def search_for_start(r,
X,
w0,
applythreshold,
hf0,
pm=(.85, .93, 1., 1.07, 1.15),
storeopt=False):
"""Search for starting values
:Parameters:
*d*
data set
*w0*
coarse starting values
*pm*
increments or decrements
"""
logging.info("Initial preoptimization")
Mwh = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
Mnh = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
M0 = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
Mwh.w = np.concatenate((Mwh.w, np.zeros(6, 'd')))
Mwh.X = X
nhind = 0
whind = 0
i = 1
for al in pm:
for lm in pm:
logging.info("::::: Optimizing from starting value %d :::::" %
(i, ))
w0 = M0.w.copy()
w0[1:hf0] *= al
p0 = M0.pi.copy()
p0[0] *= lm
p0[-1] = 1 - p0[0] - p0[1]
M_ = model.history_model(r,
X,
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
if Mwh.loglikelihood < M_.loglikelihood:
logging.info(" *model chosen for history*")
Mwh = M_
whind = i
M_ = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
if Mnh.loglikelihood < M_.loglikelihood:
logging.info(" *model chosen for independent*")
Mnh = M_
nhind = i
i += 1
logging.info("Mwh.w = %s\nMnh.w = %s" % (str(Mwh.w), str(Mnh.w)))
logging.info("Mwh.ll = %g\nMnh.ll = %s" %
(Mwh.loglikelihood, Mnh.loglikelihood))
logging.info(
"Starting values:\n with history: %d\n without history: %d\n" %
(whind, nhind))
print "X", np.shape(X)
print "Mwh", Mwh.w, np.shape(Mwh.X)
return Mnh, Mwh
def mcsimulation(generator,
collector,
nsamples,
w0,
p0,
nu0=0.,
verbose=False,
hf0=2,
applythreshold=[]):
"""run a monte carlo simulation
:Parameters:
*generator*
a callable such that generator() will generate a new pair of
responses and design matrix
*collector*
a callable such that collector(fitted_model) will be everything
that is stored from the fitted dataset
*nsamples*
how many samples are to be generated
*w0*
starting values for the optimizations
*verbose*
if true, print status bar
:Returns:
a list of the collector outputs
:Example:
>>> np.random.seed(0)
>>> generator = lambda: ((np.random.rand(10)>0.5).astype('d'),np.c_[np.ones(10),np.random.uniform(-1,1,size=10)])
>>> collector = lambda M: (M.w,M.pi)
>>> sim_results = mcsimulation ( generator, collector, 3, [10,10] )
>>> len(sim_results)
3
>>> sim_results[0]
(array([ 0.89742561, 1.36421569]), array([ 0.04270847, 0.03216249, 0.92512904]))
"""
if verbose:
status = np.arange(0, nsamples, nsamples / 10)
else:
status = [-1]
collected_output = []
for i in xrange(nsamples):
r, X = generator()
# print(np.shape(r))
# print(np.shape(X))
M = model.history_model(r,
X,
w0=w0,
p0=p0,
nu0=nu0,
hf0=hf0,
applythreshold=applythreshold,
emiter=80)
collected_output.append(collector(M))
w0 = M.w
p0 = M.pi
nu0 = M.nu
if i in status:
if verbose:
sys.stderr.write("\r%d%%" % (100 * (float(i) / nsamples), ))
sys.stderr.flush()
if verbose:
sys.stderr.write("\rDone\n")
sys.stderr.flush()
return collected_output
def analysis(d,
w0,
nsamples=200,
perm_collector=statistics.EvaluationCollector):
"""Analyze a dataset
:Parameters:
*d*
a history.DataSet instance (typically a subclass of history.DataSet)
*w0*
starting values for the first parameters in the model. The remaining parameters will
start at 0. It seems to be a good idea to give starting values for the stimulus dependent
parameters in the model and leave the parameters for history features at 0.
*nsamples*
number of samples for monte carlo procedures
*perm_collector*
the collector object for the permutation tests. The default one should do for all
experiments with a design matrix in which the first column is 1 and the second column
refers to the slope.
:Example:
>>> d,w,plotinfo = load_plaid ()
>>> results = analysis ( d, w, 10 )
>>> results.keys()
['model_nohist', 'model_w_hist', 'bootstrap', 'permutation_nh', 'permutation_wh']
"""
if d.__dict__.has_key('detection') and d.detection:
dnh = d.__class__(d.fname, threshold=len(d.th_features) > 0)
else:
dnh = d.__class__(d.fname) # reload the data without history
dwh = d
if getattr(d, 'audio', False):
easy, difficult = d.performance_filter()
else:
easy, difficult = None, None
logging.info("Fitting models")
if getattr(d, 'p0', np.array(False)).all():
# This code will only run for monkeys
M0 = model.history_model(dnh.r,
dnh.X,
applythreshold=dnh.th_features,
w0=w0,
p0=d.p0,
lm=0.1,
hf0=dnh.hf0,
emiter=40)
Mnh = model.history_model(dnh.r,
dnh.X,
applythreshold=dnh.th_features,
w0=M0.w,
p0=M0.pi,
lm=.1,
hf0=dnh.hf0,
verbose=True,
emiter=300)
logging.info("likelihood for independent responses: %g" %
(Mnh.loglikelihood, ))
Mwh = model.history_model(dwh.r,
dwh.X,
applythreshold=dwh.th_features,
w0=M0.w,
p0=M0.pi,
lm=.1,
hf0=dwh.hf0,
verbose=True,
emiter=300)
logging.info("likelihood for history model: %g" %
(Mwh.loglikelihood, ))
else:
Mnh, Mwh = search_for_start(d.r,
d.X,
w0,
d.th_features,
d.hf0,
storeopt=True)
logging.info("likelihood for independent responses: %g" %
(Mnh.loglikelihood, ))
logging.info("likelihood for history model: %g" %
(Mwh.loglikelihood, ))
print "nh", Mnh.w, Mnh.pi
print "wh", Mwh.w, Mwh.pi
print Mwh.applythreshold
# Monte Carlo testing
if nsamples > 0:
r_, X_ = dwh.permutation()
Mnh_perm, Mwh_perm = search_for_start(r_, X_, w0, d.th_features, d.hf0)
logging.info("Permutation with history")
dnh.rng.set_state(
dwh.rng.get_state()
) # Set the states of the two random number generators to the same values to get the exact same sequence of random numbers
perm_collector = statistics.EvaluationCollector(Mwh, easy, difficult)
permutation_wh = pl.array(
statistics.mcsimulation(dwh.permutation,
perm_collector,
nsamples,
Mwh_perm.w,
Mwh_perm.pi,
Mwh_perm.nu,
verbose=logging.root.level < 20,
hf0=dwh.hf0,
applythreshold=Mwh.applythreshold))
logging.info("Permutation without history")
perm_collector = statistics.EvaluationCollector(Mnh, easy, difficult)
permutation_nh = pl.array(
statistics.mcsimulation(dnh.permutation,
perm_collector,
nsamples,
Mnh_perm.w,
Mnh_perm.pi,
Mnh_perm.nu,
verbose=logging.root.level < 20,
hf0=dnh.hf0,
applythreshold=Mwh.applythreshold))
logging.info("Bootstrap")
kcollector = statistics.Kernel_and_Slope_Collector(dwh.h,
dwh.hf0,
slopeindex=range(
1, dwh.hf0))
bootstrap = pl.array(
statistics.mcsimulation(dwh.bootstrap,
kcollector,
nsamples,
Mwh.w,
Mwh.pi,
Mwh.nu,
verbose=logging.root.level < 20,
hf0=dwh.hf0,
applythreshold=Mwh.applythreshold))
else:
permutation_wh = None
permutation_nh = None
bootstrap = None
results = {
'model_nohist': Mnh,
'model_w_hist': Mwh,
'permutation_wh': permutation_wh,
'permutation_nh': permutation_nh,
'bootstrap': bootstrap
}
return results
def search_for_start(r,
X,
w0,
applythreshold,
hf0,
pm=(.85, .93, 1., 1.07, 1.15),
storeopt=False,
modulation=False,
doublemodulation=False):
"""Search for starting values
:Parameters:
*d*
data set
*w0*
coarse starting values
*pm*
increments or decrements
"""
logging.info("Initial preoptimization")
Mwh = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
Mnh = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
M0 = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
if modulation:
Mhmod = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
# add the history weights the model with history
Mwh.w = np.concatenate((Mwh.w, np.zeros(6, 'd')))
Mwh.X = copy.copy(X)
Mwh.X = Mwh.X[:, :-2] # only the part that has no modulation
# add those same weights + modulatory terms to the hmod model
Mhmod.w = np.concatenate((Mhmod.w, np.zeros(15, 'd')))
Mhmod.X = X
# to check that the sizes work
print "X", np.shape(X)
print "Mwh", Mwh.w, np.shape(Mwh.X)
print "Mhmod", Mhmod.w, np.shape(Mhmod.X)
elif doublemodulation:
Mhmod = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
lm=0.1,
hf0=hf0,
emiter=100)
# add the history weights to the model with just history
Mwh.w = np.concatenate((Mwh.w, np.zeros(6, 'd')))
Mwh.X = copy.copy(X)
Mwh.X = Mwh.X[:, :-3] # only the part that has no modulation
# add those same weights + modulatory terms to the hmod model
Mhmod.w = np.concatenate((Mhmod.w, np.zeros(24, 'd')))
Mhmod.X = X
# to check that the sizes work
print "X", np.shape(X)
print "Mwh", Mwh.w, np.shape(Mwh.X)
print "Mhmod", Mhmod.w, np.shape(Mhmod.X)
else:
Mwh.w = np.concatenate((Mwh.w, np.zeros(6, 'd')))
Mwh.X = X
Mhmod = [] # return empty
nhind = 0
whind = 0
i = 1
for al in pm:
for lm in pm:
logging.info("::::: Optimizing from starting value %d :::::" %
(i, ))
w0 = M0.w.copy()
w0[1:hf0] *= al
p0 = M0.pi.copy()
p0[0] *= lm
p0[-1] = 1 - p0[0] - p0[1]
if modulation or doublemodulation:
M_ = model.history_model(r,
X,
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
if Mhmod.loglikelihood < M_.loglikelihood:
logging.info(" *model chosen for history + modulation*")
Mhmod = M_
whind = i
if modulation:
M_ = model.history_model(r,
X[:, :-2],
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
elif doublemodulation:
M_ = model.history_model(r,
X[:, :-3],
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
if Mwh.loglikelihood < M_.loglikelihood:
logging.info(" *model chosen for history*")
Mwh = M_
whind = i
else:
M_ = model.history_model(r,
X,
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
if Mwh.loglikelihood < M_.loglikelihood:
logging.info(" *model chosen for history*")
Mwh = M_
whind = i
M_ = model.history_model(r,
X[:, :hf0],
applythreshold=applythreshold,
w0=w0,
p0=p0,
nu0=M0.nu,
lm=0.1,
hf0=hf0,
verbose=True,
emiter=300,
storeopt=storeopt)
if Mnh.loglikelihood < M_.loglikelihood:
logging.info(" *model chosen for independent*")
Mnh = M_
nhind = i
i += 1
logging.info("Mwh.w = %s\nMnh.w = %s" % (str(Mwh.w), str(Mnh.w)))
logging.info("Mwh.ll = %g\nMnh.ll = %s" %
(Mwh.loglikelihood, Mnh.loglikelihood))
logging.info(
"Starting values:\n with history: %d\n without history: %d\n" %
(whind, nhind))
# NOW, THE HISTORY ONLY MODEL HAS SIZE 22!
print "X", np.shape(X)
print "Mwh", Mwh.w, np.shape(Mwh.X)
if modulation:
print "Mhmod", Mhmod.w, np.shape(Mhmod.X)
return Mnh, Mwh, Mhmod