r'\multicolumn{2}{@{\,}X@{\,}|}{%s}' % writeFEvalsMaxPrec(float(refalgert[-1]), 2)) curline.append('%d' % refalgnbsucc) if refalgnbsucc: curline.append('/%d' % refalgnbruns) #curline.append(curline[0]) table.append(curline) extraeol.append('') #for i, gna in enumerate(zip((1, 2, 3), ('bla', 'blo', 'bli'))): #print i, gna, gno #set_trace() # Format data #if df == (5, 17): #set_trace() header = r'\providecommand{\ntables}{7}' for i, alg in enumerate(algnames): #algname, entries, irs, line, line2, succ, runs, testres1alg in zip(algnames, #data, dispersion, isBoldArray, isItalArray, nbsucc, nbruns, testres): commandname = r'\alg%stables' % numtotext(i) # header += r'\providecommand{%s}{{%s}{}}' % (commandname, str_to_latex(strip_pathname(alg))) header += r'\providecommand{%s}{\StrLeft{%s}{\ntables}}' % ( commandname, str_to_latex(strip_pathname(alg))) curline = [ commandname + r'\hspace*{\fill}' ] # each list element becomes a &-separated table entry? for j, tmp in enumerate( zip( algerts[i], algdisp[i], # j is target index isBoldArray[i], algtestres[i])): ert, dispersion, isBold, testres = tmp alignment = '@{\,}X@{\,}' if j == len(algerts[i]) - 1: alignment = '@{\,}X@{\,}|' data = ert / refalgert[j] # write star for significance against all other algorithms str_significance_subsup = '' if (len(best_alg_idx) > 0 and len(significance_versus_others) > 0 and i == best_alg_idx[j] and nbtests * significance_versus_others[j][1] < 0.05): logp = -numpy.ceil( numpy.log10( nbtests * significance_versus_others[j][1])) str_significance_subsup = r"^{%s%s}" % ( significance_vs_others_symbol, str(int(logp)) if logp > 1 else '') # moved out of the above else: this was a bug!? z, p = testres if (nbtests * p) < 0.05 and data < 1. and z < 0.: if not numpy.isinf(refalgert[j]): tmpevals = algevals[i][j].copy() tmpevals[numpy.isnan(tmpevals)] = algentries[ i].maxevals[numpy.isnan(tmpevals)] bestevals = refalgentry.detEvals([targets[j]]) bestevals, bestalgalg = (bestevals[0][0], bestevals[1][0]) bestevals[numpy.isnan( bestevals)] = refalgentry.maxevals[bestalgalg][ numpy.isnan(bestevals)] tmpevals = numpy.array(sorted( tmpevals))[0:min(len(tmpevals), len(bestevals))] bestevals = numpy.array(sorted( bestevals))[0:min(len(tmpevals), len(bestevals))] #The conditions are now that ERT < ERT_best and # all(sorted(FEvals_best) > sorted(FEvals_current)). if numpy.isinf(refalgert[j]) or all(tmpevals < bestevals): nbstars = -numpy.ceil(numpy.log10(nbtests * p)) # tmp2[-1] += r'$^{%s}$' % superscript str_significance_subsup += r'_{%s%s}' % ( significance_vs_ref_symbol, str(int(nbstars)) if nbstars > 1 else '') if str_significance_subsup: str_significance_subsup = '$%s$' % str_significance_subsup # format number in variable data if numpy.isnan(data): curline.append(r'\multicolumn{2}{%s}{.}' % alignment) else: if numpy.isinf(refalgert[j]): curline.append( r'\multicolumn{2}{%s}{\textbf{%s}\mbox{\tiny (%s)}%s}' % (alignment, writeFEvalsMaxPrec(algerts[i][j], 2), writeFEvalsMaxPrec(dispersion, precdispersion), str_significance_subsup)) continue tmp = writeFEvalsMaxPrec(data, precfloat, maxfloatrepr=maxfloatrepr) if data >= maxfloatrepr or data < 0.01: # either inf or scientific notation if numpy.isinf(data) and j == len(algerts[i]) - 1: tmp += r'\,\textit{%s}' % writeFEvalsMaxPrec( algfinaldata[i][1], 0, maxfloatrepr=maxfloatrepr) else: tmp = writeFEvalsMaxPrec(data, precscien, maxfloatrepr=data) if isBold: tmp = r'\textbf{%s}' % tmp if not numpy.isnan(dispersion): tmpdisp = dispersion / refalgert[j] if tmpdisp >= maxfloatrepr or tmpdisp < 0.005: # TODO: hack tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=tmpdisp) else: tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=maxfloatrepr) tmp += r'\mbox{\tiny (%s)}' % tmpdisp curline.append( r'\multicolumn{2}{%s}{%s%s}' % (alignment, tmp, str_significance_subsup)) else: tmp2 = tmp.split('.', 1) if len(tmp2) < 2: tmp2.append('') else: tmp2[-1] = '.' + tmp2[-1] if isBold: tmp3 = [] for k in tmp2: tmp3.append(r'\textbf{%s}' % k) tmp2 = tmp3 if not numpy.isnan(dispersion): tmpdisp = dispersion / refalgert[j] if tmpdisp >= maxfloatrepr or tmpdisp < 0.01: tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=tmpdisp) else: tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=maxfloatrepr) tmp2[-1] += (r'\mbox{\tiny (%s)}' % (tmpdisp)) tmp2[-1] += str_significance_subsup curline.extend(tmp2) curline.append('%d' % algnbsucc[i]) curline.append('/%d' % algnbruns[i]) table.append(curline) extraeol.append('') # Write table res = tableXLaTeX(table, spec=spec, extraeol=extraeol) try: filename = os.path.join( outputdir, 'pptables_f%03d_%02dD.tex' % (df[1], df[0])) f = open(filename, 'w') f.write(header + '\n') f.write(res) if verbose: print 'Wrote table in %s' % filename except: raise else: f.close()
def main(dictAlg, sortedAlgs, outputdir='.', verbose=True, function_targets_line=True): # [1, 13, 101] """Generate one table per func with results of multiple algorithms.""" """Difference with the first version: * numbers aligned using the decimal separator * premices for dispersion measure * significance test against best algorithm * table width... Takes ``targetsOfInterest`` from this file as "input argument" to compute the desired target values. ``targetsOfInterest`` might be configured via config. """ # TODO: method is long, terrible to read, split if possible if not bestalg.bestalgentries2009: bestalg.loadBBOB2009() # Sort data per dimension and function dictData = {} dsListperAlg = list(dictAlg[i] for i in sortedAlgs) for n, entries in enumerate(dsListperAlg): tmpdictdim = entries.dictByDim() for d in tmpdictdim: tmpdictfun = tmpdictdim[d].dictByFunc() for f in tmpdictfun: dictData.setdefault((d, f), {})[n] = tmpdictfun[f] nbtests = len(dictData) for df in dictData: # Generate one table per df # first update targets for each dimension-function pair if needed: targets = targetsOfInterest((df[1], df[0])) targetf = targets[-1] # best 2009 refalgentry = bestalg.bestalgentries2009[df] refalgert = refalgentry.detERT(targets) refalgevals = (refalgentry.detEvals((targetf, ))[0][0]) refalgnbruns = len(refalgevals) refalgnbsucc = numpy.sum(numpy.isnan(refalgevals) == False) # Process the data # The following variables will be lists of elements each corresponding # to an algorithm algnames = [] #algdata = [] algerts = [] algevals = [] algdisp = [] algnbsucc = [] algnbruns = [] algmedmaxevals = [] algmedfinalfunvals = [] algtestres = [] algentries = [] for n in sorted(dictData[df].keys()): entries = dictData[df][n] # the number of datasets for a given dimension and function (df) # should be strictly 1. TODO: find a way to warn # TODO: do this checking before... why wasn't it triggered by ppperprof? if len(entries) > 1: txt = ("There is more than a single entry associated with " "folder %s on %d-D f%d." % (sortedAlgs[n], df[0], df[1])) raise Exception(txt) entry = entries[0] algentries.append(entry) algnames.append(sortedAlgs[n]) evals = entry.detEvals(targets) #tmpdata = [] tmpdisp = [] tmpert = [] for i, e in enumerate(evals): succ = (numpy.isnan(e) == False) ec = e.copy() # note: here was the previous bug (changes made in e also appeared in evals !) ec[succ == False] = entry.maxevals[succ == False] ert = toolsstats.sp(ec, issuccessful=succ)[0] #tmpdata.append(ert/refalgert[i]) if succ.any(): tmp = toolsstats.drawSP(ec[succ], entry.maxevals[succ == False], [10, 50, 90], samplesize=samplesize)[0] tmpdisp.append((tmp[-1] - tmp[0])/2.) else: tmpdisp.append(numpy.nan) tmpert.append(ert) algerts.append(tmpert) algevals.append(evals) #algdata.append(tmpdata) algdisp.append(tmpdisp) algmedmaxevals.append(numpy.median(entry.maxevals)) algmedfinalfunvals.append(numpy.median(entry.finalfunvals)) #algmedmaxevals.append(numpy.median(entry.maxevals)/df[0]) #algmedfinalfunvals.append(numpy.median(entry.finalfunvals)) algtestres.append(significancetest(refalgentry, entry, targets)) # determine success probability for Df = 1e-8 e = entry.detEvals((targetf ,))[0] algnbsucc.append(numpy.sum(numpy.isnan(e) == False)) algnbruns.append(len(e)) # Process over all data # find best values... nalgs = len(dictData[df]) maxRank = 1 + numpy.floor(0.14 * nalgs) # number of algs to be displayed in bold isBoldArray = [] # Point out the best values algfinaldata = [] # Store median function values/median number of function evaluations tmptop = getTopIndicesOfColumns(algerts, maxRank=maxRank) for i, erts in enumerate(algerts): tmp = [] for j, ert in enumerate(erts): # algi targetj tmp.append(i in tmptop[j] or (nalgs > 7 and algerts[i][j] <= 3. * refalgert[j])) isBoldArray.append(tmp) algfinaldata.append((algmedfinalfunvals[i], algmedmaxevals[i])) # significance test of best given algorithm against all others best_alg_idx = numpy.array(algerts).argsort(0)[0, :] # indexed by target index significance_versus_others = significance_all_best_vs_other(algentries, targets, best_alg_idx)[0] # Create the table table = [] spec = r'@{}c@{}|*{%d}{@{\,}r@{}X@{\,}}|@{}r@{}@{}l@{}' % (len(targets)) # in case StrLeft not working: replaced c@{} with l@{ } spec = r'@{}c@{}|*{%d}{@{}r@{}X@{}}|@{}r@{}@{}l@{}' % (len(targets)) # in case StrLeft not working: replaced c@{} with l@{ } extraeol = [] # Generate header lines if with_table_heading: header = funInfos[df[1]] if funInfos else 'f%d' % df[1] table.append([r'\multicolumn{%d}{@{\,}c@{\,}}{{\textbf{%s}}}' % (2 * len(targets) + 2, header)]) extraeol.append('') if function_targets_line is True or (function_targets_line and df[1] in function_targets_line): if isinstance(targetsOfInterest, pproc.RunlengthBasedTargetValues): curline = [r'\#FEs/D'] for i in targetsOfInterest.labels(): curline.append(r'\multicolumn{2}{@{}c@{}}{%s}' % i) else: curline = [r'$\Delta f_\mathrm{opt}$'] for t in targets: curline.append(r'\multicolumn{2}{@{\,}X@{\,}}{%s}' % writeFEvals2(t, precision=1, isscientific=True)) # curline.append(r'\multicolumn{2}{@{\,}X@{}|}{%s}' # % writeFEvals2(targets[-1], precision=1, isscientific=True)) curline.append(r'\multicolumn{2}{@{}l@{}}{\#succ}') table.append(curline) extraeol.append(r'\hline') # extraeol.append(r'\hline\arrayrulecolor{tableShade}') curline = [r'ERT$_{\text{best}}$'] if with_table_heading else [r'\textbf{f%d}' % df[1]] if isinstance(targetsOfInterest, pproc.RunlengthBasedTargetValues): # write ftarget:fevals for i in xrange(len(refalgert[:-1])): temp="%.1e" %targetsOfInterest((df[1], df[0]))[i] if temp[-2]=="0": temp=temp[:-2]+temp[-1] curline.append(r'\multicolumn{2}{@{}c@{}}{\textit{%s}:%s \quad}' % (temp,writeFEvalsMaxPrec(refalgert[i], 2))) temp="%.1e" %targetsOfInterest((df[1], df[0]))[-1] if temp[-2]=="0": temp=temp[:-2]+temp[-1] curline.append(r'\multicolumn{2}{@{}c@{}|}{\textit{%s}:%s }' % (temp,writeFEvalsMaxPrec(refalgert[-1], 2))) else: # write #fevals of the reference alg for i in refalgert[:-1]: curline.append(r'\multicolumn{2}{@{}c@{}}{%s \quad}' % writeFEvalsMaxPrec(i, 2)) curline.append(r'\multicolumn{2}{@{}c@{}|}{%s}' % writeFEvalsMaxPrec(refalgert[-1], 2)) # write the success ratio for the reference alg tmp2 = numpy.sum(numpy.isnan(refalgevals) == False) # count the nb of success curline.append('%d' % (tmp2)) if tmp2 > 0: curline.append('/%d' % len(refalgevals)) table.append(curline[:]) extraeol.append('') #for i, gna in enumerate(zip((1, 2, 3), ('bla', 'blo', 'bli'))): #print i, gna, gno #set_trace() # Format data #if df == (5, 17): #set_trace() header = r'\providecommand{\ntables}{7}' for i, alg in enumerate(algnames): #algname, entries, irs, line, line2, succ, runs, testres1alg in zip(algnames, #data, dispersion, isBoldArray, isItalArray, nbsucc, nbruns, testres): commandname = r'\alg%stables' % numtotext(i) # header += r'\providecommand{%s}{{%s}{}}' % (commandname, str_to_latex(strip_pathname(alg))) header += r'\providecommand{%s}{\StrLeft{%s}{\ntables}}' % (commandname, str_to_latex(strip_pathname2(alg))) curline = [commandname + r'\hspace*{\fill}'] # each list element becomes a &-separated table entry? for j, tmp in enumerate(zip(algerts[i], algdisp[i], # j is target index isBoldArray[i], algtestres[i])): ert, dispersion, isBold, testres = tmp alignment = '@{\,}X@{\,}' if j == len(algerts[i]) - 1: alignment = '@{\,}X@{\,}|' data = ert/refalgert[j] # write star for significance against all other algorithms str_significance_subsup = '' if (len(best_alg_idx) > 0 and len(significance_versus_others) > 0 and i == best_alg_idx[j] and nbtests * significance_versus_others[j][1] < 0.05): logp = -numpy.ceil(numpy.log10(nbtests * significance_versus_others[j][1])) str_significance_subsup = r"^{%s%s}" % (significance_vs_others_symbol, str(int(logp)) if logp > 1 else '') # moved out of the above else: this was a bug!? z, p = testres if (nbtests * p) < 0.05 and data < 1. and z < 0.: if not numpy.isinf(refalgert[j]): tmpevals = algevals[i][j].copy() tmpevals[numpy.isnan(tmpevals)] = algentries[i].maxevals[numpy.isnan(tmpevals)] bestevals = refalgentry.detEvals(targets) bestevals, bestalgalg = (bestevals[0][0], bestevals[1][0]) bestevals[numpy.isnan(bestevals)] = refalgentry.maxevals[bestalgalg][numpy.isnan(bestevals)] tmpevals = numpy.array(sorted(tmpevals))[0:min(len(tmpevals), len(bestevals))] bestevals = numpy.array(sorted(bestevals))[0:min(len(tmpevals), len(bestevals))] #The conditions are now that ERT < ERT_best and # all(sorted(FEvals_best) > sorted(FEvals_current)). if numpy.isinf(refalgert[j]) or all(tmpevals < bestevals): nbstars = -numpy.ceil(numpy.log10(nbtests * p)) # tmp2[-1] += r'$^{%s}$' % superscript str_significance_subsup += r'_{%s%s}' % (significance_vs_ref_symbol, str(int(nbstars)) if nbstars > 1 else '') if str_significance_subsup: str_significance_subsup = '$%s$' % str_significance_subsup # format number in variable data if numpy.isnan(data): curline.append(r'\multicolumn{2}{%s}{.}' % alignment) else: if numpy.isinf(refalgert[j]): curline.append(r'\multicolumn{2}{%s}{\textbf{%s}\mbox{\tiny (%s)}%s}' % (alignment, writeFEvalsMaxPrec(algerts[i][j], 2), writeFEvalsMaxPrec(dispersion, precdispersion), str_significance_subsup)) continue tmp = writeFEvalsMaxPrec(data, precfloat, maxfloatrepr=maxfloatrepr) if data >= maxfloatrepr or data < 0.01: # either inf or scientific notation if numpy.isinf(data) and j == len(algerts[i]) - 1: tmp += r'\,\textit{%s}' % writeFEvalsMaxPrec(algfinaldata[i][1], 0, maxfloatrepr=maxfloatrepr) else: tmp = writeFEvalsMaxPrec(data, precscien, maxfloatrepr=data) if isBold: tmp = r'\textbf{%s}' % tmp if not numpy.isnan(dispersion): tmpdisp = dispersion/refalgert[j] if tmpdisp >= maxfloatrepr or tmpdisp < 0.005: # TODO: hack tmpdisp = writeFEvalsMaxPrec(tmpdisp, precdispersion, maxfloatrepr=tmpdisp) else: tmpdisp = writeFEvalsMaxPrec(tmpdisp, precdispersion, maxfloatrepr=maxfloatrepr) tmp += r'\mbox{\tiny (%s)}' % tmpdisp curline.append(r'\multicolumn{2}{%s}{%s%s}' % (alignment, tmp, str_significance_subsup)) else: tmp2 = tmp.split('.', 1) if len(tmp2) < 2: tmp2.append('') else: tmp2[-1] = '.' + tmp2[-1] if isBold: tmp3 = [] for k in tmp2: tmp3.append(r'\textbf{%s}' % k) tmp2 = tmp3 if not numpy.isnan(dispersion): tmpdisp = dispersion/refalgert[j] if tmpdisp >= maxfloatrepr or tmpdisp < 0.01: tmpdisp = writeFEvalsMaxPrec(tmpdisp, precdispersion, maxfloatrepr=tmpdisp) else: tmpdisp = writeFEvalsMaxPrec(tmpdisp, precdispersion, maxfloatrepr=maxfloatrepr) tmp2[-1] += (r'\mbox{\tiny (%s)}' % (tmpdisp)) tmp2[-1] += str_significance_subsup curline.extend(tmp2) curline.append('%d' % algnbsucc[i]) curline.append('/%d' % algnbruns[i]) table.append(curline) extraeol.append('') # Write table res = tableXLaTeX(table, spec=spec, extraeol=extraeol) try: filename = os.path.join(outputdir, 'pptables_f%03d_%02dD.tex' % (df[1], df[0])) f = open(filename, 'w') f.write(header + '\n') f.write(res) if verbose: print 'Wrote table in %s' % filename except: raise else: f.close()
def main(dictAlg, sortedAlgs, outputdir='.', verbose=True, function_targets_line=True): # [1, 13, 101] """Generate one table per func with results of multiple algorithms.""" """Difference with the first version: * numbers aligned using the decimal separator * premices for dispersion measure * significance test against best algorithm * table width... Takes ``targetsOfInterest`` from this file as "input argument" to compute the desired target values. ``targetsOfInterest`` might be configured via config. """ # TODO: method is long, terrible to read, split if possible if not bestalg.bestalgentries2009: bestalg.loadBBOB2009() # Sort data per dimension and function dictData = {} dsListperAlg = list(dictAlg[i] for i in sortedAlgs) for n, entries in enumerate(dsListperAlg): tmpdictdim = entries.dictByDim() for d in tmpdictdim: tmpdictfun = tmpdictdim[d].dictByFunc() for f in tmpdictfun: dictData.setdefault((d, f), {})[n] = tmpdictfun[f] nbtests = len(dictData) for df in dictData: # Generate one table per df # first update targets for each dimension-function pair if needed: targets = targetsOfInterest((df[1], df[0])) targetf = targets[-1] # best 2009 refalgentry = bestalg.bestalgentries2009[df] refalgert = refalgentry.detERT(targets) refalgevals = (refalgentry.detEvals((targetf, ))[0][0]) refalgnbruns = len(refalgevals) refalgnbsucc = numpy.sum(numpy.isnan(refalgevals) == False) # Process the data # The following variables will be lists of elements each corresponding # to an algorithm algnames = [] #algdata = [] algerts = [] algevals = [] algdisp = [] algnbsucc = [] algnbruns = [] algmedmaxevals = [] algmedfinalfunvals = [] algtestres = [] algentries = [] for n in sorted(dictData[df].keys()): entries = dictData[df][n] # the number of datasets for a given dimension and function (df) # should be strictly 1. TODO: find a way to warn # TODO: do this checking before... why wasn't it triggered by ppperprof? if len(entries) > 1: print entries txt = ("There is more than a single entry associated with " "folder %s on %d-D f%d." % (sortedAlgs[n], df[0], df[1])) raise Exception(txt) entry = entries[0] algentries.append(entry) algnames.append(sortedAlgs[n]) evals = entry.detEvals(targets) #tmpdata = [] tmpdisp = [] tmpert = [] for i, e in enumerate(evals): succ = (numpy.isnan(e) == False) ec = e.copy( ) # note: here was the previous bug (changes made in e also appeared in evals !) ec[succ == False] = entry.maxevals[succ == False] ert = toolsstats.sp(ec, issuccessful=succ)[0] #tmpdata.append(ert/refalgert[i]) if succ.any(): tmp = toolsstats.drawSP(ec[succ], entry.maxevals[succ == False], [10, 50, 90], samplesize=samplesize)[0] tmpdisp.append((tmp[-1] - tmp[0]) / 2.) else: tmpdisp.append(numpy.nan) tmpert.append(ert) algerts.append(tmpert) algevals.append(evals) #algdata.append(tmpdata) algdisp.append(tmpdisp) algmedmaxevals.append(numpy.median(entry.maxevals)) algmedfinalfunvals.append(numpy.median(entry.finalfunvals)) #algmedmaxevals.append(numpy.median(entry.maxevals)/df[0]) #algmedfinalfunvals.append(numpy.median(entry.finalfunvals)) algtestres.append(significancetest(refalgentry, entry, targets)) # determine success probability for Df = 1e-8 e = entry.detEvals((targetf, ))[0] algnbsucc.append(numpy.sum(numpy.isnan(e) == False)) algnbruns.append(len(e)) # Process over all data # find best values... nalgs = len(dictData[df]) maxRank = 1 + numpy.floor( 0.14 * nalgs) # number of algs to be displayed in bold isBoldArray = [] # Point out the best values algfinaldata = [ ] # Store median function values/median number of function evaluations tmptop = getTopIndicesOfColumns(algerts, maxRank=maxRank) for i, erts in enumerate(algerts): tmp = [] for j, ert in enumerate(erts): # algi targetj tmp.append(i in tmptop[j] or (nalgs > 7 and algerts[i][j] <= 3. * refalgert[j])) isBoldArray.append(tmp) algfinaldata.append((algmedfinalfunvals[i], algmedmaxevals[i])) # significance test of best given algorithm against all others best_alg_idx = numpy.array(algerts).argsort(0)[ 0, :] # indexed by target index significance_versus_others = significance_all_best_vs_other( algentries, targets, best_alg_idx)[0] # Create the table table = [] tableHtml = [] spec = r'@{}c@{}|*{%d}{@{\,}r@{}X@{\,}}|@{}r@{}@{}l@{}' % ( len(targets) ) # in case StrLeft not working: replaced c@{} with l@{ } spec = r'@{}c@{}|*{%d}{@{}r@{}X@{}}|@{}r@{}@{}l@{}' % ( len(targets) ) # in case StrLeft not working: replaced c@{} with l@{ } extraeol = [] # Generate header lines if with_table_heading: header = funInfos[df[1]] if funInfos else 'f%d' % df[1] table.append([ r'\multicolumn{%d}{@{\,}c@{\,}}{{\textbf{%s}}}' % (2 * len(targets) + 2, header) ]) extraeol.append('') if function_targets_line is True or (function_targets_line and df[1] in function_targets_line): if isinstance(targetsOfInterest, pproc.RunlengthBasedTargetValues): curline = [r'\#FEs/D'] curlineHtml = ['<thead>\n<tr>\n<th>#FEs/D<br>REPLACEH</th>\n'] counter = 1 for i in targetsOfInterest.labels(): curline.append(r'\multicolumn{2}{@{}c@{}}{%s}' % i) curlineHtml.append('<td>%s<br>REPLACE%d</td>\n' % (i, counter)) counter += 1 else: curline = [r'$\Delta f_\mathrm{opt}$'] curlineHtml = [ '<thead>\n<tr>\n<th>Δ f<sub>opt</sub><br>REPLACEH</th>\n' ] counter = 1 for t in targets: curline.append( r'\multicolumn{2}{@{\,}X@{\,}}{%s}' % writeFEvals2(t, precision=1, isscientific=True)) curlineHtml.append( '<td>%s<br>REPLACE%d</td>\n' % (writeFEvals2( t, precision=1, isscientific=True), counter)) counter += 1 # curline.append(r'\multicolumn{2}{@{\,}X@{}|}{%s}' # % writeFEvals2(targets[-1], precision=1, isscientific=True)) curline.append(r'\multicolumn{2}{@{}l@{}}{\#succ}') curlineHtml.append('<td>#succ<br>REPLACEF</td>\n</tr>\n</thead>\n') table.append(curline) extraeol.append(r'\hline') # extraeol.append(r'\hline\arrayrulecolor{tableShade}') curline = [r'ERT$_{\text{best}}$' ] if with_table_heading else [r'\textbf{f%d}' % df[1]] replaceValue = 'ERT<sub>best</sub>' if with_table_heading else ( '<b>f%d</b>' % df[1]) curlineHtml = [ item.replace('REPLACEH', replaceValue) for item in curlineHtml ] if isinstance(targetsOfInterest, pproc.RunlengthBasedTargetValues): # write ftarget:fevals counter = 1 for i in xrange(len(refalgert[:-1])): temp = "%.1e" % targetsOfInterest((df[1], df[0]))[i] if temp[-2] == "0": temp = temp[:-2] + temp[-1] curline.append( r'\multicolumn{2}{@{}c@{}}{\textit{%s}:%s \quad}' % (temp, writeFEvalsMaxPrec(refalgert[i], 2))) replaceValue = '<i>%s</i>:%s' % ( temp, writeFEvalsMaxPrec(refalgert[i], 2)) curlineHtml = [ item.replace('REPLACE%d' % counter, replaceValue) for item in curlineHtml ] counter += 1 temp = "%.1e" % targetsOfInterest((df[1], df[0]))[-1] if temp[-2] == "0": temp = temp[:-2] + temp[-1] curline.append(r'\multicolumn{2}{@{}c@{}|}{\textit{%s}:%s }' % (temp, writeFEvalsMaxPrec(refalgert[-1], 2))) replaceValue = '<i>%s</i>:%s' % ( temp, writeFEvalsMaxPrec(refalgert[-1], 2)) curlineHtml = [ item.replace('REPLACE%d' % counter, replaceValue) for item in curlineHtml ] else: # write #fevals of the reference alg counter = 1 for i in refalgert[:-1]: curline.append(r'\multicolumn{2}{@{}c@{}}{%s \quad}' % writeFEvalsMaxPrec(i, 2)) curlineHtml = [ item.replace('REPLACE%d' % counter, writeFEvalsMaxPrec(i, 2)) for item in curlineHtml ] counter += 1 curline.append(r'\multicolumn{2}{@{}c@{}|}{%s}' % writeFEvalsMaxPrec(refalgert[-1], 2)) curlineHtml = [ item.replace('REPLACE%d' % counter, writeFEvalsMaxPrec(refalgert[-1], 2)) for item in curlineHtml ] # write the success ratio for the reference alg tmp2 = numpy.sum( numpy.isnan(refalgevals) == False) # count the nb of success curline.append('%d' % (tmp2)) if tmp2 > 0: curline.append('/%d' % len(refalgevals)) replaceValue = '%d/%d' % (tmp2, len(refalgevals)) else: replaceValue = '%d' % tmp2 curlineHtml = [ item.replace('REPLACEF', replaceValue) for item in curlineHtml ] table.append(curline[:]) tableHtml.extend(curlineHtml[:]) tableHtml.append('<tbody>\n') extraeol.append('') #for i, gna in enumerate(zip((1, 2, 3), ('bla', 'blo', 'bli'))): #print i, gna, gno #set_trace() # Format data #if df == (5, 17): #set_trace() header = r'\providecommand{\ntables}{7}' for i, alg in enumerate(algnames): tableHtml.append('<tr>\n') #algname, entries, irs, line, line2, succ, runs, testres1alg in zip(algnames, #data, dispersion, isBoldArray, isItalArray, nbsucc, nbruns, testres): commandname = r'\alg%stables' % numtotext(i) # header += r'\providecommand{%s}{{%s}{}}' % (commandname, str_to_latex(strip_pathname(alg))) header += r'\providecommand{%s}{\StrLeft{%s}{\ntables}}' % ( commandname, str_to_latex(strip_pathname1(alg))) curline = [ commandname + r'\hspace*{\fill}' ] # each list element becomes a &-separated table entry? curlineHtml = [ '<th>%s</th>\n' % str_to_latex(strip_pathname1(alg)) ] for j, tmp in enumerate( zip( algerts[i], algdisp[i], # j is target index isBoldArray[i], algtestres[i])): ert, dispersion, isBold, testres = tmp alignment = '@{\,}X@{\,}' if j == len(algerts[i]) - 1: alignment = '@{\,}X@{\,}|' data = ert / refalgert[j] # write star for significance against all other algorithms str_significance_subsup = '' str_significance_subsup_html = '' if (len(best_alg_idx) > 0 and len(significance_versus_others) > 0 and i == best_alg_idx[j] and nbtests * significance_versus_others[j][1] < 0.05): logp = -numpy.ceil( numpy.log10( nbtests * significance_versus_others[j][1])) logp = numpy.min( (9, logp)) # not messing up the format and handling inf str_significance_subsup = r"^{%s%s}" % ( significance_vs_others_symbol, str(int(logp)) if logp > 1 else '') str_significance_subsup_html = '<sup>%s%s</sup>' % ( significance_vs_others_symbol_html, str(int(logp)) if logp > 1 else '') # moved out of the above else: this was a bug!? z, p = testres if (nbtests * p) < 0.05 and data < 1. and z < 0.: if not numpy.isinf(refalgert[j]): tmpevals = algevals[i][j].copy() tmpevals[numpy.isnan(tmpevals)] = algentries[ i].maxevals[numpy.isnan(tmpevals)] bestevals = refalgentry.detEvals(targets) bestevals, bestalgalg = (bestevals[0][0], bestevals[1][0]) bestevals[numpy.isnan( bestevals)] = refalgentry.maxevals[bestalgalg][ numpy.isnan(bestevals)] tmpevals = numpy.array(sorted( tmpevals))[0:min(len(tmpevals), len(bestevals))] bestevals = numpy.array(sorted( bestevals))[0:min(len(tmpevals), len(bestevals))] #The conditions are now that ERT < ERT_best and # all(sorted(FEvals_best) > sorted(FEvals_current)). if numpy.isinf(refalgert[j]) or all(tmpevals < bestevals): nbstars = -numpy.ceil(numpy.log10(nbtests * p)) # tmp2[-1] += r'$^{%s}$' % superscript str_significance_subsup += r'_{%s%s}' % ( significance_vs_ref_symbol, str(int(nbstars)) if nbstars > 1 else '') str_significance_subsup_html = '<sub>%s%s</sub>' % ( significance_vs_ref_symbol_html, str(int(nbstars)) if nbstars > 1 else '') if str_significance_subsup: str_significance_subsup = '$%s$' % str_significance_subsup # format number in variable data if numpy.isnan(data): curline.append(r'\multicolumn{2}{%s}{.}' % alignment) else: if numpy.isinf(refalgert[j]): curline.append( r'\multicolumn{2}{%s}{\textbf{%s}\mbox{\tiny (%s)}%s}' % (alignment, writeFEvalsMaxPrec(algerts[i][j], 2), writeFEvalsMaxPrec(dispersion, precdispersion), str_significance_subsup)) curlineHtml.append( '<td sorttable_customkey=\"%f\"><b>%s</b> (%s)%s</td>\n' % (algerts[i][j], writeFEvalsMaxPrec(algerts[i][j], 2), writeFEvalsMaxPrec(dispersion, precdispersion), str_significance_subsup_html)) continue tmp = writeFEvalsMaxPrec(data, precfloat, maxfloatrepr=maxfloatrepr) tmpHtml = writeFEvalsMaxPrec(data, precfloat, maxfloatrepr=maxfloatrepr) sortKey = data if data >= maxfloatrepr or data < 0.01: # either inf or scientific notation if numpy.isinf(data) and j == len(algerts[i]) - 1: tmp += r'\,\textit{%s}' % writeFEvalsMaxPrec( algfinaldata[i][1], 0, maxfloatrepr=maxfloatrepr) tmpHtml += '<i>%s</i>' % writeFEvalsMaxPrec( algfinaldata[i][1], 0, maxfloatrepr=maxfloatrepr) sortKey = algfinaldata[i][1] else: tmp = writeFEvalsMaxPrec(data, precscien, maxfloatrepr=data) if isBold: tmpHtml = '<b>%s</b>' % tmp tmp = r'\textbf{%s}' % tmp if not numpy.isnan(dispersion): tmpdisp = dispersion / refalgert[j] if tmpdisp >= maxfloatrepr or tmpdisp < 0.005: # TODO: hack tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=tmpdisp) else: tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=maxfloatrepr) tmp += r'\mbox{\tiny (%s)}' % tmpdisp tmpHtml += ' (%s)' % tmpdisp curline.append( r'\multicolumn{2}{%s}{%s%s}' % (alignment, tmp, str_significance_subsup)) tmpHtml = tmpHtml.replace('$\infty$', '∞') if (numpy.isinf(sortKey)): sortKey = sys.maxint curlineHtml.append( '<td sorttable_customkey=\"%f\">%s%s</td>' % (sortKey, tmpHtml, str_significance_subsup_html)) else: tmp2 = tmp.split('.', 1) if len(tmp2) < 2: tmp2.append('') else: tmp2[-1] = '.' + tmp2[-1] if isBold: tmp3 = [] tmp3html = [] for k in tmp2: tmp3.append(r'\textbf{%s}' % k) tmp3html.append('<b>%s</b>' % k) tmp2 = tmp3 tmp2html = tmp3html else: tmp2html = [] tmp2html.extend(tmp2) if not numpy.isnan(dispersion): tmpdisp = dispersion / refalgert[j] if tmpdisp >= maxfloatrepr or tmpdisp < 0.01: tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=tmpdisp) else: tmpdisp = writeFEvalsMaxPrec( tmpdisp, precdispersion, maxfloatrepr=maxfloatrepr) tmp2[-1] += (r'\mbox{\tiny (%s)}' % (tmpdisp)) tmp2html[-1] += ' (%s)' % tmpdisp tmp2[-1] += str_significance_subsup tmp2html[-1] += str_significance_subsup_html curline.extend(tmp2) tmp2html = ("").join(str(item) for item in tmp2html) tmp2html = tmp2html.replace('$\infty$', '∞') curlineHtml.append( '<td sorttable_customkey=\"%f\">%s</td>' % (data, tmp2html)) curline.append('%d' % algnbsucc[i]) curline.append('/%d' % algnbruns[i]) table.append(curline) curlineHtml.append('<td sorttable_customkey=\"%d\">%d/%d</td>\n' % (algnbsucc[i], algnbsucc[i], algnbruns[i])) tableHtml.extend(curlineHtml[:]) extraeol.append('') # Write table res = tableXLaTeX(table, spec=spec, extraeol=extraeol) try: filename = os.path.join( outputdir, 'pptables_f%03d_%02dD.tex' % (df[1], df[0])) f = open(filename, 'w') f.write(header + '\n') f.write(res) res = ("").join(str(item) for item in tableHtml) res = '\n<table class=\"sortable\" style=\"width:800px \">\n%s</table>\n<p/>\n' % res if df[0] in (5, 20): filename = os.path.join( outputdir, genericsettings.many_algorithm_file_name + '.html') lines = [] with open(filename) as infile: for line in infile: if '<!--' + 'pptablesf%03d%02dDHtml' % ( df[1], df[0]) + '-->' in line: lines.append(res) lines.append(line) with open(filename, 'w') as outfile: for line in lines: outfile.write(line) if verbose: print 'Wrote table in %s' % filename except: raise else: f.close()