def _format_for_gam2(count, time, pos):
assert np.rank(count) == np.rank(time) == (np.rank(pos) - 1)
assert count.shape[0] == time.shape[0] == pos.shape[0]
assert (count.shape[1] + 1) == time.shape[1] == pos.shape[1]
y = count.flatten()[:,None]
npt = y.size
tax, spax = 1, -1
# don't use real time as won't compare equivalent portions of trials
# t = edge2cen(time, axis=tax) # (ntrial, nbin)
# subtract offset to get all relative times
#t = (t - t[:,:1]).flatten()[:,None]
# instead use bin numbers
ntrial, nbin = count.shape
t = np.tile(np.arange(nbin, dtype=float)[None], (ntrial, 1))
t = t.flatten()[:,None]
# also create trial numbers for pulling out appropriate later on
# these don't correspond to original trial numbers because original trials
# have been permuted before getting here
tr = np.tile(np.arange(ntrial), (nbin, 1)).T.flatten()
d = kin.get_dir(pos, tax=tax, spax=spax).reshape(npt, 3)
p = edge2cen(pos, axis=tax).reshape(npt, 3)
v = kin.get_vel(pos, time, tax=tax, spax=spax).reshape(npt, 3)
sp = kin.get_speed(pos, time, tax=tax, spax=spax).flatten()[:,None]
# q is a second direction set-of-columns for deviance calculation
q = kin.get_dir(pos, tax=tax, spax=spax).reshape(npt, 3)
return np.concatenate([y,t,d,p,v,sp,q], axis=1), tr
def encode(self, bnd):
'''
Return spike counts corresponding to given direction and encoding model.
Parameters
----------
bnd :BinnedData
uses bnd.pos, bnd.bin_edges
'''
pos = bnd.pos
time = bnd.bin_edges
ntask, nrep, nedge, ndim = pos.shape
model = self.model
assert(ndim == 3)
nbin = nedge - 1
lshape = ntask, nrep, nbin
rate = np.zeros(lshape, dtype=float)
assert(model == 'kd') # only model currently implemented
rate += self.B['k'] # baseline
dr = get_dir(pos)
rate += np.dot(dr, self.B['D']) # direction
rate += np.dot(pos[...,:-1,:], self.B['P']) # position
speed = get_speed(pos, time, tax=2, spax=-1)
rate += self.B['s'] * speed
return np.exp(rate)
def encode(self, bnd):
'''
Return spike counts corresponding to given direction and encoding model.
Parameters
----------
bnd :BinnedData
uses bnd.pos, bnd.bin_edges
'''
pos = bnd.pos
time = bnd.bin_edges
ntask, nrep, nedge, ndim = pos.shape
model = self.model
assert (ndim == 3)
nbin = nedge - 1
lshape = ntask, nrep, nbin
rate = np.zeros(lshape, dtype=float)
assert (model == 'kd') # only model currently implemented
rate += self.B['k'] # baseline
dr = get_dir(pos)
rate += np.dot(dr, self.B['D']) # direction
rate += np.dot(pos[..., :-1, :], self.B['P']) # position
speed = get_speed(pos, time, tax=2, spax=-1)
rate += self.B['s'] * speed
return np.exp(rate)
def _format_for_gam(count, time, pos):
'''
Format data for gam_predict_cv, i.e. an (n, 12) array
Parameters
----------
count : ndarray
spike counts, shape (ntrial, nbin)
time : ndarray
bin_edges, shape (ntrial, nbin + 1)
pos : ndarray
positions at `bin_edges`, shape (ntrial, nbin + 1, 3)
Returns
-------
formatted_data : ndarray
shape (ntrial * nbin, 12)
dimension 1 is [count, t, dx, dy, dz, px, py, pz, vx, vy, vz, sp]
'''
assert np.rank(count) == np.rank(time) == (np.rank(pos) - 1)
assert count.shape[0] == time.shape[0] == pos.shape[0]
assert (count.shape[1] + 1) == time.shape[1] == pos.shape[1]
y = count.flatten()[:,None]
npt = y.size
tax, spax = 1, -1
# don't use real time as won't compare equivalent portions of trials
# t = edge2cen(time, axis=tax) # (ntrial, nbin)
# subtract offset to get all relative times
#t = (t - t[:,:1]).flatten()[:,None]
# instead use bin numbers
ntrial, nbin = count.shape
t = np.tile(np.arange(nbin, dtype=float)[None], (ntrial, 1))
t = t.flatten()[:,None]
d = kin.get_dir(pos, tax=tax, spax=spax).reshape(npt, 3)
p = edge2cen(pos, axis=tax).reshape(npt, 3)
v = kin.get_vel(pos, time, tax=tax, spax=spax).reshape(npt, 3)
sp = kin.get_speed(pos, time, tax=tax, spax=spax).flatten()[:,None]
# q is a second direction set-of-columns for deviance calculation
q = kin.get_dir(pos, tax=tax, spax=spax).reshape(npt, 3)
return np.concatenate([y,t,d,p,v,sp,q], axis=1)
def make_inst_dir_rate(time, pos, pd, kd, sqp={}):
'''
Create directional firing components from instantaneous direction,
preferred direction, and modulation depth.
Parameters
----------
time : ndarray
times for pos samples, shape (??)
pos : ndarray
positional data, shape (ntrial, nbin, 3)
pd : array_like
pds for a number of cells, shape (nunit, 3)
kd : array_like
modulation depth, shape (nunit)
sqp : dict, optional
parameter values for squashing function
'''
drn = get_dir(pos, tax=1, spax=-1)
rate = np.dot(drn, pd.T) * kd
set_squash_defaults(sqp, rate)
return squash(rate, sqp)
def prepare_regressors(count, pos, time, model=''):
'''
Parameters
----------
count : array_like
firing rates, shape (ntrial, nbin)
pos : array_like
positional data, shape (ntrial, nbin + 1, ndim)
these are positions at bin edges
time : array_like
bin edge times, shape (ntrial, nbin + 1)
model : string
model
Returns
-------
reg_vars : ndarray
reg_frs : ndarray
'''
count = np.asarray(count)
assert(np.rank(count) == 2)
ntrial, nbin = count.shape
assert(not np.any(np.isnan(count)))
pos = np.asarray(pos)
assert(np.rank(pos) == 3)
assert(pos.shape[0:2] == (ntrial, nbin + 1))
time = np.asarray(time)
assert(np.rank(time) == 2)
assert(time.shape == (ntrial, nbin + 1))
assert(type(model) == str)
#ntrial, nbin = count.shape
ndim = pos.shape[-1]
nvar = ntrial * nbin
endog = count.reshape(nvar) # should be equivalent to flatten
vels = None
exog = None
# add the time-invariant variables
if 'p' in model:
pos_flat = pos[:,1:,:].reshape(nvar, ndim)
exog = add_column(exog, pos_flat)
if 's' in model:
if vels == None:
vels = get_vel(pos, time, tax=1, spax=2)
spds = np.sqrt(np.sum(vels**2, axis=2))
spds_flat = spds.reshape(nvar)
exog = add_column(exog, spds_flat)
# do the potentially time-variant variables
if ('d' in model or 'v' in model):
if 'd' in model:
var = get_dir(pos, tax=1, spax=2)
elif 'v' in model:
if vels == None:
vels = get_vel(pos, time, tax=1, spax=2)
var = vels
if not 'X' in model: # time-static variable
var_flat = np.reshape(var, (nvar, ndim))
else: # time-dynamic variable
mask = np.eye(nbin)
var_flat = np.reshape(mask[None,...,None] * var[...,None,:],
(nvar, nbin * ndim))
exog = add_column(exog, var_flat)
# do the constant + dynamic direction model
if 'q' in model:
assert ('X' in model)
# for simplicity's sake, let us call this model 'kqX'
assert ('d' not in model)
assert ('v' not in model)
if 'k' in model:
exog = sm.tools.add_constant(exog, prepend=False)
offset = np.log(np.diff(time, axis=-1).reshape(nvar))
return endog, exog, offset
def test_gam_predict_cv():
dsname = 'frank-osmd'
align = 'hold'
lag = 0.1 # s
b0 = 20 # Hz
noise_sd = 1. # Hz
b_scale = 10.
nbin = 10
ds = datasets[dsname]
dc = DataCollection(ds.get_files()[:5])
unit = ds.get_units()[0]
dc.add_unit(unit, lag)
bnd = dc.make_binned(nbin=nbin, align=align, do_count=True)
ntask, nrep, nedge, ndim = bnd.pos.shape
pos = bnd.pos.reshape(ntask * nrep, nedge, -1)
drn = kin.get_dir(pos)
# simplest model
# y = b0 + Bd.D
B = np.array([.2, .6, .4]) * b_scale
y = b0 + np.dot(drn, B)
noise = np.random.normal(0, noise_sd, size=y.shape)
y += noise
bnd.set_count_from_flat(y[:,None])
out = gam_predict_cv(bnd, ['kd'], [dsname, unit, lag, align], \
family='gaussian')
have = np.mean(out.coef[0], axis=0)[:4]
want = np.array([b0, B[0], B[1], B[2]])
np.testing.assert_array_almost_equal(have, want, decimal=1)
# changing Bdt model, gaussian noise model
# y = b0 + Bdt.D
gc_type = [('A', float), ('fwhm', float), ('c', float)]
gcoeff = np.array([[(1., 1.5, 3.), (.75, 2.5, 6.)],
[(.4, 4., 2.), (.1, 2., 5.)],
[(.8, 3., 5.), (.6, 4.5, 6.5)]], dtype=gc_type)
gcoeff = gcoeff.view(np.recarray)
x = np.linspace(0, 10, nbin)
Bt = np.zeros((3, nbin))
for Bkt, gc in zip(Bt, gcoeff):
g0 = gauss1d(x, gc.A[0], gc.fwhm[0], gc.c[0])
g1 = gauss1d(x, gc.A[1], gc.fwhm[1], gc.c[1])
Bkt[:] = g0 + g1
Bt *= b_scale
y = b0 + np.sum(Bt.T[None] * drn, axis=-1)
y = np.random.normal(y, noise_sd)
bnd.set_count_from_flat(y[:,None])
out = gam_predict_cv(bnd, ['kdX'], [dsname, unit, lag, align], \
family='gaussian')
# check that actual and predicted are correlated
x = out.actual.flatten()
y = out.pred[8].flatten()
mc, rp = pearsonr(x,y)
np.testing.assert_approx_equal(mc, 1, significant=1)