def psychometric_fit(xValues, nTrials, nHits, constraints=None, alpha=0.05):
'''
Function moved from jaratoolbox/extrastats.py
Given performance for each value of parameter, estimate the curve.
This function uses psignifit (BootstrapInference)
xValues: 1-D array of size M #Numbers of frequencies
nHits: 1-D array of size M #Nnumber of hits each frequency
nTrials: 1-D array of size M #Nnumber of trials each frequency
Returns 4 values:
'''
import pypsignifit as psi
if constraints is None:
constraints = ('flat', 'Uniform(0,0.3)', 'Uniform(0,0.2)',
'Uniform(0,0.2)')
data = np.c_[xValues, nHits, nTrials]
session = psi.BootstrapInference(data,
sample=False,
priors=constraints,
nafc=1)
# session = psi.BayesInference(data,sample=False,priors=constraints,nafc=1)
# (pHit,confIntervals) = binofit(nHits,nTrials,alpha)
# return (session.estimate,pHit,confIntervals)
return session.estimate
def find_theta_and_thresh(xs, ys, zs, nsamples=500, thresh_val=0.75):
a = "Gauss(0,5)"
b = "Gauss(1,3)"
l = "Beta(1.5,12)"
data = np.array(zip(xs, ys, zs))
B = psi.BootstrapInference(data, priors=[a, b, l], nafc=2, sample=nsamples)
return np.mean(B.mcestimates, 0), Finv(thresh_val, B.estimate)
def find_theta_and_thresh(xs, ys, zs, thresh_val, nboots=0):
data = np.array(zip(xs, ys, zs))
a = "Gauss(0,5)"
b = "Gauss(1,3)"
# l = "Beta(2,20)"
l = "Beta(1.5,12)"
B = psi.BootstrapInference(data, priors=[a, b, l], nafc=2, sample=nboots)
thetas = B.mcestimates if B.mcestimates is not None else [B.estimate]
threshes = [Finv(thresh_val, th) for th in thetas]
mean_theta = np.mean(thetas, 0)
mean_thresh = Finv(thresh_val, mean_theta)
return mean_theta, mean_thresh, thetas, threshes
def integration_grid(data, run=1, dists=None):
data = np.array(data)
mprior, mmin, mmax = pfp.default_mid(data[:, 0])
wprior, wmin, wmax = pfp.default_width(data[:, 0])
lprior, lmin, lmax = pfp.default_lapse()
gprior, gmin, gmax = pfp.default_lapse()
priors = (mprior, wprior, lprior, gprior)
pdata, ppmf, pn = sfu.make_dataset_and_pmf(data, 1, "logistic", "mw0.1",
priors)
mapest = pf.BootstrapInference(data, priors, nafc=1).estimate
if run == 1:
gridsize = 9
mmin, mmax = bounds(mapest, pdata, ppmf, "m")
wmin, wmax = bounds(mapest, pdata, ppmf, "w")
lmin, lmax = bounds(mapest, pdata, ppmf, "lm")
gmin, gmax = bounds(mapest, pdata, ppmf, "gm")
grid = np.reshape(
np.mgrid[mmin:mmax:1j * gridsize, wmin:wmax:1j * gridsize,
lmin:lmax:1j * gridsize, gmin:gmax:1j * gridsize],
(4, -1))
elif run == 2:
f_m = dist2class(dists[0])
f_w = dist2class(dists[1])
f_l = dist2class(dists[2])
f_g = dist2class(dists[3])
grid = np.mgrid[.025:.975:7j, .025:.975:7j, .025:.975:7j, .025:.975:7j]
grid[0] = f_m.ppf(grid[0])
grid[1] = f_w.ppf(grid[1])
grid[2] = f_l.ppf(grid[2])
grid[3] = f_g.ppf(grid[3])
grid = np.reshape(grid, (4, -1))
gridsize = 7
post = np.reshape(
np.array(map(lambda prm: ppmf.neglpost(prm, pdata), grid.T)),
[gridsize] * pn) # negative log posterior
post = np.exp(-post) # posterior
grid = np.reshape(grid, [pn] + [gridsize] * pn)
d = [np.diff(grid[i], axis=i)[0].max() for i in xrange(pn)]
fx = [marginalize(post, d, i) for i in xrange(pn)]
x = []
for i in xrange(pn):
s = [i] + [0] * pn
s[i + 1] = slice(0, grid.shape[i + 1])
x.append(grid[s])
return x, fx, priors
def addData(self, response):
self.y.append(response)
self.i += 1
if i == self.nxInit * self.nInit:
B = psi.BootstrapInference(
self.y,
core='ab',
sigmoid='gauss',
priors=('unconstrained', 'unconstrained',
'Uniform(0.0399,0.0401)', 'Uniform(0.0399,0.0401)'),
nafc=1)
self.x = random.shuffle(
np.linspace(B.estimate - 4 * B.deviance,
B.estimate + 4 * B.deviance, nxPost))
def psychometric_fit(xValues, nTrials, nHits, constraints=None, alpha=0.05):
'''
Given performance for each value of parameter, estimate the curve.
This function uses psignifit (BootstrapInference)
http://psignifit.sourceforge.net/TUTORIAL_BOOTSTRAP.html
xValues: 1-D array of size M
nHits: 1-D array of size M
nTrials: 1-D array of size M
'''
import pypsignifit as psi
if constraints is None:
constraints = ( 'flat', 'Uniform(0,0.3)' ,'Uniform(0,0.2)', 'Uniform(0,0.2)')
data = np.c_[xValues,nHits,nTrials]
session = psi.BootstrapInference(data,sample=False, priors=constraints, nafc=1)
# session = psi.BayesInference(data,sample=False,priors=constraints,nafc=1)
# (pHit,confIntervals) = binofit(nHits,nTrials,alpha)
# return (session.estimate,pHit,confIntervals)
return session.estimate
def plot(self, fileName):
self.graph.clear()
self.graph.legend.items = []
self.graph.addLine(y=0.04)
self.graph.addLine(y=0.96)
(xDict, yDict, conditions) = readFile(fileName)
colors = ['F00', '0F0', '00F', 'AA0', '0AA', 'A0A']
for i in range(len(conditions)):
condition = conditions[i]
x = xDict[condition]
y = yDict[condition]
#print("c: {}\nx: {}\ny: {}\n".format(condition, x, y))
data = zip(x, y, [1] * len(x))
constraints = ('unconstrained', 'unconstrained',
'Uniform(0.0399,0.0401)', 'Uniform(0.0399,0.0401)')
B = psi.BootstrapInference(data,
core='ab',
sigmoid='gauss',
priors=constraints,
nafc=1)
print("est: {}, dev: {}".format(B.estimate, B.deviance))
self.graph.plot(x,
y,
pen=None,
symbolPen={'color': colors[i % 6]},
symbolSize=6,
antialias=True)
xx = np.linspace(min(x), max(x), 100)
yy = []
for p in xx:
yy.append(f(p, B.estimate[0], B.estimate[1]))
self.graph.plot(xx,
yy,
pen={
'color': colors[i % 6],
'width': 2
},
name="{}: {:.3f} +/- {:.3f}".format(
condition, B.estimate[0], B.estimate[1]))
nrefine = 2
print data
x, fx, priors = integration_grid(data)
print "x1 =", x
print "f1 =", fx
post = fit_posterior(fx, x)
for i in xrange(nrefine):
x, fx, priors = integration_grid(data, 2, post)
print post
print "x%d =" % (i + 1, ), x
print "f%d =" % (i + 1), fx
post = fit_posterior(fx, x)
f = [sfu.get_prior(p) for p in post]
mapest = pf.BootstrapInference(data, priors, core="mw0.1", nafc=1).estimate
pdata, ppmf, pn = sfu.make_dataset_and_pmf(data, 1, "logistic", "mw0.1",
priors)
samples = sample_importance_resample(post,
pdata,
ppmf,
nresample=600,
nsamples=6000)
rng = [(3, 7), (0, 6), (0, .5), (0, .5)]
hist_ax = [
pl.axes([.15 + .2 * i, .75 - .2 * i, .15, .15]) for i in xrange(4)
]
labels = ["m", "w", "lm", "gm"]
#plt.setp(pdots,ms=6,mec='k',mew=2,mfc='k')
plt.hold(True)
# -- Calculate and plot psychometric fit --
constraints = None
#constraints = ['Uniform({},{})'.format(logPossibleValues[0],logPossibleValues[-1]), 'unconstrained' ,'unconstrained', 'unconstrained']
#curveParams = extrastats.psychometric_fit(possibleValues,nTrialsEachValue, nHitsEachValue)
# -- Fit psy curve with psi.BoostrapInference object -- #
data = np.c_[logPossibleValues, nHitsEachValue, nTrialsEachValue]
# linear core 'ab': (x-a)/b; logistic sigmoid: 1/(1+np.exp(-(xval-alpha)/beta))
psyCurveInference = psi.BootstrapInference(data,
sample=False,
sigmoid='logistic',
core='ab',
priors=constraints,
nafc=1)
curveParams = psyCurveInference.estimate
deviance = psyCurveInference.deviance
predicted = psyCurveInference.predicted
(alpha, beta, lapse, guess) = list(curveParams)
xValues = logPossibleValues
xRange = xValues[-1] - xValues[1]
fitxval = np.linspace(xValues[0] - 0.2 * xRange,
xValues[-1] + 0.2 * xRange, 40)
fityval = psyCurveInference.evaluate(x=fitxval)
#fityval = extrastats.psychfun(fitxval,*curveParams)
plt.plot(fitxval,
targetFrequencyThisBlock = targetFrequency[trialsEachType[:,stimType]]
validThisBlock = valid[trialsEachType[:,stimType]]
choiceRightThisBlock = choiceRight[trialsEachType[:,stimType]]
(possibleValues,fractionHitsEachValue,ciHitsEachValue,nTrialsEachValue,nHitsEachValue)=\
behavioranalysis.calculate_psychometric(choiceRightThisBlock,targetFrequencyThisBlock,validThisBlock)
#plot psychometric curve
(pline, pcaps, pbars, pdots) = extraplots.plot_psychometric(1e-3*possibleValues,fractionHitsEachValue,
ciHitsEachValue,xTickPeriod=1)
plt.setp((pline, pcaps, pbars), color=FREQCOLORS[stimType])
plt.setp(pdots, mfc=FREQCOLORS[stimType], mec=FREQCOLORS[stimType])
allPline.append(pline)
curveLegends.append(stimLabels[stimType])
####fitting psychometric curve using BootstrapInference
data = zip(possibleValues,nHitsEachValue,nTrialsEachValue)
constraints=('flat','Uniform(0,0.3)' ,'Uniform(0,0.2)', 'Uniform(0,0.2)')
###parameters that may need to be optimized: priors,sigmoid and core functions
session = psi.BootstrapInference(data,sample=True, priors=constraints, nafc=1, sigmoid='logistic', core='ab')
fittedSessions.append(session)
plt.hold(1)
psi.psigniplot.plotMultiplePMFs(*fittedSessions)
ax2=plt.subplot(gs[thisAnimalPos+1,0])
for session in fittedSessions:
#session.sample()
psi.psigniplot.ParameterPlot(session)
plt.hold(1)
plt.show()
sys.stderr.write("\n")
for simulation in xrange(options.nsimulations):
sys.stderr.write("\nSimulation %d is running" % (simulation, ))
O = create_new_observer()
# print "\nO=",O
if not options.fixed_sequence:
random.shuffle(x)
data = O.DoAnExperiment(x, ntrials=options.blocksize)
print "\ndata =", data
print constraints
write_id_gen(outfile, simulation)
if nonparametric:
Bnpr = pypsignifit.BootstrapInference(data,
priors=constraints,
parametric=False,
**ana_kwargs)
if options.fixed_pmf:
Bnpr.estimate = OO.params
tic = time.time()
Bnpr.sample(options.nbootstrap)
toc = time.time()
count_npr += check_ci(O, Bnpr)
write_nonparametric(outfile, Bnpr, toc - tic)
print "Done npar"
if parametric:
Bpar = pypsignifit.BootstrapInference(data,
priors=constraints,
parametric=True,
**ana_kwargs)
