def getBestModel(filePath, dataPoints):
dot_pos = filePath.rfind('.')
basePath = filePath[:dot_pos]
ext = filePath[dot_pos + 1:]
modelDict = {}
for i in range(NUMBER_OF_MODELS_TO_CONSIDER):
if ext == 'py':
#print('Generating data')
synthDataGen.run(filePath, samples=dataPoints)
dataFile = basePath + '.csv'
ic = CaLingam.IC(dataFile, dataPoints)
model = ic.getModel()
dagStr = model.getEquivalencyKey()
if not dagStr in modelDict:
modelDict[dagStr] = []
modelDict[dagStr].append(model)
# Find model based on most common structure
mostCommon = None
mostCommonCount = 0
for modelList in modelDict.values():
if len(modelList) > mostCommonCount:
mostCommonCount = len(modelList)
mostCommon = modelList[0]
model = mostCommon
print('Most common model = ', mostCommonCount, '/',
NUMBER_OF_MODELS_TO_CONSIDER)
return model
def run(s1, s2):
global firstTime, CUM_DEPENDENCE
global SNR, SNR_DB, CUM_TIME
reset = True
if firstTime:
#print('std ratio = ', s1D.std() / s2D.std())
firstTime = False
else:
reset = False
try:
synthDataGen.run(FILE + '.py',
samples=DATA_POINTS,
maxDifficulty=MAX_DIFFICULTY,
reset=reset)
except:
pass
dr = getData.DataReader(input=FILE + '.csv', limit=DATA_POINTS)
s1D = np.array(dr.getSeries(s1))
s2D = np.array(dr.getSeries(s2))
dependence = IndHSIC.scoreDependence(s1D, s2D)
#dependence = IndLn.scoreDependence(s1D, s2D)
CUM_DEPENDENCE = CUM_DEPENDENCE + dependence
#print('s1 = ', s1D.mean(), s1D.std())
#print('s2 = ', s2D.mean(), s2D.std())
if firstTime:
#print('std ratio = ', s1D.std() / s2D.std())
firstTime = False
coefs = np.polyfit(s1D, s2D.T, 1)
x1coef, x0coef = coefs
Dprint('coefs = ', coefs)
cs2D = s1D * x1coef + x0coef
# coef = lstsq(np.array([list(s1D)]).T, s2D)[0][0]
# Dprint('coef = ', coef)
# cs2D = s1D * coef
res = s2D - cs2D
signal = cs2D.var(ddof=1)
noise = res.var(ddof=1)
snr = signal / noise
if snr < SNR:
SNR = snr
SNR_DB = 10 * math.log(snr, 10)
start = time.time()
if IND_TYPE == 'PL' or IND_TYPE == 'CCM':
dep = scoreDependence(s1D, s2D)
else:
dep = scoreDependence(s1D, res)
end = time.time()
duration = end - start
CUM_TIME += duration
Dprint('Residual Dependence for ', s1, '-->', s2, ' = ', dep)
return dep
def analyzeOneLingam(self, dataCount, gen=True, valData=None):
if gen:
synthDataGen.run(self.genFile, dataCount)
dr = getData.DataReader(self.testFile, dataCount)
vars = dr.getSeriesNames()
if valData is None:
valData = synthDataGen.getValidation()
lowestSNR = 10**100
cumSNR = 0
pairCount = 0
lowestPair = None
stats = {}
nodeDiffs = {}
for tuple in valData:
successor, predecessors = tuple
if len(predecessors) == 0:
continue
for predecessor in predecessors:
nodeDiffs[predecessor] = {}
pD = np.array(dr.getSeries(predecessor))
sD = np.array(dr.getSeries(successor))
coefs = np.polyfit(pD, sD.T, 1)
x1coef, x0coef = coefs
#Dprint('coefs = ', coefs)
cs2D = pD * x1coef + x0coef
res = sD - cs2D
# Note that signal and noise are reversed from what you might expect.
# In this case, the variance of the error term is the signal we are looking for,
# while the linear component is actually the noise
noise = cs2D.var(ddof=1)
signal = res.var(ddof=1)
#print('pair = ', predecessor, '--->', successor)
#print('noise = ', noise)
#print('signal = ', signal)
snr = signal / noise
#print('snr = ', snr)
nodeDiffs[predecessor][successor] = 10* math.log(1.0/snr, 10)
if snr < lowestSNR:
lowestSNR = snr
lowestPair = (predecessor, successor)
cumSNR += snr
pairCount += 1
stats['minSnr'] = 10*math.log(lowestSNR,10)
avgSNR = cumSNR/float(pairCount)
stats['avgSnr'] = 10*math.log(avgSNR, 10)
stats['weakestPair'] = lowestPair
difficulty = max([10 * math.log(1.0 / lowestSNR, 10),0.0])
stats['difficulty'] = difficulty
stats['weakestPairing'] = lowestPair
stats['variableDifficulty'] = nodeDiffs
stats['normDifficulty'] = 100.0 * difficulty / (dataCount**.5)
return stats
def run(dataDefFile, var1, var2, datapoints):
tau = TAU
print('Generating data using: ', dataDefFile)
fileExt = dataDefFile.split('.')[-1]
if fileExt == 'py':
# Data Definition File
outFile = synthDataGen.run(dataDefFile, datapoints)
elif fileExt == 'csv':
# Data file
outFile = dataDefFile
else:
print('*** Invalid file type = ', fileExt)
return
d = getData.DataReader(outFile, limit=datapoints)
fig = plt.figure()
ax = fig.add_axes([.1,.1,.8,.8],projection='3d')
# If var1 and var2 are not specified, then build a master manifold out of three vars at a time. Otherwise, build shadow manifolds
# from the two specified vars
if var1 is None:
vars = d.getSeriesNames()
vars.sort()
print('Vars = ', vars)
colors = ['b', 'g', 'r','o']
for i in range(4):
if len(vars) < i*3+3:
break
X = d.getSeries(vars[i*3])
Y = d.getSeries(vars[i*3+1])
Z = d.getSeries(vars[i*3+2])
if standard:
X = standardize.standardize(X)
Y = standardize.standardize(Y)
Z = standardize.standardize(Z)
color = colors[i]
ax.plot(X, Y, Z)
else:
var1D = d.getSeries(var1)
if standard:
var1D = standardize.standardize(var1D)
X1 = var1D[:-2*tau]
Y1 = var1D[tau:-tau]
Z1 = var1D[2*tau:]
var2D = d.getSeries(var2)
if standard:
var2D = standardize.standardize(var2D)
X2 = var2D[:-2*tau]
Y2 = var2D[tau:-tau]
Z2 = var2D[2*tau:]
#plotData(d, datapoints)
ax.plot(X1,Y1,Z1)
ax.plot(X2,Y2,Z2, 'r')
#ax.plot(X1, Y1, Z2, 'g')
# `ax` is a 3D-aware axis instance because of the projection='3d' keyword argument to add_subplot
#ax = fig.add_subplot(1, 2, 1, projection='3d')
plt.show()
def calibrateOneCITest(self, testType, filePath):
synthDataGen.run(filePath + '.py', samples=self.datacount)
exec('import ' + testType)
module = eval(testType)
SA = analyzeSEM.SemAnalyzer(filePath + '.py', self.datacount)
reader = getData.DataReader(filePath + '.csv', self.datacount)
dependencies, independencies = SA.getCondDependencies()
# print('dependencies = ', dependencies)
# print('independencies = ', independencies)
errors = 0
errorTerms = {}
items = 0
for item in dependencies:
x, y, z = item
X = reader.getSeries(x)
Y = reader.getSeries(y)
Z = reader.getSeries(z)
ind = module.isIndependent(X, Y, Z)
if ind:
print('Error -- ', x, 'and', y, 'Should be dependent given', z)
self.err1Count += 1
errors += 1
errorTerms[item] = 1
self.testCount += 1
for item in independencies:
x, y, z = item
X = reader.getSeries(x)
Y = reader.getSeries(y)
Z = reader.getSeries(z)
ind = module.isIndependent(X, Y, Z)
if not ind:
print('Error -- ', x, 'and', y, 'Should be independent given',
z)
self.err2Count += 1
errors += 1
errorTerms[item] = 1
self.testCount += 1
#print('Rating = ', (1 - (errors / items))*100, '%')
print('Errors for file: ', filePath, '=', errors,
list(errorTerms.keys()))
return
def run(dataDefFile, datapoints):
print('Generating data using: ', dataDefFile)
fileExt = dataDefFile.split('.')[-1]
if fileExt == 'py':
# Data Definition File
outFile = synthDataGen.run(dataDefFile, datapoints)
elif fileExt == 'csv':
# Data file
outFile = dataDefFile
else:
print('*** Invalid file type = ', fileExt)
return
d = getData.DataReader(outFile, limit=datapoints)
print('Vars = ', d.getSeriesNames())
plotData(d, datapoints)
def CaLingamTest(testType, paramSet):
global VALIDATION_ERRORS, CAUSAL_ORDERS
VALIDATION_ERRORS = {}
CAUSAL_ORDERS = {}
maxDifficulty = MAX_DIFFICULTY
fails = 0.0
datafile, runs, dataPoints, validation = paramSet[:4]
valData = None
totalDuration = 0
datafileRootName = datafile.split('.')[-2].split('\\')[-1]
#print('dfrn = ', datafileRootName)
sa = analyzeSEM.SemAnalyzer(datafileRootName)
difficulty = 0
diffNorm = 0
for i in range(runs):
if i == 0:
reset = True
if RESET_COUNT > 0 and i > 0 and i / RESET_COUNT == int(
i / RESET_COUNT):
reset = True
print()
print('Previous SEM:')
print(synthDataGen.getSEM())
#print('Resetting')
else:
reset = False
if i == 0:
reset = True
prune = True
if validation == 'SynthOrderVal':
# Suppress pruning in order to increase perf when only testing order.
prune = False
if testType == 'synth':
outFile = synthDataGen.run(datafile,
samples=dataPoints,
reset=reset,
maxDifficulty=maxDifficulty)
valData = synthDataGen.getValidation()
if i == 0 or reset:
saStats = sa.analyzeOne(dataPoints, gen=False, valData=valData)
difficulty = round(saStats['difficulty'], 2)
diffNorm = round(saStats['normDifficulty'], 2)
print('difficulty = ', difficulty, ', norm difficulty = ',
diffNorm)
elif testType == 'live':
outFile = datafile
else:
print('*** Invalid Test Type = ', testType)
return
startTime = time.time()
c = CaLingam.IC(outFile, limit=dataPoints, prune=prune)
dag = c.getDAG()
endTime = time.time()
duration = endTime - startTime
totalDuration += duration
if len(validation) > 0:
result = eval(validation + '(paramSet, dag, valData)')
else:
corder = str.join('||', dag.getVarOrder()[:5])
if not corder in CAUSAL_ORDERS:
CAUSAL_ORDERS[corder] = 1
#print('Causal Order = ', dag.getVarOrder())
# If we get a new causal order after the first time, it is considered an error
if len(CAUSAL_ORDERS) > 1:
result = 0
else:
result = 1
else:
# Got an existing causal order. Consider that success
count = CAUSAL_ORDERS[corder]
count += 1
CAUSAL_ORDERS[corder] = count
result = 1
if result:
#print ('Success', )
print('.', end='', flush=True)
else:
print('x', end='', flush=True)
fails += 1
print()
reliability = round(1.0 - (fails / float(runs)), 2)
#stats = 'Errors: ' + str(int(fails)) + ' / ' + str(int(i+1)) + ' -- Reliability: ' + str((1.0 - (fails / (i+1))) * 100) + '%'
stats = 'Errors: ' + str(int(fails)) + ' / ' + str(
int(runs)) + ' -- Reliability: ' + str(
reliability * 100) + '% avg duration: ' + str(
round(totalDuration / runs,
2)) + ' sec' + ' difficulty: ' + str(
difficulty) + ' diffNorm: ' + str(
diffNorm) + ' strength: ' + str(
round(reliability * diffNorm, 2))
print('Stats = ', stats)
if fails > 0:
if len(validation) > 0:
# Sort validation_errors to show the most common first
counts = []
keys = []
for key in VALIDATION_ERRORS.keys():
count = VALIDATION_ERRORS[key]
keys.append(key)
counts.append(count)
tuples = list(zip(counts, keys))
tuples.sort()
tuples.reverse()
keys = [key for (count, key) in tuples]
errStrs = []
for key in keys:
errStrs.append(key + ':' + str(VALIDATION_ERRORS[key]))
errStr = str.join(', ', errStrs)
print('ValidationErrors = ', errStr)
else:
maxKey = None
maxKeyCount = 0
totalKeyCount = 0
keys = CAUSAL_ORDERS.keys()
for key in keys:
keyCount = CAUSAL_ORDERS[key]
if keyCount > maxKeyCount:
maxKeyCount = keyCount
maxKey = key
totalKeyCount += keyCount
print('Most Common Order (', maxKeyCount / totalKeyCount * 100,
'%) = ', maxKey)
print()
print('CausalOrders = ', str(CAUSAL_ORDERS))
return