def parse_args(xvar='laps', yvar='evidence'):
''' Returns Namespace of parsed arguments retrieved from command line
'''
parser = argparse.ArgumentParser()
parser.add_argument('dataName', type=str, default='AsteriskK8')
parser.add_argument('jpath', type=str, default='demo*')
parser.add_argument('--xvar', type=str, default=xvar,
choices=list(LabelMap.keys()),
help="name of x axis variable to plot.")
parser.add_argument('--yvar', type=str, default=yvar,
#choices=LabelMap.keys(),
help="name of y axis variable to plot.")
helpMsg = "ids of trials/runs to plot from given job." + \
" Example: '4' or '1,2,3' or '2-6'."
parser.add_argument(
'--taskids', type=str, default=None, help=helpMsg)
parser.add_argument(
'--savefilename', type=str, default=None,
help="location where to save figure (absolute path directory)")
args, unkList = parser.parse_known_args()
argDict = BNPYArgParser.arglist_to_kwargs(unkList, doConvertFromStr=False)
argDict.update(args.__dict__)
argDict['jpathPattern'] = os.path.join(os.environ['BNPYOUTDIR'],
args.dataName,
args.jpath)
del argDict['dataName']
del argDict['jpath']
return argDict
def parse_args():
''' Returns Namespace of parsed arguments retrieved from command line
'''
parser = argparse.ArgumentParser()
parser.add_argument('dataName', type=str, default='AsteriskK8')
parser.add_argument('jpath', type=str, default='demo*')
helpMsg = "ids of trials/runs to plot from given job." + \
" Example: '4' or '1,2,3' or '2-6'."
parser.add_argument(
'--taskids',
type=str, default=None,
help=helpMsg)
parser.add_argument(
'--savefilename', type=str, default=None,
help="location where to save figure (absolute path directory)")
parser.add_argument('--fileSuffix', type=str, default='PredLik.mat')
args, unkList = parser.parse_known_args()
argDict = BNPYArgParser.arglist_to_kwargs(unkList)
argDict.update(args.__dict__)
argDict['jpathPattern'] = os.path.join(os.environ['BNPYOUTDIR'],
args.dataName,
args.jpath)
del argDict['dataName']
del argDict['jpath']
return argDict
def parse_args(**kwargs):
''' Returns Namespace of parsed arguments retrieved from command line
'''
parser = argparse.ArgumentParser()
parser.add_argument('dataName', type=str, default='AsteriskK8')
parser.add_argument('jpathPattern', type=str, default='demo*')
parser.add_argument('--xvar', type=str, default=None,
help="name of x axis variable to plot.")
parser.add_argument('--yvar', type=str, default='evidence',
choices=list(LabelMap.keys()),
help="name of y axis variable to plot.")
parser.add_argument('--lvar', type=str, default=None,
help="quantity that varies across lines")
parser.add_argument('--pvar', type=str, default=None,
help="quantity that varies across subplots")
parser.add_argument('--taskids', type=str, default='all',
help="specify which task to plot (all, .best, .worst, etc)")
parser.add_argument(
'--savefilename', type=str, default=None,
help="location where to save figure (absolute path directory)")
args, unkList = parser.parse_known_args()
argDict = BNPYArgParser.arglist_to_kwargs(unkList)
argDict.update(args.__dict__)
argDict.update(kwargs)
argDict['jpathPattern'] = os.path.join(os.environ['BNPYOUTDIR'],
args.dataName,
args.jpathPattern)
del argDict['dataName']
for key in argDict:
if key.endswith('vals'):
if not isinstance(argDict[key], list):
argDict[key] = argDict[key].split(',')
return argDict
def parse_args():
''' Parse cmd line arguments
'''
parser = argparse.ArgumentParser()
BNPYArgParser.addRequiredVizArgsToParser(parser)
BNPYArgParser.addStandardVizArgsToParser(parser)
parser.add_argument('--lap', default=None, type=float,
help="Specific lap at which to plot parameters." \
+ " If exact lap not available, instead plots nearest lap.")
parser.add_argument('--doPlotData', action='store_true', default=False,
help="If present, also plot training data.")
parser.add_argument('--doPlotTruth', action='store_true', default=False,
help="If present, also plot true model params that generated data.")
parser.add_argument('--doSort', action='store_true', default=False,
help="If present, sort parameters by global appearance probabilities.")
args = parser.parse_args()
return args
def parse_args():
''' Returns Namespace of parsed arguments retrieved from command line
'''
parser = argparse.ArgumentParser()
BNPYArgParser.addRequiredVizArgsToParser(parser)
BNPYArgParser.addStandardVizArgsToParser(parser)
parser.add_argument('--xvar', type=str, default='laps',
help="name of x axis variable to plot. one of {iters,laps,times}")
parser.add_argument('--traceEvery', type=str, default=None,
help="Specifies how often to plot data points. For example, traceEvery=10 only plots data points associated with laps divisible by 10.")
parser.add_argument('--legendnames', type=str, default=None,
help="optional names to show on legend in place of jobnames")
args = parser.parse_args()
args.algNames = args.algNames.split(',')
args.jobnames = args.jobnames.split(',')
if args.legendnames is not None:
args.legendnames = args.legendnames.split(',')
#assert len(args.legendnames) == len(args.jobnames) * len(args.algNames)
return args
def plot_all_tasks_for_job(jobpath, args, jobname=None, color=None):
''' Create line plot in current matplotlib figure
for each task/run of the designated jobpath
'''
if not os.path.exists(jobpath):
raise ValueError("No such path: %s" % (jobpath))
taskids = BNPYArgParser.parse_task_ids(jobpath, args.taskids)
xAll = list()
yAll = list()
xLocs = list()
yLocs = list()
for tt, taskid in enumerate(taskids):
xs = np.loadtxt(os.path.join(jobpath, taskid, args.xvar+'.txt'))
ys = np.loadtxt(os.path.join(jobpath, taskid, 'evidence.txt'))
# remove first-lap of moVB, since ELBO is not accurate
if jobpath.count('moVB') > 0 and args.xvar == 'laps':
mask = xs >= 1.0
xs = xs[mask]
ys = ys[mask]
if args.traceEvery is not None:
mask = bnpy.util.isEvenlyDivisibleFloat(xs, args.traceEvery)
xs = xs[mask]
ys = ys[mask]
plotargs = dict(markersize=10, linewidth=2, label=None,
color=color, markeredgecolor=color)
if tt == 0:
plotargs['label'] = jobname
pylab.plot(xs, ys, '.-', **plotargs)
if len(ys) > 0:
xLocs.append(xs[-1])
yLocs.append(ys[-1])
yAll.extend(ys[1:])
xAll.extend(xs[1:])
# Zoom in to the useful part of the ELBO trace
if len(yAll) > 0:
global YMin, YMax
ymin = np.percentile(yAll, 1)
ymax = np.max(yAll)
if YMin is None:
YMin = ymin
YMax = ymax
else:
YMin = np.minimum(ymin, YMin)
YMax = np.maximum(YMax, ymax)
blankmargin = 0.08*(YMax - YMin)
pylab.ylim( [YMin, YMax + blankmargin])
pylab.xlabel(XLabelMap[args.xvar])
pylab.ylabel('log evidence')
def plot_all_tasks_for_job(jobpath, args, jobname=None, color=None):
''' Create line plot in current matplotlib figure
for each task/run of the designated jobpath
'''
if not os.path.exists(jobpath):
raise ValueError("No such path: %s" % (jobpath))
taskids = BNPYArgParser.parse_task_ids(jobpath, args.taskids)
xAll = list()
yAll = list()
xLocs = list()
yLocs = list()
for tt, taskid in enumerate(taskids):
xs = np.loadtxt(os.path.join(jobpath, taskid, args.xvar+'.txt'))
try:
ys = np.loadtxt(os.path.join(jobpath, taskid, 'K.txt'))
except IOError:
MatDict = scipy.io.loadmat(os.path.join(jobpath,taskid, 'AllocPrior.mat'))
Kfixed = int(MatDict['K'])
ys = Kfixed* np.ones(len(xs))
if args.traceEvery is not None:
mask = bnpy.util.isEvenlyDivisibleFloat(xs, args.traceEvery)
xs = xs[mask]
ys = ys[mask]
plotargs = dict(markersize=10, linewidth=2, label=None,
color=color, markeredgecolor=color)
if tt == 0:
plotargs['label'] = jobname
pylab.plot(xs, ys, '.-', **plotargs)
if len(ys) > 0:
xLocs.append(xs[-1])
yLocs.append(ys[-1])
yAll.extend(ys[1:])
xAll.extend(xs[1:])
# Zoom in to the useful part of the ELBO trace
if len(yAll) > 0:
global YMax
ymax = np.max(yAll)
if YMax is None:
YMax = ymax
else:
YMax = np.maximum(YMax, ymax)
blankmargin = 0.05*(YMax)
pylab.ylim( [0, YMax + blankmargin])
pylab.xlabel(XLabelMap[args.xvar])
pylab.ylabel('K')
def parse_jobpath_and_taskids(args):
rootpath = os.path.join(os.environ['BNPYOUTDIR'], args.dataName,
args.allocModelName, args.obsModelName)
jobpath = os.path.join(rootpath, args.algNames, args.jobnames)
if not os.path.exists(jobpath):
raise ValueError("No such path: %s" % (jobpath))
taskids = BNPYArgParser.parse_task_ids(jobpath, args.taskids)
# Verify that the intended savefile will work as expected!
if args.savefilename is not None:
if args.savefilename.count('%') and len(taskids) > 1:
try:
args.savefilename % ('1')
except TypeError:
raise ValueError("Missing or bad format string in savefilename %s" %
(args.savefilename)
)
return jobpath, taskids
def plotCompsForJob(jobpath='', taskids=[1], lap=None, **kwargs):
''' Show plot of learned clusters from run(s) saved results on disk
'''
# Verify given absolute path is valid.
jobpath_originalarg = jobpath
if not os.path.isdir(jobpath):
# Fallback: try to prepend BNPYOUTDIR to handle "shortcut" names
jobpath = os.path.join(os.environ['BNPYOUTDIR'], jobpath)
if not os.path.isdir(jobpath):
raise ValueError('Not valid path: ' + jobpath_originalarg)
taskids = BNPYArgParser.parse_task_ids(jobpath, taskids)
for tt, taskid in enumerate(taskids):
if tt == 0 and isinstance(taskid, str):
if taskid.startswith('.'):
rankTasksForSingleJobOnDisk(jobpath)
taskpath = os.path.join(jobpath, str(taskid))
plotCompsForTask(taskpath, lap=lap, **kwargs)
if 'block' in kwargs:
pylab.show(block=kwargs['block'])
legNames = ['%s=%s' % (plotkey, x) for x in RangeMap[plotkey]]
# Build list of final jpaths in order of decided legend
keepListFinal = list()
for x in RangeMap[plotkey]:
for jID, jdict in enumerate(keepListD):
if jdict[plotkey] == x:
keepListFinal.append(keepListP[jID])
else:
keepListFinal = keepListP[:1]
legNames = [None]
if verbose:
print('\nLegend entries for selected jobs (auto-selected)')
for name in legNames:
print(name)
return keepListFinal, legNames
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('dataName', default='AsteriskK8')
parser.add_argument('jobName', default='bm')
args, unkList = parser.parse_known_args()
reqDict = BNPYArgParser.arglist_to_kwargs(unkList, doConvertFromStr=False)
jpath = os.path.join(os.environ['BNPYOUTDIR'], args.dataName, args.jobName)
keepJobs, legNames = filterJobs(jpath, verbose=1, **reqDict)
def plot_all_tasks_for_job(jobpath, label, taskids=None,
lineType='.-',
spreadLineType='--',
color=None,
yvar='avgLikScore',
xvar='laps',
markersize=10,
linewidth=2,
minLap=0,
showFinalPt=0,
fileSuffix='PredLik.mat',
xjitter=None,
prefix='predlik',
colorID=0,
**kwargs):
''' Create line plot in current figure for each task/run of jobpath
'''
if not os.path.exists(jobpath):
print('PATH NOT FOUND', jobpath)
return None
if not yvar.startswith('avg') and yvar.count('Kactive') == 0:
yvar = 'avg' + yvar
if not yvar.endswith('Score') and yvar.count('Kactive') == 0:
yvar = yvar + 'Score'
if color is None:
color = Colors[colorID % len(Colors)]
taskids = BNPYArgParser.parse_task_ids(jobpath, taskids)
for tt, taskid in enumerate(taskids):
taskoutpath = os.path.join(jobpath, taskid)
hpaths = glob.glob(os.path.join(taskoutpath, '*' + fileSuffix))
txtpaths = glob.glob(os.path.join(taskoutpath, 'predlik-*.txt'))
ys_hi = None
ys_lo = None
if len(txtpaths) > 0:
if fileSuffix.endswith('.txt'):
suffix = '-' + fileSuffix
else:
suffix = '.txt'
if xvar.count('lap'):
xs = np.loadtxt(
os.path.join(taskoutpath, prefix + '-lapTrain.txt'))
elif xvar.count('K'):
xs = np.loadtxt(os.path.join(taskoutpath, prefix + '-K.txt'))
elif xvar.count('time'):
xs = np.loadtxt(os.path.join(
taskoutpath, prefix + '-timeTrain.txt'))
else:
raise ValueError("Unrecognized xvar: " + xvar)
if yvar.count('Kactive') and not yvar.count('Percentile'):
ys = np.loadtxt(os.path.join(taskoutpath,
prefix + '-' + yvar + 'Percentile50.txt'))
ys_lo = np.loadtxt(os.path.join(taskoutpath,
prefix + '-' + yvar + 'Percentile10.txt'))
ys_hi = np.loadtxt(os.path.join(taskoutpath,
prefix + '-' + yvar + 'Percentile90.txt'))
else:
ys = np.loadtxt(
os.path.join(taskoutpath, prefix + '-' + yvar + suffix))
if minLap > 0 and taskoutpath.count('fix'):
mask = laps > minLap
xs = xs[mask]
ys = ys[mask]
elif len(hpaths) > 0:
hpaths.sort()
basenames = [x.split(os.path.sep)[-1] for x in hpaths]
xs = np.asarray([float(x[3:11]) for x in basenames])
ys = np.zeros_like(xs)
for ii, hpath in enumerate(hpaths):
MatVars = scipy.io.loadmat(hpath)
ys[ii] = float(MatVars['avgPredLL'])
else:
raise ValueError(
'Pred Lik data unavailable for job\n' + taskoutpath)
plotargs = dict(markersize=markersize, linewidth=linewidth, label=None,
color=color, markeredgecolor=color,
)
plotargs.update(kwargs)
if tt == 0:
plotargs['label'] = label
if xjitter is not None:
xs = xs + xjitter
pylab.plot(xs, ys, lineType, **plotargs)
if ys_lo is not None:
del plotargs['label']
pylab.plot(xs, ys_lo, spreadLineType, **plotargs)
pylab.plot(xs, ys_hi, spreadLineType, **plotargs)
if showFinalPt:
pylab.plot(xs[-1], ys[-1], '.', **plotargs)
pylab.xlabel(XLabelMap[xvar])
pylab.ylabel(YLabelMap[yvar])
def run(dataName=None, allocModelName=None, obsModelName=None, algName=None, \
doSaveToDisk=True, doWriteStdOut=True,
taskID=None, **kwargs):
''' Fit specified model to data with learning algorithm.
Usage
-------
To fit a Gauss MixModel to a custom dataset defined in matrix X
>> Data = bnpy.data.XData(X)
>> hmodel = run(Data, 'MixModel', 'Gauss', 'EM', K=3, nLap=10)
To load a dataset specified in a specific script
For example, 2D toy data in demodata/AsteriskK8.py
>> hmodel = run('AsteriskK8', 'MixModel', 'Gauss', 'VB', K=3)
To run 5 tasks (separate initializations) and get best of 5 runs:
>> opts = dict(K=8, nLap=100, printEvery=0)
>> hmodel = run('AsteriskK8','MixModel','Gauss','VB', nTask=5, **opts)
Args
-------
dataName : either one of
* bnpy Data object,
* string filesystem path of Data module within BNPYDATADIR
allocModelName : string name of allocation (latent structure) model
{MixModel, DPMixModel, AdmixModel, HMM, etc.}
obsModelName : string name of observation (likelihood) model
{Gauss, ZMGauss, WordCount, etc.}
**kwargs : keyword args defining properties of the model or alg
see Doc for details [TODO]
Returns
-------
hmodel : best model fit to the dataset (across nTask runs)
LP : local parameters of that best model on the dataset
evBound : log evidence (ELBO) for the best model on the dataset
scalar, real value where larger value implies better model
'''
hasReqArgs = dataName is not None
hasReqArgs &= allocModelName is not None
hasReqArgs &= obsModelName is not None
hasReqArgs &= algName is not None
if hasReqArgs:
ReqArgs = dict(dataName=dataName,
allocModelName=allocModelName,
obsModelName=obsModelName,
algName=algName)
else:
ReqArgs = BNPYArgParser.parseRequiredArgs()
dataName = ReqArgs['dataName']
allocModelName = ReqArgs['allocModelName']
obsModelName = ReqArgs['obsModelName']
algName = ReqArgs['algName']
KwArgs, UnkArgs = BNPYArgParser.parseKeywordArgs(ReqArgs, **kwargs)
jobname = KwArgs['OutputPrefs']['jobname']
if taskID is None:
starttaskid = KwArgs['OutputPrefs']['taskid']
else:
starttaskid = taskID
KwArgs['OutputPrefs']['taskid'] = taskID
nTask = KwArgs['OutputPrefs']['nTask']
bestInfo = None
bestEvBound = -np.inf
for taskid in range(starttaskid, starttaskid + nTask):
hmodel, LP, Info = _run_task_internal(jobname, taskid, nTask, ReqArgs,
KwArgs, UnkArgs, dataName,
allocModelName, obsModelName,
algName, doSaveToDisk,
doWriteStdOut)
if (Info['evBound'] > bestEvBound):
bestModel = hmodel
bestLP = LP
bestEvBound = Info['evBound']
bestInfo = Info
return bestModel, bestLP, bestInfo
def plotSingleJob(dataName,
jobname,
taskids='1',
lap=None,
showELBOInTitle=True,
cmap='gray',
title='',
mixZs=False):
''' Visualize results of single run
'''
# Parse the jobpath, and create example task paths
jobpath = os.path.join(os.path.expandvars('$BNPYOUTDIR'), dataName,
jobname)
if isinstance(taskids, str):
taskids = BNPYArgParser.parse_task_ids(jobpath, taskids)
elif isinstance(taskids, int):
taskids = [str(taskids)]
taskpath = os.path.join(jobpath, taskids[0])
# Load data, with same dataset size prefs as specified at inference time.
dataKwargs = bnpy.ioutil.DataReader.loadDataKwargsFromDisk(taskpath)
Data = bnpy.ioutil.DataReader.loadDataFromSavedTask(taskpath)
AdjMat = np.squeeze(Data.toAdjacencyMatrix())
if hasattr(Data, 'TrueParams'):
if 'nodeZ' in Data.TrueParams:
sortids = np.argsort(Data.TrueParams['nodeZ'])
print 'Sorting nodes by true labels...'
elif 'pi' in Data.TrueParams:
sortids = np.argsort(Data.TrueParams['pi'].argmax(axis=1))
else:
sortids = np.arange(AdjMaj.shape[0])
# Rearrange the rows/cols of AdjMat
AdjMat = AdjMat[sortids, :]
AdjMat = AdjMat[:, sortids]
if hasattr(Data, 'nodeNames'):
nodeNames = [Data.nodeNames[s] for s in sortids]
else:
nodeNames = None
# Show the true adj mat and the estimated side-by-side
# First, the true adjacency matrix
ncols = len(taskids) + 1
pylab.subplots(nrows=1, ncols=ncols, figsize=(3 * ncols, 3))
pylab.subplot(1, ncols, 1)
pylab.imshow(AdjMat, cmap='Greys', interpolation='nearest', vmin=0, vmax=1)
if len(nodeNames) < 25:
pylab.gca().set_yticks(np.arange(len(nodeNames)))
pylab.gca().set_yticklabels(nodeNames)
for tt, taskid in enumerate(taskids):
taskoutpath = os.path.join(jobpath, taskid) + os.path.sep
# Load the model for the current task at specified lap
hmodel, curLap = bnpy.ioutil.ModelReader.loadModelForLap(
taskoutpath, lap)
# Compute expected state-state edge prob matrix Ew
Ew = hmodel.obsModel.Post.lam1 / \
(hmodel.obsModel.Post.lam1 + hmodel.obsModel.Post.lam0)
isAssortative = str(type(hmodel.allocModel)).count('Assort')
if isAssortative:
K = hmodel.allocModel.K
Ew_tmp = hmodel.allocModel.epsilon * np.ones((K, K, Ew.shape[-1]))
for k in xrange(K):
Ew_tmp[k, k] = Ew[k]
Ew = Ew_tmp
taskAdjMat = np.zeros((Data.nNodes, Data.nNodes, Data.dim))
useLP = 0
if useLP:
LP = hmodel.calc_local_params(Data)
for eid, (s, t) in enumerate(Data.edges):
resp_st = LP['resp'][eid]
if isAssortative:
taskAdjMat[s, t] = np.sum(resp_st[:, np.newaxis] * Ew,
axis=0)
else:
assert np.allclose(resp_st.sum(), 1.0)
taskAdjMat[s, t] = np.sum(resp_st[:, :, np.newaxis] * Ew,
axis=(0, 1))
else:
Epi = np.exp(hmodel.allocModel.E_logPi())
for eid, (s, t) in enumerate(Data.edges):
for d in xrange(Data.dim):
taskAdjMat[s, t,
d] = np.inner(Epi[s, :],
np.dot(Ew[:, :, d], Epi[t, :]))
assert taskAdjMat.min() >= 0
assert taskAdjMat.max() <= 1.0
taskAdjMat = np.squeeze(taskAdjMat)
taskAdjMat = taskAdjMat[sortids, :]
taskAdjMat = taskAdjMat[:, sortids]
pylab.subplot(1, ncols, 2 + tt)
pylab.imshow(taskAdjMat,
cmap='Greys',
interpolation='nearest',
vmin=0,
vmax=1)
def plotSingleJob(
dataset,
jobname,
taskids='1',
lap='final',
sequences=[1],
showELBOInTitle=False,
dispTrue=True,
aspectFactor=4.0,
specialStateIDs=None,
seqNames=None,
cmap='Set1',
maxT=None,
colorManyToOne=False,
):
'''
Returns the array of Data corresponding to a single sequence to display
If dispTrue = True, the true labels will be shown underneath the
estimated labels
'''
# Make sequences zero-indexed
if isinstance(sequences, str):
sequences = np.asarray([int(x) for x in args.sequences.split(',')],
dtype=np.int32)
sequences = np.asarray(sequences, dtype=np.int32)
if np.min(sequences) < 1:
raise ValueError('Sequences need to be one-index.\n' +
'Valid values are 1,2,...N.')
sequences -= 1
# Determine the jobpath and taskids
jobpath = os.path.join(os.path.expandvars('$BNPYOUTDIR'), dataset, jobname)
if isinstance(taskids, str):
if taskids.startswith('.'):
taskids = [taskids]
else:
taskids = BNPYArgParser.parse_task_ids(jobpath, taskids)
elif isinstance(taskids, int):
taskids = [str(taskids)]
datasetPrefFile = os.path.join(jobpath, taskids[0],
'args-DatasetPrefs.txt')
datasetPrefs = dict()
if os.path.exists(datasetPrefFile):
with open(datasetPrefFile, 'r') as f:
for line in f.readlines():
fields = line.strip().split(' ')
if len(fields) != 2:
continue
datasetPrefs[fields[0]] = fields[1]
# Load Data from its python module
Datamod = imp.load_source(
dataset, os.path.expandvars('$BNPYDATADIR/' + dataset + '.py'))
if dataset == 'SpeakerDiar':
if len(sequences) > 1:
raise ValueError(
'Joint modeling of several sequences makes no sense')
Data = Datamod.get_data(meetingNum=sequences[0] + 1, **datasetPrefs)
jobpath = jobpath.replace('SpeakerDiar',
'SpeakerDiar' + str(sequences[0] + 1))
sequences[0] = 0
else:
Data = Datamod.get_data(**datasetPrefs)
# Determine the maximum length among any of the sequences to be plotted
if maxT is None:
Ts = Data.doc_range[sequences + 1] - Data.doc_range[sequences]
maxT = np.max(Ts)
# Define the number of pixels used by vertical space of figure
NUM_STACK = int(np.ceil(maxT / float(aspectFactor)))
if dispTrue:
NUM_STACK /= 2
f, axes = plt.subplots(len(sequences),
len(taskids),
sharex='col',
sharey='row')
# For singleton case, make sure that axes is index-able
if len(sequences) == 1 and len(taskids) == 1:
axes = [axes]
for tt, taskidstr in enumerate(taskids):
if tt == 0 and taskidstr.startswith('.'):
rankTasksForSingleJobOnDisk(jobpath)
path = os.path.join(jobpath, taskidstr) + os.path.sep
# Figure out which lap to use
if lap == 'final':
lapsFile = open(path + 'laps-saved-params.txt')
curLap = lapsFile.readlines()
curLap = float(curLap[-1])
lapsFile.close()
else:
curLap = int(lap)
if showELBOInTitle:
hdists = np.loadtxt(os.path.join(path, 'hamming-distance.txt'))
hlaps = np.loadtxt(os.path.join(path, 'laps-saved-params.txt'))
Keffvals = np.loadtxt(os.path.join(path, 'Keff-saved-params.txt'))
# Determine scalar values to display
loc = np.argmin(np.abs(hlaps - curLap))
hdist = hdists[loc]
Kefffinal = Keffvals[loc]
try:
Kvals = np.loadtxt(os.path.join(path, 'K.txt'))
ELBOscores = np.loadtxt(os.path.join(path, 'evidence.txt'))
laps = np.loadtxt(os.path.join(path, 'laps.txt'))
loc = np.argmin(np.abs(laps - curLap))
ELBO = ELBOscores[loc]
Kfinal = Kvals[loc]
except IOError:
ELBO = 0.0
Kfinal = Kefffinal
# Load in the saved Data from $BNPYOUTDIR
try:
filename = 'Lap%08.3fMAPStateSeqsAligned.mat' % curLap
zHatBySeq = scipy.io.loadmat(path + filename)
key1 = 'zHatBySeqAligned'
key2 = 'zHatBySeq'
if key1 in zHatBySeq:
zHatBySeq = convertStateSeq_MAT2list(zHatBySeq[key1])
elif key2 in zHatBySeq:
zHatBySeq = convertStateSeq_MAT2list(zHatBySeq[key2])
else:
raise IOError
except IOError:
filename = 'Lap%08.3fMAPStateSeqs.mat' % curLap
zHatBySeq = scipy.io.loadmat(path + filename)
zHatBySeq = convertStateSeq_MAT2list(zHatBySeq['zHatBySeq'])
if specialStateIDs is not None:
zHatBySeq = relabelAllSequences(zHatBySeq, specialStateIDs)
# Find maximum number of states we need to display
nSeq = len(zHatBySeq)
Kmax = np.max([zHatBySeq[i].max() for i in xrange(nSeq)])
hasGroundTruth = False
vmin = 0
Kignore = 0
if hasattr(Data, 'TrueParams') and 'Z' in Data.TrueParams:
hasGroundTruth = True
Kmax = np.maximum(Data.TrueParams['Z'].max(), Kmax)
uLabels = np.unique(Data.TrueParams['Z'])
Kignore = np.sum(uLabels < 0)
if Kignore > 0:
for k in range(1, Kignore + 1):
print 'ignoring state %d Ttrue = %d' % (
-k, np.sum(Data.TrueParams['Z'] == -k))
if colorManyToOne:
# For each state in zHat, find best true sequence
Zflat = convertStateSeq_list2flat(zHatBySeq, Data)
ZflatA = -1 * np.ones_like(Zflat)
for uID in np.unique(Zflat):
overlap = np.zeros(uLabels.size)
for ii, trueID in enumerate(uLabels):
overlap[ii] = np.sum(
np.logical_and(Data.TrueParams['Z'] == trueID,
Zflat == uID))
bestii = overlap.argmax()
ZflatA[Zflat == uID] = uLabels[bestii]
zHatBySeq = convertStateSeq_flat2list(ZflatA, Data)
# In case there's only one sequence, make sure it's index-able
for ii, seqNum in enumerate(sequences):
image = np.tile(zHatBySeq[seqNum], (NUM_STACK, 1))
# Add the true labels to the image (if they exist)
if hasGroundTruth and dispTrue:
start = Data.doc_range[seqNum]
stop = Data.doc_range[seqNum + 1]
img_trueZ = np.tile(Data.TrueParams['Z'][start:stop],
(NUM_STACK, 1))
if dispTrue == 2:
image = img_trueZ # Show only true labels
else:
image = np.vstack((image, img_trueZ))
image = image[:, :maxT]
if len(sequences) == 1 or len(taskids) == 1:
cur_ax = axes[ii + tt]
else:
cur_ax = axes[ii, tt]
if hasattr(cmap, 'N'):
vmax = cmap.N
else:
vmax = Kmax
cur_ax.imshow(Kignore + image + .0001,
interpolation='nearest',
vmin=vmin,
vmax=vmax,
cmap=cmap)
if tt == 0:
if seqNames is not None:
h = cur_ax.set_ylabel('%s' % (seqNames[ii]), fontsize=13)
h.set_rotation(0)
elif len(sequences) > 4:
cur_ax.set_ylabel('%d' % (seqNum + 1), fontsize=13)
else:
cur_ax.set_ylabel('Seq. %d' % (seqNum + 1), fontsize=13)
if ii == 0:
if showELBOInTitle:
fmtSpec = "ELBO: %.3f K=%d Keff=%d "
if hdist > 0.01:
fmtSpec += "dist=%.2f"
elif hdist > 0.001:
fmtSpec += "dist=%.3f"
else:
fmtSpec += "dist=%.4f"
title = fmtSpec % (ELBO, Kfinal, Kefffinal, hdist)
cur_ax.set_title(title)
cur_ax.set_xlim([0, maxT])
cur_ax.set_ylim([0, image.shape[0]])
cur_ax.set_yticks([])
# ... end loop over sequences
return axes, zHatBySeq
def plotSingleLineAcrossJobsByXVar(jpathPattern,
label='',
xvar=None,
xvals=None,
xlabel=None,
yvar='evidence',
lineStyle='.-',
taskids='all',
lineID=0,
lvar='',
**kwargs):
''' Create line plot in current figure for job matching the pattern
Iterates over each xval in provided list of values.
Each one corresponds to a single saved job.
Post Condition
--------------
Current axes have one line added.
'''
prefixfilepath = os.path.sep.join(jpathPattern.split(os.path.sep)[:-1])
PPListMap = makePPListMapFromJPattern(jpathPattern)
if xvals is None:
xvals = PPListMap[xvar]
xs = np.zeros(len(xvals))
ys = np.zeros(len(xvals))
jpathList = makeListOfJPatternsWithSpecificVals(
PPListMap,
prefixfilepath=prefixfilepath,
key=xvar,
vals=xvals,
**kwargs)
plotargs = copy.deepcopy(DefaultLinePlotKwArgs)
# Plot all tasks as faint points with no connections
for i, jobpath in enumerate(jpathList):
if not os.path.exists(jobpath):
raise ValueError("PATH NOT FOUND: %s" % (jobpath))
x = float(xvals[i])
for key in plotargs:
if key in kwargs:
plotargs[key] = kwargs[key]
plotargs['markeredgecolor'] = plotargs['color']
alltaskids = BNPYArgParser.parse_task_ids(jobpath, taskids)
for tid in alltaskids:
y = loadYValFromDisk(jobpath, tid, yvar=yvar)
pylab.plot(x, y, '.', **plotargs)
# Plot top-ranked tasks as solid points connected by line
for i, jobpath in enumerate(jpathList):
rankTasksForSingleJobOnDisk(os.path.join(jobpath))
x = float(xvals[i])
y = loadYValFromDisk(jobpath, '.best', yvar=yvar)
assert isinstance(x, float)
assert isinstance(y, float)
xs[i] = x
ys[i] = y
plotargs = copy.deepcopy(DefaultLinePlotKwArgs)
for key in plotargs:
if key in kwargs:
plotargs[key] = kwargs[key]
plotargs['markeredgecolor'] = plotargs['color']
plotargs['label'] = label
pylab.plot(xs, ys, lineStyle, **plotargs)
if lineID == 0:
if xlabel is None:
xlabel = xvar
pylab.xlabel(xlabel)
pylab.ylabel(LabelMap[yvar])
def run(dataName=None,
allocModelName=None,
obsModelName=None,
algName=None,
doSaveToDisk=True,
doWriteStdOut=True,
taskID=None,
**kwargs):
""" Fit specified model to data with learning algorithm.
Args
-------
dataName : either one of
* bnpy Data object,
* string name of python file within BNPYDATADIR
allocModelName : string name of allocation (latent structure) model
obsModelName : string name of observation (likelihood) model
**kwargs : keyword args defining properties of the model or alg
Returns
-------
hmodel : best model fit to the dataset (across nTask runs)
Info : dict of information about this best model
"""
hasReqArgs = dataName is not None
hasReqArgs &= allocModelName is not None
hasReqArgs &= obsModelName is not None
hasReqArgs &= algName is not None
if hasReqArgs:
ReqArgs = dict(dataName=dataName,
allocModelName=allocModelName,
obsModelName=obsModelName,
algName=algName)
else:
ReqArgs = BNPYArgParser.parseRequiredArgs()
dataName = ReqArgs['dataName']
allocModelName = ReqArgs['allocModelName']
obsModelName = ReqArgs['obsModelName']
algName = ReqArgs['algName']
KwArgs, UnkArgs = BNPYArgParser.parseKeywordArgs(ReqArgs, **kwargs)
KwArgs['OutputPrefs']['doSaveToDisk'] = doSaveToDisk
KwArgs['OutputPrefs']['doWriteStdOut'] = doWriteStdOut
jobname = KwArgs['OutputPrefs']['jobname']
# Update stored numerical options via keyword args
bnpy.util.NumericUtil.UpdateConfig(**UnkArgs)
if taskID is None:
starttaskid = KwArgs['OutputPrefs']['taskid']
else:
starttaskid = taskID
KwArgs['OutputPrefs']['taskid'] = taskID
nTask = KwArgs['OutputPrefs']['nTask']
best_info_dict = None
best_loss = np.inf
for taskid in range(starttaskid, starttaskid + nTask):
hmodel, info_dict = _run_task_internal(jobname, taskid, nTask, ReqArgs,
KwArgs, UnkArgs, dataName,
allocModelName, obsModelName,
algName, doSaveToDisk,
doWriteStdOut)
if (taskid == starttaskid or info_dict['loss'] < best_loss):
bestModel = hmodel
best_loss = info_dict['loss']
best_info_dict = info_dict
return bestModel, best_info_dict
def run(dataName=None, allocModelName=None, obsModelName=None, algName=None, \
doSaveToDisk=True, doWriteStdOut=True,
taskID=None, **kwargs):
''' Fit specified model to data with learning algorithm.
Usage
-------
To fit a Gauss MixModel to a custom dataset defined in matrix X
>> Data = bnpy.data.XData(X)
>> hmodel = run(Data, 'MixModel', 'Gauss', 'EM', K=3, nLap=10)
To load a dataset specified in a specific script
For example, 2D toy data in demodata/AsteriskK8.py
>> hmodel = run('AsteriskK8', 'MixModel', 'Gauss', 'VB', K=3)
To run 5 tasks (separate initializations) and get best of 5 runs:
>> opts = dict(K=8, nLap=100, printEvery=0)
>> hmodel = run('AsteriskK8','MixModel','Gauss','VB', nTask=5, **opts)
Args
-------
dataName : either one of
* bnpy Data object,
* string filesystem path of Data module within BNPYDATADIR
allocModelName : string name of allocation (latent structure) model
{MixModel, DPMixModel, AdmixModel, HMM, etc.}
obsModelName : string name of observation (likelihood) model
{Gauss, ZMGauss, WordCount, etc.}
**kwargs : keyword args defining properties of the model or alg
see Doc for details [TODO]
Returns
-------
hmodel : best model fit to the dataset (across nTask runs)
LP : local parameters of that best model on the dataset
evBound : log evidence (ELBO) for the best model on the dataset
scalar, real value where larger value implies better model
'''
hasReqArgs = dataName is not None
hasReqArgs &= allocModelName is not None
hasReqArgs &= obsModelName is not None
hasReqArgs &= algName is not None
if hasReqArgs:
ReqArgs = dict(dataName=dataName, allocModelName=allocModelName,
obsModelName=obsModelName, algName=algName)
else:
ReqArgs = BNPYArgParser.parseRequiredArgs()
dataName = ReqArgs['dataName']
allocModelName = ReqArgs['allocModelName']
obsModelName = ReqArgs['obsModelName']
algName = ReqArgs['algName']
KwArgs, UnkArgs = BNPYArgParser.parseKeywordArgs(ReqArgs, **kwargs)
jobname = KwArgs['OutputPrefs']['jobname']
if taskID is None:
starttaskid = KwArgs['OutputPrefs']['taskid']
else:
starttaskid = taskID
KwArgs['OutputPrefs']['taskid'] = taskID
nTask = KwArgs['OutputPrefs']['nTask']
bestInfo = None
bestEvBound = -np.inf
for taskid in range(starttaskid, starttaskid + nTask):
hmodel, LP, Info = _run_task_internal(jobname, taskid, nTask,
ReqArgs, KwArgs, UnkArgs,
dataName, allocModelName, obsModelName, algName,
doSaveToDisk, doWriteStdOut)
if (Info['evBound'] > bestEvBound):
bestModel = hmodel
bestLP = LP
bestEvBound = Info['evBound']
bestInfo = Info
return bestModel, bestLP, bestInfo
def plot_all_tasks_for_job(jobpath, label, taskids=None,
color=None,
colorID=0,
density=2,
yvar='evidence',
markersize=10,
linewidth=2,
linestyle='-',
drawLineToXMax=None,
showOnlyAfterLap=0,
xvar='laps',
**kwargs):
''' Create line plot in current figure for each task/run of jobpath
'''
if not os.path.exists(jobpath):
if not jobpath.startswith(os.path.sep):
jobpath_tmp = os.path.join(os.environ['BNPYOUTDIR'], jobpath)
if not os.path.exists(jobpath_tmp):
raise ValueError("PATH NOT FOUND: %s" % (jobpath))
jobpath = jobpath_tmp
if color is None:
color = Colors[colorID % len(Colors)]
taskids = BNPYArgParser.parse_task_ids(jobpath, taskids)
if yvar == 'hamming-distance':
yspfile = os.path.join(jobpath, taskids[0], yvar + '-saved-params.txt')
if xvar == 'laps' and os.path.isfile(yspfile):
xvar = 'laps-saved-params'
for tt, taskid in enumerate(taskids):
xs = None
ys = None
laps = None
try:
var_ext = ''
ytxtfile = os.path.join(jobpath, taskid, yvar + '.txt')
if not os.path.isfile(ytxtfile):
var_ext = '-saved-params'
ytxtfile = os.path.join(
jobpath, taskid, yvar + var_ext + '.txt')
ys = np.loadtxt(ytxtfile)
if ytxtfile.count('saved-params'):
laptxtfile = os.path.join(jobpath, taskid, 'laps-saved-params.txt')
else:
laptxtfile = os.path.join(jobpath, taskid, 'laps.txt')
except IOError as e:
# TODO: when is this code needed?
# xs, ys = loadXYFromTopicModelFiles(jobpath, taskid)
try:
if isinstance(xs, np.ndarray) and yvar.count('Keff'):
ys = loadKeffForTask(
os.path.join(jobpath, taskid), **kwargs)
assert xs.size == ys.size
else:
# Heldout metrics
xs, ys = loadXYFromTopicModelSummaryFiles(
jobpath, taskid, xvar=xvar, yvar=yvar)
if showOnlyAfterLap and showOnlyAfterLap > 0:
laps, _ = loadXYFromTopicModelSummaryFiles(
jobpath, taskid, xvar='laps', yvar=yvar)
except ValueError:
try:
xs, ys = loadXYFromTopicModelSummaryFiles(jobpath, taskid)
except ValueError:
raise e
if yvar == 'hamming-distance' or yvar == 'Keff':
if xvar == 'laps-saved-params':
# fix off-by-one error, if we save an extra dist on final lap
if xs.size == ys.size - 1:
ys = ys[:-1]
elif ys.size == xs.size - 1:
xs = xs[:-1] # fix off-by-one error, if we quit early
elif xs.size != ys.size:
# Try to subsample both time series at laps where they
# intersect
laps_x = np.loadtxt(os.path.join(jobpath, taskid, 'laps.txt'))
laps_y = np.loadtxt(os.path.join(jobpath, taskid,
'laps-saved-params.txt'))
assert xs.size == laps_x.size
if ys.size == laps_y.size - 1:
laps_y = laps_y[:-1]
xs = xs[np.in1d(laps_x, laps_y)]
ys = ys[np.in1d(laps_y, laps_x)]
if xs.size != ys.size:
raise ValueError('Dimension mismatch. len(xs)=%d, len(ys)=%d'
% (xs.size, ys.size))
# Cleanup laps data. Verify that it is sorted, with no collisions.
if xvar == 'laps':
diff = xs[1:] - xs[:-1]
goodIDs = np.flatnonzero(diff >= 0)
if len(goodIDs) < xs.size - 1:
print('WARNING: looks like multiple runs writing to this file!')
print(jobpath)
print('Task: ', taskid)
print(len(goodIDs), xs.size - 1)
xs = np.hstack([xs[goodIDs], xs[-1]])
ys = np.hstack([ys[goodIDs], ys[-1]])
if xvar == 'laps' and yvar == 'evidence':
mask = xs >= 1.0
xs = xs[mask]
ys = ys[mask]
elif showOnlyAfterLap:
# print "Filtering for data recorded at lap >= %s" % (
# showOnlyAfterLap)
if laps is None:
laps = np.loadtxt(laptxtfile)
mask = laps >= showOnlyAfterLap
xs = xs[mask]
ys = ys[mask]
# Force plot density (data points per lap) to desired specification
# This avoids making plots that have huge file sizes,
# due to too much content in the given display space
if xvar == 'laps' and xs.size > 20 and np.sum(xs > 5) > 10:
if (xs[-1] - xs[9]) != 0:
curDensity = (xs.size - 10) / (xs[-1] - xs[9])
else:
curDensity = density
while curDensity > density and xs.size > 11:
# Thin xs and ys data by a factor of 2
# while preserving the first 10 data points
xs = np.hstack([xs[:10], xs[10::2]])
ys = np.hstack([ys[:10], ys[10::2]])
curDensity = (xs.size - 10) / (xs[-1] - xs[9])
plotargs = dict(
markersize=markersize,
linewidth=linewidth,
linestyle=linestyle,
label=None,
color=color, markeredgecolor=color)
for key in kwargs:
if key in plotargs:
plotargs[key] = kwargs[key]
if tt == 0:
plotargs['label'] = label
pylab.plot(xs, ys, **plotargs)
if drawLineToXMax:
xs_dashed = np.asarray([xs[-1], drawLineToXMax])
ys_dashed = np.asarray([ys[-1], ys[-1]])
plotargs['label'] = None
pylab.plot(xs_dashed, ys_dashed, '--', **plotargs)
pylab.xlabel(LabelMap[xvar])
if yvar in LabelMap:
yLabelStr = LabelMap[yvar]
if yvar == 'Keff' and 'effCountThr' in kwargs:
effCountThr = float(kwargs['effCountThr'])
yLabelStr = yLabelStr + ' > %s' % (str(effCountThr))
pylab.ylabel(yLabelStr)
