def build(dictAlg, sortedAlg=None):
"""Merge datasets in an algorithm portfolio.
:param dict dictAlg: dictionary of data sets with algorithm name for
keys
:param seq sortedAlgs: sequence for sorting the entries of
:py:data:`dictAlg`, if not provided,
dictAlg.keys() will be instead
:returns: an instance of :py:class:`DataSetList` with the porfolio
data sets
"""
if not sortedAlg:
sortedAlg = dictAlg.keys()
tmpres = []
for f, i in pp.dictAlgByFun(dictAlg).iteritems():
for d, j in pp.dictAlgByDim(i).iteritems():
tmp = []
if sortedAlg:
tmplist = list(j[k] for k in sortedAlg)
else:
tmplist = j.values()
for k in tmplist:
assert len(k) == 1 # one element list
tmp.append(k[0])
try:
tmpres.append(DataSet(tmp))
except Usage, err:
print >>sys.stderr, err.msg
def generate(dictalg):
"""Generates dictionary of best algorithm data set.
"""
#dsList, sortedAlgs, dictAlg = processInputArgs(args, verbose=verbose)
res = {}
for f, i in pproc.dictAlgByFun(dictalg).iteritems():
for d, j in pproc.dictAlgByDim(i).iteritems():
tmp = BestAlgSet(j)
res[(d, f)] = tmp
return res
def generate(dictalg):
"""Generates dictionary of best algorithm data set.
"""
# dsList, sortedAlgs, dictAlg = processInputArgs(args, verbose=verbose)
res = {}
for f, i in pproc.dictAlgByFun(dictalg).iteritems():
for d, j in pproc.dictAlgByDim(i).iteritems():
tmp = BestAlgSet(j)
res[(d, f)] = tmp
return res
def main(argv=None):
"""
"""
if argv is None:
argv = sys.argv[1:]
try:
try:
opts, args = getopt.getopt(argv, "h", ["help"])
except getopt.error, msg:
raise Usage(msg)
args = algs
# if not (args):
# usage()
# sys.exit()
outputdir = 'bestAlg'
verbose = True
#Process options
for o, a in opts:
if o in ("-h", "--help"):
usage()
sys.exit()
else:
assert False, "unhandled option"
dsList, sortedAlgs, dictAlg = processInputArgs(args, verbose=verbose)
#set_trace()
if not os.path.exists(outputdir):
os.mkdir(outputdir)
if verbose:
print 'Folder %s was created.' % (outputdir)
res = {}
for f, i in dictAlgByFun(dictAlg).iteritems():
for d, j in dictAlgByDim(i).iteritems():
tmp = BestAlgSet(j)
picklefilename = os.path.join(
outputdir, 'bestalg_f%03d_%02d.pickle' % (f, d))
fid = open(picklefilename, 'w')
pickle.dump(tmp, fid, 2)
fid.close()
res[(d, f)] = tmp
picklefilename = os.path.join(outputdir, 'bestalg.pickle')
fid = open(picklefilename, 'w')
pickle.dump(res, fid, 2)
fid.close()
def all_single_functions(dictAlg, sortedAlgs=None, outputdir='.', verbose=0):
dictFG = pp.dictAlgByFun(dictAlg)
for fg, tmpdictAlg in dictFG.iteritems():
dictDim = pp.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
single_fct_output_dir = (
outputdir.rstrip(os.sep) + os.sep +
'pprldmany-single-functions'
# + os.sep + ('f%03d' % fg)
)
if not os.path.exists(single_fct_output_dir):
os.makedirs(single_fct_output_dir)
main(entries,
order=sortedAlgs,
outputdir=single_fct_output_dir,
info=('f%03d_%02dD' % (fg, d)),
verbose=verbose)
def all_single_functions(dictAlg, sortedAlgs=None, outputdir='.',
verbose=0):
dictFG = pp.dictAlgByFun(dictAlg)
for fg, tmpdictAlg in dictFG.iteritems():
dictDim = pp.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
single_fct_output_dir = (outputdir.rstrip(os.sep) + os.sep +
'pprldmany-single-functions'
# + os.sep + ('f%03d' % fg)
)
if not os.path.exists(single_fct_output_dir):
os.makedirs(single_fct_output_dir)
main(entries,
order=sortedAlgs,
outputdir=single_fct_output_dir,
info=('f%03d_%02dD' % (fg, d)),
verbose=verbose)
def main(dictAlg,
order=None,
outputdir='.',
info='default',
dimension=None,
verbose=True):
"""Generates a figure showing the performance of algorithms.
From a dictionary of :py:class:`DataSetList` sorted by algorithms,
generates the cumulative distribution function of the bootstrap
distribution of ERT for algorithms on multiple functions for
multiple targets altogether.
:param dict dictAlg: dictionary of :py:class:`DataSetList` instances
one instance is equivalent to one algorithm,
:param list targets: target function values
:param list order: sorted list of keys to dictAlg for plotting order
:param str outputdir: output directory
:param str info: output file name suffix
:param bool verbose: controls verbosity
"""
global x_limit # late assignment of default, because it can be set to None in config
global divide_by_dimension # not fully implemented/tested yet
if 'x_limit' not in globals() or x_limit is None:
x_limit = x_limit_default
tmp = pp.dictAlgByDim(dictAlg)
# tmp = pp.DictAlg(dictAlg).by_dim()
if len(tmp) != 1 and dimension is None:
raise ValueError('We never integrate over dimension.')
if dimension is not None:
if dimension not in tmp.keys():
raise ValueError('dimension %d not in dictAlg dimensions %s' %
(dimension, str(tmp.keys())))
tmp = {dimension: tmp[dimension]}
dim = tmp.keys()[0]
divisor = dim if divide_by_dimension else 1
algorithms_with_data = [a for a in dictAlg.keys() if dictAlg[a] != []]
dictFunc = pp.dictAlgByFun(dictAlg)
# Collect data
# Crafting effort correction: should we consider any?
CrEperAlg = {}
for alg in algorithms_with_data:
CrE = 0.
if 1 < 3 and dictAlg[alg][0].algId == 'GLOBAL':
tmp = dictAlg[alg].dictByNoise()
assert len(tmp.keys()) == 1
if tmp.keys()[0] == 'noiselessall':
CrE = 0.5117
elif tmp.keys()[0] == 'nzall':
CrE = 0.6572
CrEperAlg[alg] = CrE
if CrE != 0.0:
print 'Crafting effort for', alg, 'is', CrE
dictData = {} # list of (ert per function) per algorithm
dictMaxEvals = {} # list of (maxevals per function) per algorithm
bestERT = [] # best ert per function
# funcsolved = [set()] * len(targets) # number of functions solved per target
xbest2009 = []
maxevalsbest2009 = []
for f, dictAlgperFunc in dictFunc.iteritems():
if function_IDs and f not in function_IDs:
continue
# print target_values((f, dim))
for j, t in enumerate(target_values((f, dim))):
# for j, t in enumerate(genericsettings.current_testbed.ecdf_target_values(1e2, f)):
# funcsolved[j].add(f)
for alg in algorithms_with_data:
x = [np.inf] * perfprofsamplesize
runlengthunsucc = []
try:
entry = dictAlgperFunc[alg][
0] # one element per fun and per dim.
evals = entry.detEvals([t])[0]
assert entry.dim == dim
runlengthsucc = evals[np.isnan(evals) == False] / divisor
runlengthunsucc = entry.maxevals[np.isnan(evals)] / divisor
if len(runlengthsucc) > 0:
x = toolsstats.drawSP(runlengthsucc,
runlengthunsucc,
percentiles=[50],
samplesize=perfprofsamplesize)[1]
except (KeyError, IndexError):
#set_trace()
warntxt = (
'Data for algorithm %s on function %d in %d-D ' %
(alg, f, dim) + 'are missing.\n')
warnings.warn(warntxt)
dictData.setdefault(alg, []).extend(x)
dictMaxEvals.setdefault(alg, []).extend(runlengthunsucc)
if displaybest2009:
#set_trace()
if not bestalg.bestalgentries2009:
bestalg.loadBBOB2009()
bestalgentry = bestalg.bestalgentries2009[(dim, f)]
bestalgevals = bestalgentry.detEvals(target_values((f, dim)))
# print bestalgevals
for j in range(len(bestalgevals[0])):
if bestalgevals[1][j]:
evals = bestalgevals[0][j]
#set_trace()
assert dim == bestalgentry.dim
runlengthsucc = evals[np.isnan(evals) == False] / divisor
runlengthunsucc = bestalgentry.maxevals[
bestalgevals[1][j]][np.isnan(evals)] / divisor
x = toolsstats.drawSP(runlengthsucc,
runlengthunsucc,
percentiles=[50],
samplesize=perfprofsamplesize)[1]
else:
x = perfprofsamplesize * [np.inf]
runlengthunsucc = []
xbest2009.extend(x)
maxevalsbest2009.extend(runlengthunsucc)
if order is None:
order = dictData.keys()
# Display data
lines = []
if displaybest2009:
args = {
'ls': '-',
'linewidth': 6,
'marker': 'D',
'markersize': 11.,
'markeredgewidth': 1.5,
'markerfacecolor': refcolor,
'markeredgecolor': refcolor,
'color': refcolor,
'label': 'best 2009',
'zorder': -1
}
lines.append(
plotdata(np.array(xbest2009),
x_limit,
maxevalsbest2009,
CrE=0.,
**args))
def algname_to_label(algname, dirname=None):
"""to be extended to become generally useful"""
if isinstance(algname, (tuple, list)): # not sure this is needed
return ' '.join([str(name) for name in algname])
return str(algname)
for i, alg in enumerate(order):
try:
data = dictData[alg]
maxevals = dictMaxEvals[alg]
except KeyError:
continue
args = styles[(i) % len(styles)]
args['linewidth'] = 1.5
args['markersize'] = 12.
args['markeredgewidth'] = 1.5
args['markerfacecolor'] = 'None'
args['markeredgecolor'] = args['color']
args['label'] = algname_to_label(alg)
#args['markevery'] = perfprofsamplesize # option available in latest version of matplotlib
#elif len(show_algorithms) > 0:
#args['color'] = 'wheat'
#args['ls'] = '-'
#args['zorder'] = -1
# plotdata calls pprldistr.plotECDF which calls ppfig.plotUnifLog... which does the work
lines.append(
plotdata(np.array(data),
x_limit,
maxevals,
CrE=CrEperAlg[alg],
**args))
labels, handles = plotLegend(lines, x_limit)
if True: # isLateXLeg:
fileName = os.path.join(outputdir, 'pprldmany_%s.tex' % (info))
with open(fileName, 'w') as f:
f.write(r'\providecommand{\nperfprof}{7}')
algtocommand = {} # latex commands
for i, alg in enumerate(order):
tmp = r'\alg%sperfprof' % pptex.numtotext(i)
f.write(
r'\providecommand{%s}{\StrLeft{%s}{\nperfprof}}' %
(tmp,
toolsdivers.str_to_latex(
toolsdivers.strip_pathname2(algname_to_label(alg)))))
algtocommand[algname_to_label(alg)] = tmp
if displaybest2009:
tmp = r'\algzeroperfprof'
f.write(r'\providecommand{%s}{best 2009}' % (tmp))
algtocommand['best 2009'] = tmp
commandnames = []
for label in labels:
commandnames.append(algtocommand[label])
# f.write(headleg)
if len(
order
) > 28: # latex sidepanel won't work well for more than 25 algorithms, but original labels are also clipped
f.write(
r'\providecommand{\perfprofsidepanel}{\mbox{%s}\vfill\mbox{%s}}'
% (commandnames[0], commandnames[-1]))
else:
fontsize_command = r'\tiny{}' if len(order) > 19 else ''
f.write(r'\providecommand{\perfprofsidepanel}{{%s\mbox{%s}' %
(fontsize_command,
commandnames[0])) # TODO: check len(labels) > 0
for i in range(1, len(labels)):
f.write('\n' + r'\vfill \mbox{%s}' % commandnames[i])
f.write('}}\n')
# f.write(footleg)
if verbose:
print 'Wrote right-hand legend in %s' % fileName
figureName = os.path.join(outputdir, 'pprldmany_%s' % (info))
#beautify(figureName, funcsolved, x_limit*x_annote_factor, False, fileFormat=figformat)
beautify()
text = 'f%s' % (ppfig.consecutiveNumbers(sorted(dictFunc.keys())))
text += ',%d-D' % dim # TODO: this is strange when different dimensions are plotted
plt.text(0.01,
0.98,
text,
horizontalalignment="left",
verticalalignment="top",
transform=plt.gca().transAxes)
if len(dictFunc) == 1:
plt.title(' '.join(
(str(dictFunc.keys()[0]),
genericsettings.current_testbed.short_names[dictFunc.keys()[0]])))
a = plt.gca()
plt.xlim(xmin=1e-0, xmax=x_limit**annotation_space_end_relative)
xticks, labels = plt.xticks()
tmp = []
for i in xticks:
tmp.append('%d' % round(np.log10(i)))
a.set_xticklabels(tmp)
if save_figure:
ppfig.saveFigure(figureName, verbose=verbose)
if len(dictFunc) == 1:
ppfig.save_single_functions_html(
os.path.join(outputdir, 'pprldmany'),
'', # algorithms names are clearly visible in the figure
add_to_names='_%02dD' % (dim),
algorithmCount=ppfig.AlgorithmCount.NON_SPECIFIED)
if close_figure:
plt.close()
plt.rc("axes", **inset.rcaxes)
plt.rc("xtick", **inset.rctick)
plt.rc("ytick", **inset.rctick)
plt.rc("font", **inset.rcfont)
plt.rc("legend", **inset.rclegend)
#convergence plots
if isConv:
ppconverrorbars.main(dictAlg,outputdir,verbose)
# Performance profiles
if isPer:
config.config()
# ECDFs per noise groups
dictNoi = pproc.dictAlgByNoi(dictAlg)
for ng, tmpdictAlg in dictNoi.iteritems():
dictDim = pproc.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
# pprldmany.main(entries, inset.summarized_target_function_values,
# from . import config
# config.config()
pprldmany.main(entries, # pass expensive flag here?
order=sortedAlgs,
outputdir=outputdir,
info=('%02dD_%s' % (d, ng)),
verbose=verbose)
# ECDFs per function groups
dictFG = pproc.dictAlgByFuncGroup(dictAlg)
for fg, tmpdictAlg in dictFG.iteritems():
dictDim = pproc.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
pprldmany.main(entries,
plt.rc("axes", **inset.rcaxes)
plt.rc("xtick", **inset.rctick)
plt.rc("ytick", **inset.rctick)
plt.rc("font", **inset.rcfont)
plt.rc("legend", **inset.rclegend)
#convergence plots
if isConv:
ppconverrorbars.main(dictAlg, outputdir, verbose)
# Performance profiles
if isPer:
# ECDFs per noise groups
dictNoi = pproc.dictAlgByNoi(dictAlg)
for ng, tmpdictAlg in dictNoi.iteritems():
dictDim = pproc.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
# pprldmany.main(entries, inset.summarized_target_function_values,
# from . import config
# config.config()
pprldmany.main(
entries, # pass expensive flag here?
order=sortedAlgs,
outputdir=outputdir,
info=('%02dD_%s' % (d, ng)),
verbose=verbose)
# ECDFs per function groups
dictFG = pproc.dictAlgByFuncGroup(dictAlg)
for fg, tmpdictAlg in dictFG.iteritems():
dictDim = pproc.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
def main(dictAlg, isBiobjective, order=None, outputdir='.', info='default',
dimension=None, verbose=True):
"""Generates a figure showing the performance of algorithms.
From a dictionary of :py:class:`DataSetList` sorted by algorithms,
generates the cumulative distribution function of the bootstrap
distribution of ERT for algorithms on multiple functions for
multiple targets altogether.
:param dict dictAlg: dictionary of :py:class:`DataSetList` instances
one instance is equivalent to one algorithm,
:param list targets: target function values
:param list order: sorted list of keys to dictAlg for plotting order
:param str outputdir: output directory
:param str info: output file name suffix
:param bool verbose: controls verbosity
"""
global x_limit # late assignment of default, because it can be set to None in config
global divide_by_dimension # not fully implemented/tested yet
if 'x_limit' not in globals() or x_limit is None:
x_limit = x_limit_default
tmp = pp.dictAlgByDim(dictAlg)
# tmp = pp.DictAlg(dictAlg).by_dim()
if len(tmp) != 1 and dimension is None:
raise ValueError('We never integrate over dimension.')
if dimension is not None:
if dimension not in tmp.keys():
raise ValueError('dimension %d not in dictAlg dimensions %s'
% (dimension, str(tmp.keys())))
tmp = {dimension: tmp[dimension]}
dim = tmp.keys()[0]
divisor = dim if divide_by_dimension else 1
algorithms_with_data = [a for a in dictAlg.keys() if dictAlg[a] != []]
dictFunc = pp.dictAlgByFun(dictAlg)
# Collect data
# Crafting effort correction: should we consider any?
CrEperAlg = {}
for alg in algorithms_with_data:
CrE = 0.
if 1 < 3 and dictAlg[alg][0].algId == 'GLOBAL':
tmp = dictAlg[alg].dictByNoise()
assert len(tmp.keys()) == 1
if tmp.keys()[0] == 'noiselessall':
CrE = 0.5117
elif tmp.keys()[0] == 'nzall':
CrE = 0.6572
CrEperAlg[alg] = CrE
if CrE != 0.0:
print 'Crafting effort for', alg, 'is', CrE
dictData = {} # list of (ert per function) per algorithm
dictMaxEvals = {} # list of (maxevals per function) per algorithm
bestERT = [] # best ert per function
# funcsolved = [set()] * len(targets) # number of functions solved per target
xbest2009 = []
maxevalsbest2009 = []
for f, dictAlgperFunc in dictFunc.iteritems():
if function_IDs and f not in function_IDs:
continue
# print target_values((f, dim))
for j, t in enumerate(target_values((f, dim))):
# for j, t in enumerate(genericsettings.current_testbed.ecdf_target_values(1e2, f)):
# funcsolved[j].add(f)
for alg in algorithms_with_data:
x = [np.inf] * perfprofsamplesize
runlengthunsucc = []
try:
entry = dictAlgperFunc[alg][0] # one element per fun and per dim.
evals = entry.detEvals([t])[0]
assert entry.dim == dim
runlengthsucc = evals[np.isnan(evals) == False] / divisor
runlengthunsucc = entry.maxevals[np.isnan(evals)] / divisor
if len(runlengthsucc) > 0:
x = toolsstats.drawSP(runlengthsucc, runlengthunsucc,
percentiles=[50],
samplesize=perfprofsamplesize)[1]
except (KeyError, IndexError):
#set_trace()
warntxt = ('Data for algorithm %s on function %d in %d-D '
% (alg, f, dim)
+ 'are missing.\n')
warnings.warn(warntxt)
dictData.setdefault(alg, []).extend(x)
dictMaxEvals.setdefault(alg, []).extend(runlengthunsucc)
displaybest2009 = not isBiobjective #disabled until we find the bug
if displaybest2009:
#set_trace()
bestalgentries = bestalg.loadBestAlgorithm(isBiobjective)
bestalgentry = bestalgentries[(dim, f)]
bestalgevals = bestalgentry.detEvals(target_values((f, dim)))
# print bestalgevals
for j in range(len(bestalgevals[0])):
if bestalgevals[1][j]:
evals = bestalgevals[0][j]
#set_trace()
assert dim == bestalgentry.dim
runlengthsucc = evals[np.isnan(evals) == False] / divisor
runlengthunsucc = bestalgentry.maxevals[bestalgevals[1][j]][np.isnan(evals)] / divisor
x = toolsstats.drawSP(runlengthsucc, runlengthunsucc,
percentiles=[50],
samplesize=perfprofsamplesize)[1]
else:
x = perfprofsamplesize * [np.inf]
runlengthunsucc = []
xbest2009.extend(x)
maxevalsbest2009.extend(runlengthunsucc)
if order is None:
order = dictData.keys()
# Display data
lines = []
if displaybest2009:
args = {'ls': '-', 'linewidth': 6, 'marker': 'D', 'markersize': 11.,
'markeredgewidth': 1.5, 'markerfacecolor': refcolor,
'markeredgecolor': refcolor, 'color': refcolor,
'label': 'best 2009', 'zorder': -1}
lines.append(plotdata(np.array(xbest2009), x_limit, maxevalsbest2009,
CrE = 0., **args))
def algname_to_label(algname, dirname=None):
"""to be extended to become generally useful"""
if isinstance(algname, (tuple, list)): # not sure this is needed
return ' '.join([str(name) for name in algname])
return str(algname)
for i, alg in enumerate(order):
try:
data = dictData[alg]
maxevals = dictMaxEvals[alg]
except KeyError:
continue
args = styles[(i) % len(styles)]
args['linewidth'] = 1.5
args['markersize'] = 12.
args['markeredgewidth'] = 1.5
args['markerfacecolor'] = 'None'
args['markeredgecolor'] = args['color']
args['label'] = algname_to_label(alg)
#args['markevery'] = perfprofsamplesize # option available in latest version of matplotlib
#elif len(show_algorithms) > 0:
#args['color'] = 'wheat'
#args['ls'] = '-'
#args['zorder'] = -1
# plotdata calls pprldistr.plotECDF which calls ppfig.plotUnifLog... which does the work
lines.append(plotdata(np.array(data), x_limit, maxevals,
CrE=CrEperAlg[alg], **args))
labels, handles = plotLegend(lines, x_limit)
if True: # isLateXLeg:
fileName = os.path.join(outputdir,'pprldmany_%s.tex' % (info))
with open(fileName, 'w') as f:
f.write(r'\providecommand{\nperfprof}{7}')
algtocommand = {} # latex commands
for i, alg in enumerate(order):
tmp = r'\alg%sperfprof' % pptex.numtotext(i)
f.write(r'\providecommand{%s}{\StrLeft{%s}{\nperfprof}}' %
(tmp, toolsdivers.str_to_latex(
toolsdivers.strip_pathname2(algname_to_label(alg)))))
algtocommand[algname_to_label(alg)] = tmp
if displaybest2009:
tmp = r'\algzeroperfprof'
f.write(r'\providecommand{%s}{best 2009}' % (tmp))
algtocommand['best 2009'] = tmp
commandnames = []
for label in labels:
commandnames.append(algtocommand[label])
# f.write(headleg)
if len(order) > 28: # latex sidepanel won't work well for more than 25 algorithms, but original labels are also clipped
f.write(r'\providecommand{\perfprofsidepanel}{\mbox{%s}\vfill\mbox{%s}}'
% (commandnames[0], commandnames[-1]))
else:
fontsize_command = r'\tiny{}' if len(order) > 19 else ''
f.write(r'\providecommand{\perfprofsidepanel}{{%s\mbox{%s}' %
(fontsize_command, commandnames[0])) # TODO: check len(labels) > 0
for i in range(1, len(labels)):
f.write('\n' + r'\vfill \mbox{%s}' % commandnames[i])
f.write('}}\n')
# f.write(footleg)
if verbose:
print 'Wrote right-hand legend in %s' % fileName
figureName = os.path.join(outputdir,'pprldmany_%s' % (info))
#beautify(figureName, funcsolved, x_limit*x_annote_factor, False, fileFormat=figformat)
beautify()
text = ppfig.consecutiveNumbers(sorted(dictFunc.keys()), 'f')
text += ',%d-D' % dim # TODO: this is strange when different dimensions are plotted
plt.text(0.01, 0.98, text, horizontalalignment="left",
verticalalignment="top", transform=plt.gca().transAxes)
if len(dictFunc) == 1:
plt.title(' '.join((str(dictFunc.keys()[0]),
genericsettings.current_testbed.short_names[dictFunc.keys()[0]])))
a = plt.gca()
plt.xlim(xmin=1e-0, xmax=x_limit**annotation_space_end_relative)
xticks, labels = plt.xticks()
tmp = []
for i in xticks:
tmp.append('%d' % round(np.log10(i)))
a.set_xticklabels(tmp)
if save_figure:
ppfig.saveFigure(figureName, verbose=verbose)
if len(dictFunc) == 1:
ppfig.save_single_functions_html(
os.path.join(outputdir, 'pprldmany'),
'', # algorithms names are clearly visible in the figure
add_to_names='_%02dD' %(dim),
algorithmCount=ppfig.AlgorithmCount.NON_SPECIFIED
)
if close_figure:
plt.close()
def main(dictAlg, order=None, outputdir='.', info='default',
verbose=True):
"""Generates a figure showing the performance of algorithms.
From a dictionary of :py:class:`DataSetList` sorted by algorithms,
generates the cumulative distribution function of the bootstrap
distribution of ERT for algorithms on multiple functions for
multiple targets altogether.
:param dict dictAlg: dictionary of :py:class:`DataSetList` instances
one instance is equivalent to one algorithm,
:param list targets: target function values
:param list order: sorted list of keys to dictAlg for plotting order
:param str outputdir: output directory
:param str info: output file name suffix
:param bool verbose: controls verbosity
"""
global x_limit # late assignment of default, because it can be set to None in config
if 'x_limit' not in globals() or x_limit is None:
x_limit = x_limit_default
tmp = pp.dictAlgByDim(dictAlg)
# tmp = pp.DictAlg(dictAlg).by_dim()
if len(tmp) != 1:
raise Exception('We never integrate over dimension.')
dim = tmp.keys()[0]
algorithms_with_data = [a for a in dictAlg.keys() if dictAlg[a] != []]
dictFunc = pp.dictAlgByFun(dictAlg)
# Collect data
# Crafting effort correction: should we consider any?
CrEperAlg = {}
for alg in algorithms_with_data:
CrE = 0.
if 1 < 3 and dictAlg[alg][0].algId == 'GLOBAL':
tmp = dictAlg[alg].dictByNoise()
assert len(tmp.keys()) == 1
if tmp.keys()[0] == 'noiselessall':
CrE = 0.5117
elif tmp.keys()[0] == 'nzall':
CrE = 0.6572
CrEperAlg[alg] = CrE
if CrE != 0.0:
print 'Crafting effort for', alg, 'is', CrE
dictData = {} # list of (ert per function) per algorithm
dictMaxEvals = {} # list of (maxevals per function) per algorithm
bestERT = [] # best ert per function
# funcsolved = [set()] * len(targets) # number of functions solved per target
xbest2009 = []
maxevalsbest2009 = []
for f, dictAlgperFunc in dictFunc.iteritems():
if function_IDs and f not in function_IDs:
continue
# print target_values((f, dim))
for j, t in enumerate(target_values((f, dim))):
# for j, t in enumerate(genericsettings.current_testbed.ecdf_target_values(1e2, f)):
# funcsolved[j].add(f)
for alg in algorithms_with_data:
x = [np.inf] * perfprofsamplesize
runlengthunsucc = []
try:
entry = dictAlgperFunc[alg][0] # one element per fun and per dim.
evals = entry.detEvals([t])[0]
runlengthsucc = evals[np.isnan(evals) == False] / entry.dim
runlengthunsucc = entry.maxevals[np.isnan(evals)] / entry.dim
if len(runlengthsucc) > 0:
x = toolsstats.drawSP(runlengthsucc, runlengthunsucc,
percentiles=[50],
samplesize=perfprofsamplesize)[1]
except (KeyError, IndexError):
#set_trace()
warntxt = ('Data for algorithm %s on function %d in %d-D '
% (alg, f, dim)
+ 'are missing.\n')
warnings.warn(warntxt)
dictData.setdefault(alg, []).extend(x)
dictMaxEvals.setdefault(alg, []).extend(runlengthunsucc)
if displaybest2009:
#set_trace()
if not bestalg.bestalgentries2009:
bestalg.loadBBOB2009()
bestalgentry = bestalg.bestalgentries2009[(dim, f)]
bestalgevals = bestalgentry.detEvals(target_values((f, dim)))
# print bestalgevals
for j in range(len(bestalgevals[0])):
if bestalgevals[1][j]:
evals = bestalgevals[0][j]
#set_trace()
runlengthsucc = evals[np.isnan(evals) == False] / bestalgentry.dim
runlengthunsucc = bestalgentry.maxevals[bestalgevals[1][j]][np.isnan(evals)] / bestalgentry.dim
x = toolsstats.drawSP(runlengthsucc, runlengthunsucc,
percentiles=[50],
samplesize=perfprofsamplesize)[1]
else:
x = perfprofsamplesize * [np.inf]
runlengthunsucc = []
xbest2009.extend(x)
maxevalsbest2009.extend(runlengthunsucc)
if order is None:
order = dictData.keys()
# Display data
lines = []
if displaybest2009:
args = {'ls': '-', 'linewidth': 6, 'marker': 'D', 'markersize': 11.,
'markeredgewidth': 1.5, 'markerfacecolor': refcolor,
'markeredgecolor': refcolor, 'color': refcolor,
'label': 'best 2009', 'zorder': -1}
lines.append(plotdata(np.array(xbest2009), x_limit, maxevalsbest2009,
CrE = 0., **args))
for i, alg in enumerate(order):
try:
data = dictData[alg]
maxevals = dictMaxEvals[alg]
except KeyError:
continue
args = styles[(i) % len(styles)]
args['linewidth'] = 1.5
args['markersize'] = 12.
args['markeredgewidth'] = 1.5
args['markerfacecolor'] = 'None'
args['markeredgecolor'] = args['color']
args['label'] = alg
#args['markevery'] = perfprofsamplesize # option available in latest version of matplotlib
#elif len(show_algorithms) > 0:
#args['color'] = 'wheat'
#args['ls'] = '-'
#args['zorder'] = -1
lines.append(plotdata(np.array(data), x_limit, maxevals,
CrE=CrEperAlg[alg], **args))
labels, handles = plotLegend(lines, x_limit)
if True: #isLateXLeg:
fileName = os.path.join(outputdir,'pprldmany_%s.tex' % (info))
try:
f = open(fileName, 'w')
f.write(r'\providecommand{\nperfprof}{7}')
algtocommand = {}
for i, alg in enumerate(order):
tmp = r'\alg%sperfprof' % pptex.numtotext(i)
f.write(r'\providecommand{%s}{\StrLeft{%s}{\nperfprof}}' % (tmp, toolsdivers.str_to_latex(toolsdivers.strip_pathname2(alg))))
algtocommand[alg] = tmp
commandnames = []
if displaybest2009:
tmp = r'\algzeroperfprof'
f.write(r'\providecommand{%s}{best 2009}' % (tmp))
algtocommand['best 2009'] = tmp
for l in labels:
commandnames.append(algtocommand[l])
# f.write(headleg)
f.write(r'\providecommand{\perfprofsidepanel}{\mbox{%s}' % commandnames[0]) # TODO: check len(labels) > 0
for i in range(1, len(labels)):
f.write('\n' + r'\vfill \mbox{%s}' % commandnames[i])
f.write('}\n')
# f.write(footleg)
if verbose:
print 'Wrote right-hand legend in %s' % fileName
except:
raise # TODO: Does this make sense?
else:
f.close()
figureName = os.path.join(outputdir,'pprldmany_%s' % (info))
#beautify(figureName, funcsolved, x_limit*x_annote_factor, False, fileFormat=figformat)
beautify()
text = 'f%s' % (ppfig.consecutiveNumbers(sorted(dictFunc.keys())))
text += ',%d-D' % dim
plt.text(0.01, 0.98, text, horizontalalignment="left",
verticalalignment="top", transform=plt.gca().transAxes)
a = plt.gca()
plt.xlim(xmin=1e-0, xmax=x_limit**annotation_space_end_relative)
xticks, labels = plt.xticks()
tmp = []
for i in xticks:
tmp.append('%d' % round(np.log10(i)))
a.set_xticklabels(tmp)
ppfig.saveFigure(figureName, verbose=verbose)
plt.close()
def extractBestAlgorithms(args = algs2009, f_factor=2,
target_lb=1e-8, target_ub=1e22):
"""Returns (and prints) per dimension a list of algorithms within
algorithm list args that contains an algorithm if for any
dimension/target/function pair this algorithm:
- is the best algorithm wrt ERT
- its own ERT lies within a factor f_factor of the best ERT
- there is no algorithm within a factor of f_factor of the best ERT
and the current algorithm is the second best.
"""
# TODO: use pproc.TargetValues class as input target values
# default target values:
targets = pproc.TargetValues(
10**np.arange(np.log10(max((1e-8, target_lb))),
np.log10(target_ub) + 1e-9, 0.2))
# there should be a simpler way to express this to become the
# interface of this function
print 'Loading algorithm data from given algorithm list...\n'
verbose = True
dsList, sortedAlgs, dictAlg = pproc.processInputArgs(args, verbose=verbose)
print 'This may take a while (depending on the number of algorithms)'
selectedAlgsPerProblem = {}
for f, i in pproc.dictAlgByFun(dictAlg).iteritems():
for d, j in pproc.dictAlgByDim(i).iteritems():
selectedAlgsPerProblemDF = []
best = BestAlgSet(j)
for i in range(0, len(best.target)):
t = best.target[i]
# if ((t <= target_ub) and (t >= target_lb)):
if toolsstats.in_approximately(t,
targets((f, d), discretize=True)):
# add best for this target:
selectedAlgsPerProblemDF.append(best.algs[i])
# add second best or all algorithms that have an ERT
# within a factor of f_factor of the best:
secondbest_ERT = np.infty
secondbest_str = ''
secondbest_included = False
for astring in j:
currdictalg = dictAlg[astring].dictByDim()
if currdictalg.has_key(d):
curralgdata = currdictalg[d][f-1]
currERT = curralgdata.detERT([t])[0]
if (astring != best.algs[i]):
if (currERT < secondbest_ERT):
secondbest_ERT = currERT
secondbest_str = astring
if (currERT <= best.detERT([t])[0] * f_factor):
selectedAlgsPerProblemDF.append(astring)
secondbest_included = True
if not (secondbest_included) and (secondbest_str != ''):
selectedAlgsPerProblemDF.append(secondbest_str)
if len(selectedAlgsPerProblemDF) > 0:
selectedAlgsPerProblem[(d, f)] = selectedAlgsPerProblemDF
print 'pre-processing of function', f, 'done.'
print 'loading of best algorithm(s) data done.'
countsperalgorithm = {}
for (d, f) in selectedAlgsPerProblem:
print 'dimension:', d, ', function:', f
setofalgs = set(selectedAlgsPerProblem[d,f])
# now count how often algorithm a is best for the extracted targets
for a in setofalgs:
# use setdefault to initialize with zero if a entry not existant:
countsperalgorithm.setdefault((d, a), 0)
countsperalgorithm[(d,a)] += selectedAlgsPerProblem[d,f].count(a)
selectedalgsperdimension = {}
for (d,a) in sorted(countsperalgorithm):
if not selectedalgsperdimension.has_key(d):
selectedalgsperdimension[d] = []
selectedalgsperdimension[d].append((countsperalgorithm[(d,a)], a))
for d in sorted(selectedalgsperdimension):
print d, 'D:'
for (count, alg) in sorted(selectedalgsperdimension[d], reverse=True):
print count, alg
print '\n'
print " done."
return selectedalgsperdimension
def extractBestAlgorithms(args=algs2009,
f_factor=2,
target_lb=1e-8,
target_ub=1e22):
"""Returns (and prints) per dimension a list of algorithms within
algorithm list args that contains an algorithm if for any
dimension/target/function pair this algorithm:
- is the best algorithm wrt ERT
- its own ERT lies within a factor f_factor of the best ERT
- there is no algorithm within a factor of f_factor of the best ERT
and the current algorithm is the second best.
"""
# TODO: use pproc.TargetValues class as input target values
# default target values:
targets = pproc.TargetValues(10**np.arange(
np.log10(max((1e-8, target_lb))),
np.log10(target_ub) + 1e-9, 0.2))
# there should be a simpler way to express this to become the
# interface of this function
print 'Loading algorithm data from given algorithm list...\n'
verbose = True
dsList, sortedAlgs, dictAlg = pproc.processInputArgs(args, verbose=verbose)
print 'This may take a while (depending on the number of algorithms)'
selectedAlgsPerProblem = {}
for f, i in pproc.dictAlgByFun(dictAlg).iteritems():
for d, j in pproc.dictAlgByDim(i).iteritems():
selectedAlgsPerProblemDF = []
best = BestAlgSet(j)
for i in range(0, len(best.target)):
t = best.target[i]
# if ((t <= target_ub) and (t >= target_lb)):
if toolsstats.in_approximately(
t, targets((f, d), discretize=True)):
# add best for this target:
selectedAlgsPerProblemDF.append(best.algs[i])
# add second best or all algorithms that have an ERT
# within a factor of f_factor of the best:
secondbest_ERT = np.infty
secondbest_str = ''
secondbest_included = False
for astring in j:
currdictalg = dictAlg[astring].dictByDim()
if currdictalg.has_key(d):
curralgdata = currdictalg[d][f - 1]
currERT = curralgdata.detERT([t])[0]
if (astring != best.algs[i]):
if (currERT < secondbest_ERT):
secondbest_ERT = currERT
secondbest_str = astring
if (currERT <= best.detERT([t])[0] * f_factor):
selectedAlgsPerProblemDF.append(astring)
secondbest_included = True
if not (secondbest_included) and (secondbest_str != ''):
selectedAlgsPerProblemDF.append(secondbest_str)
if len(selectedAlgsPerProblemDF) > 0:
selectedAlgsPerProblem[(d, f)] = selectedAlgsPerProblemDF
print 'pre-processing of function', f, 'done.'
print 'loading of best algorithm(s) data done.'
countsperalgorithm = {}
for (d, f) in selectedAlgsPerProblem:
print 'dimension:', d, ', function:', f
setofalgs = set(selectedAlgsPerProblem[d, f])
# now count how often algorithm a is best for the extracted targets
for a in setofalgs:
# use setdefault to initialize with zero if a entry not existant:
countsperalgorithm.setdefault((d, a), 0)
countsperalgorithm[(d, a)] += selectedAlgsPerProblem[d, f].count(a)
selectedalgsperdimension = {}
for (d, a) in sorted(countsperalgorithm):
if not selectedalgsperdimension.has_key(d):
selectedalgsperdimension[d] = []
selectedalgsperdimension[d].append((countsperalgorithm[(d, a)], a))
for d in sorted(selectedalgsperdimension):
print d, 'D:'
for (count, alg) in sorted(selectedalgsperdimension[d], reverse=True):
print count, alg
print '\n'
print " done."
return selectedalgsperdimension
def main(dictAlg, sortedAlgs=None, target=ftarget_default, outputdir='ppdata', verbose=True):
"""From a DataSetList, returns figures showing the scaling: ERT/dim vs dim.
One function and one target per figure.
``target`` can be a scalar, a list with one element or a
``pproc.TargetValues`` instance with one target.
``sortedAlgs`` is a list of string-identifies (folder names)
"""
# target becomes a TargetValues "list" with one element
target = pproc.TargetValues.cast([target] if numpy.isscalar(target) else target)
latex_commands_filename = os.path.join(outputdir, 'bbob_pproc_commands.tex')
assert isinstance(target, pproc.TargetValues)
if len(target) != 1:
raise ValueError('only a single target can be managed in ppfigs, ' + str(len(target)) + ' targets were given')
dictFunc = pproc.dictAlgByFun(dictAlg)
if sortedAlgs is None:
sortedAlgs = sorted(dictAlg.keys())
if not os.path.isdir(outputdir):
os.mkdir(outputdir)
for f in dictFunc:
filename = os.path.join(outputdir,'ppfigs_f%03d' % (f))
handles = []
fix_styles(len(sortedAlgs)) #
for i, alg in enumerate(sortedAlgs):
dictDim = dictFunc[f][alg].dictByDim() # this does not look like the most obvious solution
#Collect data
dimert = []
ert = []
dimnbsucc = []
ynbsucc = []
nbsucc = []
dimmaxevals = []
maxevals = []
dimmedian = []
medianfes = []
for dim in sorted(dictDim):
assert len(dictDim[dim]) == 1
entry = dictDim[dim][0]
data = generateData(entry, target((f, dim))[0]) # TODO: here we might want a different target for each function
if 1 < 3 or data[2] == 0: # No success
dimmaxevals.append(dim)
maxevals.append(float(data[3])/dim)
if data[2] > 0:
dimmedian.append(dim)
medianfes.append(data[4]/dim)
dimert.append(dim)
ert.append(float(data[0])/dim)
if data[1] < 1.:
dimnbsucc.append(dim)
ynbsucc.append(float(data[0])/dim)
nbsucc.append('%d' % data[2])
# Draw lines
tmp = plt.plot(dimert, ert, **styles[i]) #label=alg, )
plt.setp(tmp[0], markeredgecolor=plt.getp(tmp[0], 'color'))
# For legend
# tmp = plt.plot([], [], label=alg.replace('..' + os.sep, '').strip(os.sep), **styles[i])
tmp = plt.plot([], [], label=alg.split(os.sep)[-1], **styles[i])
plt.setp(tmp[0], markersize=12.,
markeredgecolor=plt.getp(tmp[0], 'color'))
if dimmaxevals:
tmp = plt.plot(dimmaxevals, maxevals, **styles[i])
plt.setp(tmp[0], markersize=20, #label=alg,
markeredgecolor=plt.getp(tmp[0], 'color'),
markeredgewidth=1,
markerfacecolor='None', linestyle='None')
handles.append(tmp)
#tmp2 = plt.plot(dimmedian, medianfes, ls='', marker='+',
# markersize=30, markeredgewidth=5,
# markeredgecolor=plt.getp(tmp, 'color'))[0]
#for i, n in enumerate(nbsucc):
# plt.text(dimnbsucc[i], numpy.array(ynbsucc[i])*1.85, n,
# verticalalignment='bottom',
# horizontalalignment='center')
if not bestalg.bestalgentries2009:
bestalg.loadBBOB2009()
bestalgdata = []
dimbestalg = list(df[0] for df in bestalg.bestalgentries2009 if df[1] == f)
dimbestalg.sort()
dimbestalg2 = []
for d in dimbestalg:
entry = bestalg.bestalgentries2009[(d, f)]
tmp = entry.detERT(target((f, d)))[0]
if numpy.isfinite(tmp):
bestalgdata.append(float(tmp)/d)
dimbestalg2.append(d)
tmp = plt.plot(dimbestalg2, bestalgdata, color=refcolor, linewidth=10,
marker='d', markersize=25, markeredgecolor=refcolor, zorder=-1
#label='best 2009',
)
handles.append(tmp)
if show_significance: # plot significance-stars
xstar, ystar = [], []
dims = sorted(pproc.dictAlgByDim(dictFunc[f]))
for i, dim in enumerate(dims):
datasets = pproc.dictAlgByDim(dictFunc[f])[dim]
assert all([len(datasets[ialg]) == 1 for ialg in sortedAlgs if datasets[ialg]])
dsetlist = [datasets[ialg][0] for ialg in sortedAlgs if datasets[ialg]]
if len(dsetlist) > 1:
arzp, arialg = toolsstats.significance_all_best_vs_other(dsetlist, target((f, dim)))
if arzp[0][1] * len(dims) < show_significance:
ert = dsetlist[arialg[0]].detERT(target((f, dim)))[0]
if ert < numpy.inf:
xstar.append(dim)
ystar.append(ert/dim)
plt.plot(xstar, ystar, 'k*', markerfacecolor=None, markeredgewidth=2, markersize=0.5*styles[0]['markersize'])
if funInfos:
plt.gca().set_title(funInfos[f])
isLegend = False
if legend:
plotLegend(handles)
elif 1 < 3:
if f in (1, 24, 101, 130) and len(sortedAlgs) < 6: # 6 elements at most in the boxed legend
isLegend = True
beautify(legend=isLegend, rightlegend=legend)
plt.text(plt.xlim()[0], plt.ylim()[0], 'target ' + target.label_name() + ': ' + target.label(0)) # TODO: check
saveFigure(filename, verbose=verbose)
plt.close()
# generate commands in tex file:
try:
abc = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
alg_definitions = []
for i in range(len(sortedAlgs)):
symb = r'{%s%s}' % (color_to_latex(styles[i]['color']),
marker_to_latex(styles[i]['marker']))
alg_definitions.append((', ' if i > 0 else '') + '%s:%s' % (symb, '\\algorithm' + abc[i % len(abc)]))
toolsdivers.prepend_to_file(latex_commands_filename,
[#'\\providecommand{\\bbobppfigsftarget}{\\ensuremath{10^{%s}}}'
# % target.loglabel(0), # int(numpy.round(numpy.log10(target))),
'\\providecommand{\\bbobppfigslegend}[1]{',
scaling_figure_caption(target),
'Legend: '] + alg_definitions + ['}']
)
toolsdivers.prepend_to_file(latex_commands_filename,
['\\providecommand{\\bbobECDFslegend}[1]{',
ecdfs_figure_caption(target), '}']
)
if verbose:
print 'Wrote commands and legend to %s' % filename
# this is obsolete (however check templates)
filename = os.path.join(outputdir,'ppfigs.tex')
f = open(filename, 'w')
f.write('% Do not modify this file: calls to post-processing software'
+ ' will overwrite any modification.\n')
f.write('Legend: ')
for i in range(0, len(sortedAlgs)):
symb = r'{%s%s}' % (color_to_latex(styles[i]['color']),
marker_to_latex(styles[i]['marker']))
f.write((', ' if i > 0 else '') + '%s:%s' % (symb, writeLabels(sortedAlgs[i])))
f.close()
if verbose:
print '(obsolete) Wrote legend in %s' % filename
except IOError:
raise
handles.append(tmp)
if funInfos:
plt.gca().set_title(funInfos[f])
beautify(rightlegend=legend)
if legend:
plotLegend(handles)
else:
if f in (1, 24, 101, 130):
plt.legend()
saveFigure(filename, figFormat=genericsettings.fig_formats, verbose=verbose)
plt.close()
def main(argv=None):
"""Main routine for post-processing the data of multiple algorithms.
Keyword arguments:
argv -- list of strings containing options and arguments. If not provided,
sys.argv is accessed.
argv must list folders containing BBOB data files. Each of these folders
should correspond to the data of ONE algorithm and should be listed in
algorithmshortinfos.txt, a file from the bbob_pproc.compall package listing
the information of various algorithms treated using bbob_pproc.dataoutput
Furthermore, argv can begin with, in any order, facultative option flags
listed below.
-h, --help
display this message
-v, --verbose
verbose mode, prints out operations. When not in verbose mode, no
output is to be expected, except for errors.
-o, --output-dir OUTPUTDIR
change the default output directory ('defaultoutputdirectory') to
OUTPUTDIR
--noise-free, --noisy
restrain the post-processing to part of the data set only. Actually
quicken the post-processing since it loads only part of the pickle
files.
--tab-only, --perfprof-only
these options can be used to output respectively the comparison
tex tables or the performance profiles only.
A combination of any two of these options results in
no output.
Exceptions raised:
Usage -- Gives back a usage message.
Examples:
* Calling the runcompall.py interface from the command line:
$ python bbob_pproc/runcompall.py -v
* Loading this package and calling the main from the command line
(requires that the path to this package is in python search path):
$ python -m bbob_pproc.runcompall -h
This will print out this help message.
* From the python interactive shell (requires that the path to this
package is in python search path):
>>> from bbob_pproc import runcompall
>>> runcompall.main('-o outputfolder folder1 folder2'.split())
This will execute the post-processing on the data found in folder1
and folder2.
The -o option changes the output folder from the default cmpalldata to
outputfolder.
* Generate post-processing data for some algorithms:
$ python runcompall.py AMALGAM BFGS CMA-ES
"""
if argv is None:
argv = sys.argv[1:]
try:
try:
opts, args = getopt.getopt(argv, "hvo:",
["help", "output-dir=", "noisy",
"noise-free", "perfprof-only",
"tab-only", "verbose"])
except getopt.error, msg:
raise Usage(msg)
if not (args):
usage()
sys.exit()
verbose = False
outputdir = 'cmpalldata'
isNoisy = False
isNoiseFree = False
isPer = True
isTab = True
#Process options
for o, a in opts:
if o in ("-v","--verbose"):
verbose = True
elif o in ("-h", "--help"):
usage()
sys.exit()
elif o in ("-o", "--output-dir"):
outputdir = a
elif o == "--noisy":
isNoisy = True
elif o == "--noise-free":
isNoiseFree = True
elif o == "--tab-only":
isPer = False
isEff = False
elif o == "--perfprof-only":
isEff = False
isTab = False
else:
assert False, "unhandled option"
if (not verbose):
warnings.simplefilter('ignore')
print ("BBOB Post-processing: will generate comparison " +
"data in folder %s" % outputdir)
print " this might take several minutes."
dsList, sortedAlgs, dictAlg = processInputArgs(args, verbose=verbose)
if not dsList:
sys.exit()
for i in dictAlg:
if isNoisy and not isNoiseFree:
dictAlg[i] = dictAlg[i].dictByNoise().get('nzall', DataSetList())
elif isNoiseFree and not isNoisy:
dictAlg[i] = dictAlg[i].dictByNoise().get('noiselessall', DataSetList())
tmp = set((j.algId, j.comment) for j in dictAlg[i])
for j in tmp:
if not dataoutput.isListed(j):
dataoutput.updateAlgorithmInfo(j, verbose=verbose)
for i in dsList:
if not i.dim in (2, 3, 5, 10, 20):
continue
# Deterministic algorithms
if i.algId in ('Original DIRECT', ):
tmpInstancesOfInterest = instancesOfInterestDet
else:
tmpInstancesOfInterest = instancesOfInterest
if ((dict((j, i.itrials.count(j)) for j in set(i.itrials)) <
tmpInstancesOfInterest) and
(dict((j, i.itrials.count(j)) for j in set(i.itrials)) <
instancesOfInterest2010)):
warnings.warn('The data of %s do not list ' %(i) +
'the correct instances ' +
'of function F%d or the ' %(i.funcId) +
'correct number of trials for each.')
# group targets:
dictTarget = {}
for t in sorted(set(single_target_function_values + summarized_target_function_values)):
tmpdict = dict.fromkeys(((f, d) for f in range(0, 25) + range(101, 131) for d in (2, 3, 5, 10, 20, 40)), t)
stmp = 'E'
if t == 1:
stmp = 'E-'
# dictTarget['_f' + stmp + '%2.1f' % numpy.log10(t)] = (tmpdict, )
if t in single_target_function_values:
dictTarget['_f' + stmp + '%02d' % numpy.log10(t)] = (tmpdict, )
if t in summarized_target_function_values:
dictTarget.setdefault('_allfs', []).append(tmpdict)
if not os.path.exists(outputdir):
os.mkdir(outputdir)
if verbose:
print 'Folder %s was created.' % (outputdir)
# Performance profiles
if isPer:
dictNoi = pproc.dictAlgByNoi(dictAlg)
for ng, tmpdictAlg in dictNoi.iteritems():
dictDim = pproc.dictAlgByDim(tmpdictAlg)
for d, entries in dictDim.iteritems():
for k, t in dictTarget.iteritems():
#set_trace()
ppperfprof.main(entries, target=t, order=sortedAlgs,
plotArgs=algPlotInfos,
outputdir=outputdir,
info=('%02d%s_%s' % (d, k, ng)),
verbose=verbose)
organizeRTDpictures.do(outputdir)
print "ECDFs of ERT figures done."
if isTab:
allmintarget, allertbest = detTarget(dsList)
pptables.tablemanyalgonefunc(dictAlg, allmintarget, allertbest,
sortedAlgs, outputdir)
print "Comparison tables done."
def main(dictAlg, sortedAlgs, target=1e-8, outputdir='ppdata', verbose=True):
"""From a DataSetList, returns figures showing the scaling: ERT/dim vs dim.
One function and one target per figure.
sortedAlgs is a list of string-identifies (folder names)
"""
dictFunc = pproc.dictAlgByFun(dictAlg)
for f in dictFunc:
filename = os.path.join(outputdir,'ppfigs_f%03d' % (f))
handles = []
fix_styles(len(sortedAlgs)) #
for i, alg in enumerate(sortedAlgs):
dictDim = dictFunc[f][alg].dictByDim()
#Collect data
dimert = []
ert = []
dimnbsucc = []
ynbsucc = []
nbsucc = []
dimmaxevals = []
maxevals = []
dimmedian = []
medianfes = []
for dim in sorted(dictDim):
assert len(dictDim[dim]) == 1
entry = dictDim[dim][0]
data = generateData(entry, target) # TODO: here we might want a different target for each function
if 1 < 3 or data[2] == 0: # No success
dimmaxevals.append(dim)
maxevals.append(float(data[3])/dim)
if data[2] > 0:
dimmedian.append(dim)
medianfes.append(data[4]/dim)
dimert.append(dim)
ert.append(float(data[0])/dim)
if data[1] < 1.:
dimnbsucc.append(dim)
ynbsucc.append(float(data[0])/dim)
nbsucc.append('%d' % data[2])
# Draw lines
tmp = plt.plot(dimert, ert, **styles[i]) #label=alg, )
plt.setp(tmp[0], markeredgecolor=plt.getp(tmp[0], 'color'))
# For legend
# tmp = plt.plot([], [], label=alg.replace('..' + os.sep, '').strip(os.sep), **styles[i])
tmp = plt.plot([], [], label=alg.split(os.sep)[-1], **styles[i])
plt.setp(tmp[0], markersize=12.,
markeredgecolor=plt.getp(tmp[0], 'color'))
if dimmaxevals:
tmp = plt.plot(dimmaxevals, maxevals, **styles[i])
plt.setp(tmp[0], markersize=20, #label=alg,
markeredgecolor=plt.getp(tmp[0], 'color'),
markeredgewidth=1,
markerfacecolor='None', linestyle='None')
handles.append(tmp)
#tmp2 = plt.plot(dimmedian, medianfes, ls='', marker='+',
# markersize=30, markeredgewidth=5,
# markeredgecolor=plt.getp(tmp, 'color'))[0]
#for i, n in enumerate(nbsucc):
# plt.text(dimnbsucc[i], numpy.array(ynbsucc[i])*1.85, n,
# verticalalignment='bottom',
# horizontalalignment='center')
if not bestalg.bestalgentries2009:
bestalg.loadBBOB2009()
bestalgdata = []
dimbestalg = list(df[0] for df in bestalg.bestalgentries2009 if df[1] == f)
dimbestalg.sort()
dimbestalg2 = []
for d in dimbestalg:
entry = bestalg.bestalgentries2009[(d, f)]
tmp = entry.detERT([target])[0]
if numpy.isfinite(tmp):
bestalgdata.append(float(tmp)/d)
dimbestalg2.append(d)
tmp = plt.plot(dimbestalg2, bestalgdata, color=refcolor, linewidth=10,
marker='d', markersize=25, markeredgecolor=refcolor, zorder=-1
#label='best 2009',
)
handles.append(tmp)
if show_significance: # plot significance-stars
xstar, ystar = [], []
dims = sorted(pproc.dictAlgByDim(dictFunc[f]))
for i, dim in enumerate(dims):
datasets = pproc.dictAlgByDim(dictFunc[f])[dim]
assert all([len(datasets[ialg]) == 1 for ialg in sortedAlgs if datasets[ialg]])
dsetlist = [datasets[ialg][0] for ialg in sortedAlgs if datasets[ialg]]
if len(dsetlist) > 1:
arzp, arialg = toolsstats.significance_all_best_vs_other(dsetlist, [target])
if arzp[0][1] * len(dims) < 0.05:
ert = dsetlist[arialg[0]].detERT([target])[0]
if ert < numpy.inf:
xstar.append(dim)
ystar.append(ert/dim)
plt.plot(xstar, ystar, 'k*', markerfacecolor=None, markeredgewidth=2, markersize=0.5*styles[0]['markersize'])
if funInfos:
plt.gca().set_title(funInfos[f])
isLegend = False
if legend:
plotLegend(handles)
elif 1 < 3:
if f in (1, 24, 101, 130) and len(sortedAlgs) < 6: # 6 elements at most in the boxed legend
isLegend = True
beautify(legend=isLegend, rightlegend=legend)
plt.text(plt.xlim()[0], plt.ylim()[0], 'ftarget=%.0e' % target)
saveFigure(filename, verbose=verbose)
plt.close()
# generate commands in tex file:
try:
abc = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
alg_definitions = []
for i in range(len(sortedAlgs)):
symb = r'{%s%s}' % (color_to_latex(styles[i]['color']),
marker_to_latex(styles[i]['marker']))
alg_definitions.append((', ' if i > 0 else '') + '%s:%s' % (symb, '\\algorithm' + abc[i % len(abc)]))
filename = os.path.join(outputdir, 'bbob_pproc_commands.tex')
toolsdivers.prepend_to_file(filename,
['\\providecommand{\\bbobppfigsftarget}{\\ensuremath{10^{%d}}}'
% int(numpy.round(numpy.log10(target))),
'\\providecommand{\\bbobppfigslegend}[1]{',
scaling_figure_legend,
'Legend: '] + alg_definitions + ['}']
)
if verbose:
print 'Wrote commands and legend to %s' % filename
# this is obsolete (however check templates)
filename = os.path.join(outputdir,'ppfigs.tex')
f = open(filename, 'w')
f.write('% Do not modify this file: calls to post-processing software'
+ ' will overwrite any modification.\n')
f.write('Legend: ')
for i in range(0, len(sortedAlgs)):
symb = r'{%s%s}' % (color_to_latex(styles[i]['color']),
marker_to_latex(styles[i]['marker']))
f.write((', ' if i > 0 else '') + '%s:%s' % (symb, writeLabels(sortedAlgs[i])))
f.close()
if verbose:
print '(obsolete) Wrote legend in %s' % filename
except IOError:
raise
handles.append(tmp)
if funInfos:
plt.gca().set_title(funInfos[f])
beautify(rightlegend=legend)
if legend:
plotLegend(handles)
else:
if f in (1, 24, 101, 130):
plt.legend()
saveFigure(filename, figFormat=genericsettings.fig_formats, verbose=verbose)
plt.close()