def test_noresample_stimulus():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0,360,90)
num_blank_steps = 0
num_bar_steps = 30
ecc = 12
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 1.0
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1,2,3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width, scale_factor, tr_length, dtype)
# stimulus
npt.assert_equal(stimulus.stim_arr.shape[0:2],stimulus.stim_arr0.shape[0:2])
npt.assert_equal(stimulus.deg_x.shape,stimulus.deg_x0.shape)
npt.assert_equal(stimulus.deg_y.shape,stimulus.deg_y0.shape)
def generate_stimulus(bars):
tr_length = 1.0
scale_factor = 1.0
dtype = c_int16
stimulus = VisualStimulus(bars, VIEWING_DISTANCE, SCREEN_WIDTH, scale_factor, tr_length, dtype)
return stimulus
def __init__(self, dm, viewing_distance, screen_width, tr_dm):
self.stimulus = VisualStimulus(stim_arr=dm,
viewing_distance=viewing_distance,
screen_width=screen_width,
scale_factor=1,
tr_length=tr_dm,
dtype=np.short)
self.model_func = OGModel.GaussianModel(stimulus=self.stimulus,
hrf_model=utils.spm_hrf)
self.model_func.hrf_delay = 0
self.predictions = None
def test_resurrect_model():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0,360,90)
thetas = np.insert(thetas,0,-1)
thetas = np.append(thetas,-1)
num_blank_steps = 20
num_bar_steps = 20
ecc = 10
tr_length = 1.5
frames_per_tr = 1.0
scale_factor = 0.50
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance, screen_width,
thetas, num_bar_steps, num_blank_steps, ecc, clip=0.01)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width, scale_factor, tr_length, dtype)
# set cache grids
x_grid = utils.grid_slice(-10, 10, 5)
y_grid = utils.grid_slice(-10, 10, 5)
s_grid = utils.grid_slice(0.55,5.25, 5)
grids = (x_grid, y_grid, s_grid,)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0
model.mask_size = 5
# seed rng
np.random.seed(4932)
# cache the model
cache = model.cache_model(grids, ncpus=3)
# resurrect cached model
cached_model_path = '/tmp/og_cached_model.pkl'
model = og.GaussianModel(stimulus, utils.double_gamma_hrf, cached_model_path=cached_model_path)
model.hrf_delay = 0
model.mask_size = 5
# make sure the same
nt.assert_true(np.sum([c[0] for c in cache] - model.cached_model_timeseries) == 0)
nt.assert_true(np.sum([c[1] for c in cache] - model.cached_model_parameters) == 0)
def test_cache_model_Ns():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0,360,90)
thetas = np.insert(thetas,0,-1)
thetas = np.append(thetas,-1)
num_blank_steps = 20
num_bar_steps = 20
ecc = 10
tr_length = 1.5
frames_per_tr = 1.0
scale_factor = 0.50
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
Ns = 5
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance, screen_width,
thetas, num_bar_steps, num_blank_steps, ecc, clip=0.01)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width, scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0
model.mask_size = 5
# set cache grids
x_grid = (-10, 10)
y_grid = (-10, 10)
s_grid = (0.55,5.25)
grids = (x_grid, y_grid, s_grid,)
# seed rng
np.random.seed(4932)
# cache the pRF model
cache = model.cache_model(grids, Ns=Ns, ncpus=3)
# save it out
pickle.dump(cache, open('/tmp/og_cached_model.pkl','wb'))
# make sure its the right size
cached_model = pickle.load(open('/tmp/og_cached_model.pkl','rb'))
nt.assert_equal(np.sum([c[0] for c in cache]),np.sum([c[0] for c in cached_model]))
def __init__(self,
data,
fit_model,
visual_design,
screen_distance,
screen_width,
scale_factor,
tr,
bound_grids,
grid_steps,
bound_fits,
n_jobs,
hrf,
sg_filter_window_length=210,
sg_filter_polyorder=3,
sg_filter_deriv=0,
nr_TRs=103):
# immediately convert nans to nums
self.data = np.nan_to_num(data)
self.data_var = self.data.var(axis=-1)
self.n_units = self.data.shape[0]
self.n_timepoints = self.data.shape[1]
self.hrf = hrf
self.nr_TRs = nr_TRs
self.stimulus = VisualStimulus(stim_arr=visual_design,
viewing_distance=screen_distance,
screen_width=screen_width,
scale_factor=scale_factor,
tr_length=tr,
dtype=np.short)
#assert self.n_timepoints == self.stimulus.run_length, \
# "Data and design matrix do not have the same nr of timepoints, %i vs %i!"%(self.n_timepoints, self.stimulus.run_length)
if fit_model == 'gauss':
self.model_func = GaussianModel(stimulus=self.stimulus,
hrf=self.hrf,
nr_TRs=self.nr_TRs)
self.model_func.hrf_delay = 0
self.predictions = None
self.fit_model = fit_model
self.bound_grids = bound_grids
self.grid_steps = grid_steps
self.bound_fits = bound_fits
self.n_jobs = n_jobs
def fit_voxel(tup):
(ii, vx, js) = tup
stdat = js['Stimulus']
if pimms.is_list(stdat): stdat = stdat[0]
height = stdat['fieldofviewVert']
width = stdat['fieldofviewHorz']
### STIMULUS
# First get a viewing distance and screen size
dist = 100 # 100 cm is arbitrary
stim_width = 2 * dist * np.tan(np.pi/180 * width/2)
stimulus = VisualStimulus(stim, dist, stim_width, 1.0, float(js['TR']), ctypes.c_int16)
if fixed_hrf is not False:
model = og_nohrf.GaussianModel(stimulus, utils.double_gamma_hrf)
model.hrf_delay = fixed_hrf
else: model = og.GaussianModel(stimulus, utils.double_gamma_hrf)
### FIT
## define search grids
# these define min and max of the edge of the initial brute-force search.
x_grid = (-width/2,width/2)
y_grid = (-height/2,height/2)
s_grid = (1/stimulus.ppd + 0.25, 5.25)
h_grid = (-1.0, 1.0)
## define search bounds
# these define the boundaries of the final gradient-descent search.
x_bound = (-width, width)
y_bound = (-height, height)
s_bound = (1/stimulus.ppd, 12.0) # smallest sigma is a pixel
b_bound = (1e-8,None)
u_bound = (None,None)
h_bound = (-3.0,3.0)
## package the grids and bounds
if fixed_hrf is not False:
grids = (x_grid, y_grid, s_grid)
bounds = (x_bound, y_bound, s_bound, b_bound, u_bound,)
else:
grids = (x_grid, y_grid, s_grid, h_grid)
bounds = (x_bound, y_bound, s_bound, h_bound, b_bound, u_bound,)
## fit the response
# auto_fit = True fits the model on assignment
# verbose = 0 is silent
# verbose = 1 is a single print
# verbose = 2 is very verbose
if fixed_hrf is not False:
fit = og_nohrf.GaussianFit(model, vx, grids, bounds, Ns=Ns,
voxel_index=(ii, 1, 1), auto_fit=True, verbose=2)
else:
fit = og.GaussianFit(model, vx, grids, bounds, Ns=Ns,
voxel_index=(ii, 1, 1), auto_fit=True, verbose=2)
return (ii, vx) + tuple(fit.overloaded_estimate) + (fit.prediction,)
def test_resample_stimulus():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0,360,90)
num_blank_steps = 0
num_bar_steps = 30
ecc = 12
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1,2,3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width, scale_factor, tr_length, dtype)
# grab the stimulus dimensions
stim_dims = stimulus.stim_arr.shape
stim_coarse_dims = stimulus.stim_arr0.shape
# assert
nt.assert_true(stim_coarse_dims[0]/stim_dims[0] == scale_factor)
nt.assert_true(stim_coarse_dims[1]/stim_dims[1] == scale_factor)
nt.assert_true(stim_coarse_dims[2] == stim_dims[2])
npt.assert_array_equal(np.unique(stimulus.stim_arr0), [0, 1])
# make sure the duty-cycle doesn't change with resampling
npt.assert_almost_equal(np.sum(stimulus.stim_arr0==1)/np.sum(stimulus.stim_arr0>-1),np.sum(stimulus.stim_arr==1)/np.sum(stimulus.stim_arr>-1),3)
def read_prf_stimuli(n_vols=(224, 224)):
img = loadmat('/Fridge/users/margriet/stimuli/BAIR_pRF/bar_apertures.mat')
images = img['bar_apertures']
baseline = int(11.9 / 0.85)
b_img = zeros((images.shape[0], images.shape[1], baseline))
c = concatenate((b_img, images, b_img), 2)
all_img = []
for n_vol in n_vols:
all_img.append(c[:, :, :n_vol].copy())
st = concatenate(all_img, 2)
st = resize(st, (100, 100), anti_aliasing=True)
stimulus = VisualStimulus(st,
112,
32,
scale_factor=0.1,
tr_length=0.85,
dtype=c_short)
return stimulus
def setup_data_from_h5(data_file,
n_pix,
extent=[-5,5],
screen_distance=225,
screen_width=69.0,
rsq_threshold=0.5,
TR=0.945,
cv_fold=1,
n_folds=6,
use_median=True,
mask_name = 'V1'):
hdf5_file = get_figshare_data(data_file)
############################################################################################################################################
# getting the data
############################################################################################################################################
# timecourses are single-run psc data, either original or leave-one-out.
timecourse_data_single_run = roi_data_from_hdf(['*psc'],mask_name, hdf5_file,'psc').astype(np.float64)
timecourse_data_loo = roi_data_from_hdf(['*loo'],mask_name, hdf5_file,'loo').astype(np.float64)
timecourse_data_all_psc = roi_data_from_hdf(['*av'],mask_name, hdf5_file,'all_psc').astype(np.float64)
# prfs are per-run, as fit using the loo data
all_prf_data = roi_data_from_hdf(['*all'],mask_name, hdf5_file,'all_prf').astype(np.float64)
prf_data = roi_data_from_hdf(['*all'],mask_name, hdf5_file,'prf').astype(np.float64).reshape((all_prf_data.shape[0], -1, all_prf_data.shape[-1]))
dm=create_visual_designmatrix_all(n_pixels=n_pix)
if use_median:
dm_crossv = dm
else:
dm_crossv = np.tile(dm,(1,1,n_folds-1))
# apply pixel mask to design matrix, as we also do this to the prf profiles.
mask = dm.sum(axis = -1, dtype = bool)
dm = dm[mask,:]
# voxel mask for crossvalidation
# only count those voxels here that have positive rsq
rsq_crossv = np.mean(prf_data[:,:,-1], axis=1) * np.sign(np.mean(prf_data[:,:,4], axis=1))
rsq_mask_crossv = rsq_crossv > rsq_threshold
# determine amount of trs
nr_TRs = int(timecourse_data_single_run.shape[-1] / n_folds)
# now, mask the data and select for this fold
test_data = timecourse_data_single_run[rsq_mask_crossv,nr_TRs*cv_fold:nr_TRs*(cv_fold+1)]
if use_median:
train_data = timecourse_data_loo[rsq_mask_crossv,nr_TRs*cv_fold:nr_TRs*(cv_fold+1)]
else:
train_data = np.delete(timecourse_data_single_run[rsq_mask_crossv,:], np.s_[nr_TRs*cv_fold:nr_TRs*(cv_fold+1)], axis=1)
# set up the prf_data variable needed for decoding
prf_cv_fold_data = prf_data[rsq_mask_crossv,cv_fold]
# rescale the data
#sign=np.sign(train_data)
#train_data -= prf_cv_fold_data[:, 5, np.newaxis]
#train_data /= prf_cv_fold_data[:, 4, np.newaxis]
#train_data = sign * np.abs(train_data)**(1.0/prf_cv_fold_data[:, 3, np.newaxis])
#test_data -= prf_cv_fold_data[:, 5, np.newaxis]
#test_data /= prf_cv_fold_data[:, 4, np.newaxis]
############################################################################################################################################
# setting up prf timecourses - NOTE, this is for the 'all' situation, so should be really done on a run-by-run basis using a run's
# loo data and prf parameters. A test set would then be taken from the single_run data as this hasn't been used for that run's fit.
############################################################################################################################################
# set up model with hrf etc.
def my_spmt(delay, tr):
return spmt(np.arange(0, 33, tr))
# we're going to use these popeye convenience functions
# because they are fast, and because they were used in the fitting procedure
stimulus = VisualStimulus(dm_crossv,
screen_distance,
screen_width,
1.0,
TR,
ctypes.c_int16)
css_model = CompressiveSpatialSummationModelFiltered(stimulus, my_spmt)
css_model.hrf_delay = 0
############################################################################################################################################
# setting up prf spatial profiles for subsequent covariances, now some per-run stuff was done
############################################################################################################################################
# indices into prf output array:
# 0: X
# 1: Y
# 2: s (size, standard deviation of gauss)
# 3: n (nonlinearity power)
# 4: a (amplitude)
# 5: b (baseline, intercept value)
# 6: rsq per-run
# 7: rsq across all
#
deg_x, deg_y = np.meshgrid(np.linspace(extent[0], extent[1], n_pix, endpoint=True), np.linspace(
extent[0], extent[1], n_pix, endpoint=True))
rfs = generate_og_receptive_fields( prf_data[rsq_mask_crossv, cv_fold, 0],
prf_data[rsq_mask_crossv, cv_fold, 1],
prf_data[rsq_mask_crossv, cv_fold, 2],
np.ones((rsq_mask_crossv.sum())),
deg_x,
deg_y)
#this step is used in the css model
rfs_normal = rfs / ((2 * np.pi * prf_data[rsq_mask_crossv, cv_fold, 2]**2) * 1 /np.diff(css_model.stimulus.deg_x[0, 0:2])**2)
#rfs **= prf_cv_fold_data[:, 3]
#WARNING: CSS-like normalisation does not work well at all. simply divide by the sum for simplicity?
#for i in range(rfs_normal.shape[2]):
# rfs_normal[:,:,i]/=np.sum(rfs_normal[:,:,i])
#pl.imshow(rfs[:,:,3])
#print(np.sum(rfs_normal[:,:,0]))
#print(np.sum(rfs_normal[:,:,1]))
#print(np.sum(rfs_normal[:,:,2]))
#print(np.sum(rfs_normal[:,:,3]))
#print(np.sum(rfs_normal[:,:,4]))
#pl.colorbar()
#(however in very original decoding, masking was done only at the end.i.e. W had all the pixels in the square.)
#shouldnt have an impact but remember to check if it does. ask tomas.
# convert to 1D array and mask with circular mask (tested, works)
rfs_normal = rfs_normal.reshape((np.prod(mask.shape),-1))[mask.ravel(),:]
rfs = rfs.reshape((np.prod(mask.shape),-1))[mask.ravel(),:]
#rfs **= prf_cv_fold_data[:, 3, np.newaxis].T
############################################################################################################################################
# setting up prf spatial profiles for the decoding step, creating linear_predictor_ip
############################################################################################################################################
# and then we try to use this:
W=rfs_normal.T
#linear_predictor_ip=np.zeros((W.shape[0], W.shape[1]+1))
#linear_predictor_ip[:,1:]=np.copy(W)
#do some rescalings. This affects decoding quite a lot!
#linear_predictor_ip **= np.tile(prf_cv_fold_data[:, 3], (linear_predictor_ip.shape[1],1)).T
# #at this point (after power raising but before multiplication/subtraction) the css model convolves with hrf.
#linear_predictor_ip *= np.tile(prf_cv_fold_data[:, 4], (linear_predictor_ip.shape[1],1)).T
#linear_predictor_ip += np.tile(prf_cv_fold_data[:, 5], (linear_predictor_ip.shape[1],1)).T
############################################################################################################################################
# no convolution simple prediction. use to test difference between CSS model (time dependent)
# and respective nonlinear, time-independent model that we use in decoding ("simple prediction")
############################################################################################################################################
simple_prediction= np.dot(rfs_normal.T, dm_crossv.reshape((np.prod(mask.shape),-1))[mask.ravel(),:])
simple_prediction **= prf_cv_fold_data[:, 3, np.newaxis]
simple_prediction *= prf_cv_fold_data[:, 4, np.newaxis]
simple_prediction += prf_cv_fold_data[:, 5, np.newaxis]
hrf = css_model.hrf_model(css_model.hrf_delay, css_model.stimulus.tr_length)
#for i in range(simple_prediction.shape[0]):
#a=np.max(simple_prediction[i,:])
#simple_prediction[i,:] = fftconvolve(simple_prediction[i,:], hrf)[0:simple_prediction.shape[1]]
#simple_prediction[i,:] -= savgol_filter(simple_prediction[i,:], window_length=css_model.sg_filter_window_length, polyorder=css_model.sg_filter_order, deriv=0, mode='nearest')
css_prediction=np.zeros((rsq_mask_crossv.sum(),train_data.shape[1]))
for g, vox_prf_pars in enumerate(prf_cv_fold_data):
css_prediction[g] = css_model.generate_prediction(
x=vox_prf_pars[0], y=vox_prf_pars[1], sigma=vox_prf_pars[2], n=vox_prf_pars[3], beta=vox_prf_pars[4], baseline=vox_prf_pars[5])
#b=np.max(css_prediction[g,:])
#simple_prediction[:,g] *= b/a
all_residuals_css = train_data - css_prediction
all_residuals_simple = train_data - simple_prediction
print("Shapiro-Wilk normality test (if second value is large, residuals are normal):", sp.stats.shapiro(all_residuals_css))
print("CSS resid: ",np.sum(all_residuals_css))
print("simple model (no hrf) resid: ",np.sum(all_residuals_simple))
# print(np.max(simple_prediction[i,:])/a)
# some quick visualization
f = pl.figure(figsize=(17,5))
s = f.add_subplot(211)
pl.plot(css_prediction[np.argmax(rsq_crossv[rsq_mask_crossv])], label='CSS prediction')
pl.plot(train_data[np.argmax(rsq_crossv[rsq_mask_crossv])], label='data')
pl.plot(simple_prediction[np.argmax(rsq_crossv[rsq_mask_crossv])], label='Simple prediction')
#pl.plot(all_residuals_css[np.argmax(rsq_crossv[rsq_mask_crossv])], label='resid')
pl.legend()
s.set_title('best voxel')
s = f.add_subplot(212)
pl.plot(css_prediction[np.argmin(rsq_crossv[rsq_mask_crossv])], label='CSS prediction')
pl.plot(train_data[np.argmin(rsq_crossv[rsq_mask_crossv])], label='data')
pl.plot(simple_prediction[np.argmin(rsq_crossv[rsq_mask_crossv])], label='Simple prediction')
#pl.plot(all_residuals_css[np.argmin(rsq_crossv[rsq_mask_crossv])], label='resid')
pl.legend()
s.set_title('worst voxel given this threshold')
all_residual_covariance_css = np.cov(all_residuals_css)
return prf_cv_fold_data, W, all_residuals_css, all_residual_covariance_css, test_data, mask
def test_og_fit():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 90)
thetas = np.insert(thetas, 0, -1)
thetas = np.append(thetas, -1)
num_blank_steps = 30
num_bar_steps = 30
ecc = 12
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 1.0
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0
model.mask_size = 5
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
hrf_delay = 0.66
beta = 2.5
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, hrf_delay, beta, baseline)
# set search grid
x_grid = (-10, 10)
y_grid = (-10, 10)
s_grid = (0.25, 5.25)
h_grid = (-1.0, 1.0)
# set search bounds
x_bound = (-12.0, 12.0)
y_bound = (-12.0, 12.0)
s_bound = (0.001, 12.0)
h_bound = (-1.5, 1.5)
b_bound = (1e-8, None)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
h_grid,
)
bounds = (x_bound, y_bound, s_bound, h_bound, b_bound, m_bound)
# fit the response
fit = og.GaussianFit(model, data, grids, bounds, Ns=5)
# coarse fit
npt.assert_almost_equal(fit.x0, -5.0)
npt.assert_almost_equal(fit.y0, 5.0)
npt.assert_almost_equal(fit.s0, 2.75)
npt.assert_almost_equal(fit.hrf0, 0.5)
# the baseline/beta should be 0/1 when regressed data vs. estimate
(m, b) = np.polyfit(fit.scaled_ballpark_prediction, data, 1)
npt.assert_almost_equal(m, 1.0)
npt.assert_almost_equal(b, 0.0)
# assert equivalence
npt.assert_almost_equal(fit.x, x)
npt.assert_almost_equal(fit.y, y)
npt.assert_almost_equal(fit.hrf_delay, hrf_delay)
npt.assert_almost_equal(fit.sigma, sigma)
npt.assert_almost_equal(fit.beta, beta)
# test receptive field
rf = generate_og_receptive_field(x, y, sigma, fit.model.stimulus.deg_x,
fit.model.stimulus.deg_y)
rf /= (2 * np.pi * sigma**2) * 1 / np.diff(model.stimulus.deg_x[0, 0:2])**2
npt.assert_almost_equal(np.round(rf.sum()),
np.round(fit.receptive_field.sum()))
# test model == fit RF
npt.assert_almost_equal(
np.round(fit.model.generate_receptive_field(x, y, sigma).sum()),
np.round(fit.receptive_field.sum()))
raise ValueError(
'BOLD Image and Stimulus JSON do not have the same number of data points'
)
fields = ('theta', 'rho', 'sigma', 'hrfdelay', 'beta', 'baseline')
res = {k: [] for k in fields}
res['pred'] = []
for (ii, vx, js) in zip(range(len(bold)), bold, stim_json):
stdat = js['Stimulus']
if pimms.is_list(stdat): stdat = stdat[0]
height = stdat['fieldofviewVert']
width = stdat['fieldofviewHorz']
### STIMULUS
# First get a viewing distance and screen size
dist = 100 # 100 cm is arbitrary
stim_width = 2 * dist * np.tan(np.pi / 180 * width / 2)
stimulus = VisualStimulus(stim, dist, stim_width, 1.0, float(js['TR']),
ctypes.c_int16)
model = og.GaussianModel(stimulus, utils.double_gamma_hrf)
### FIT
## define search grids
# these define min and max of the edge of the initial brute-force search.
x_grid = (-width / 2, width / 2)
y_grid = (-height / 2, height / 2)
s_grid = (1 / stimulus.ppd + 0.25, 5.25)
h_grid = (-1.0, 1.0)
## define search bounds
# these define the boundaries of the final gradient-descent search.
x_bound = (-width, width)
y_bound = (-height, height)
s_bound = (1 / stimulus.ppd, 12.0) # smallest sigma is a pixel
b_bound = (1e-8, None)
u_bound = (None, None)
def test_strf_fit():
viewing_distance = 38
screen_width = 25
thetas = np.tile(np.arange(0,360,90),2)
thetas = np.insert(thetas,0,-1)
thetas = np.append(thetas,-1)
num_blank_steps = 20
num_bar_steps = 20
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 200
pixels_across = 200
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1,2,3)
auto_fit = True
verbose = 1
projector_hz = 480
tau = 0.00875
mask_size = 5
hrf = 0.25
# create the sweeping bar stimulus in memory
stim = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(stim, viewing_distance, screen_width, scale_factor, tr_length, dtype)
stimulus.fps = projector_hz
flicker_vec = np.zeros_like(stim[0,0,:]).astype('uint8')
flicker_vec[1*20:5*20] = 1
flicker_vec[5*20:9*20] = 2
stimulus.flicker_vec = flicker_vec
stimulus.flicker_hz = [10,20]
# initialize the gaussian model
model = strf.SpatioTemporalModel(stimulus, utils.spm_hrf)
model.tau = tau
model.hrf_delay = hrf
model.mask_size = mask_size
# generate a random pRF estimate
x = -2.24
y = 1.58
sigma = 1.23
weight = 0.90
beta = 1.0
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, weight, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-8.0,7.0,5)
y_grid = utils.grid_slice(-8.0,7.0,5)
s_grid = utils.grid_slice(0.75,3.0,5)
w_grid = utils.grid_slice(0.05,0.95,5)
# set search bounds
x_bound = (-10,10)
y_bound = (-10,10)
s_bound = (1/stimulus.ppd,10)
w_bound = (1e-8,1.0)
b_bound = (1e-8,1e5)
u_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (x_grid, y_grid, s_grid, w_grid,)
bounds = (x_bound, y_bound, s_bound, w_bound, b_bound, u_bound)
# fit the response
fit = strf.SpatioTemporalFit(model, data, grids, bounds)
# coarse fit
ballpark = [-0.5,
3.25,
2.4375,
0.94999999999999996,
1.0292,
-0.24999999999999992]
npt.assert_almost_equal((fit.x0,fit.y0,fit.sigma0,fit.weight0,fit.beta0,fit.baseline0),ballpark,4)
# fine fit
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 2)
npt.assert_almost_equal(fit.weight, weight, 2)
npt.assert_almost_equal(fit.beta, beta, 2)
npt.assert_almost_equal(fit.baseline, baseline, 2)
# overloaded
npt.assert_almost_equal(fit.overloaded_estimate, [ 2.5272803, 2.7411676, 1.23 , 0.9 , 1. , -0.25 ], 2)
m_rf = fit.model.m_rf(fit.model.tau)
p_rf = fit.model.p_rf(fit.model.tau)
npt.assert_almost_equal(simps(np.abs(m_rf)),simps(p_rf),2)
# responses
m_resp = fit.model.generate_m_resp(fit.model.tau)
p_resp = fit.model.generate_p_resp(fit.model.tau)
npt.assert_(np.max(m_resp,0)[0]<np.max(m_resp,0)[1])
npt.assert_(np.max(p_resp,0)[0]>np.max(p_resp,0)[1])
# amps
npt.assert_(fit.model.m_amp[0]<fit.model.m_amp[1])
npt.assert_(fit.model.p_amp[0]>fit.model.p_amp[1])
# receptive field
rf = generate_og_receptive_field(x, y, sigma, fit.model.stimulus.deg_x, fit.model.stimulus.deg_y)
rf /= (2 * np.pi * sigma**2) * 1/np.diff(model.stimulus.deg_x[0,0:2])**2
npt.assert_almost_equal(np.round(rf.sum()), np.round(fit.receptive_field.sum()))
# test model == fit RF
npt.assert_almost_equal(np.round(fit.model.generate_receptive_field(x,y,sigma).sum()), np.round(fit.receptive_field.sum()))
dms = []
for ix, row in df.iterrows():
dm = create_visual_designmatrix_all(
bar_width=0.125,
iti_duration=0,
thetas=[row.direction],
n_pixels=n_pixels,
nr_timepoints=row.bar_pass_duration
) # Note that we make a design matrix at level of seconds
dms.append(dm)
dm = np.concatenate(dms, -1)
stimulus = VisualStimulus(dm, distance_screen, size_cm, 0.50, 1.0,
ctypes.c_int16)
hrf = utils.spm_hrf(0, 1.)
bold_dm = fftconvolve(stimulus.stim_arr, hrf[np.newaxis, np.newaxis, :],
'full')
bold_dm = np.moveaxis(bold_dm, -1, 0)[:total_duration]
bold_dm = resample(bold_dm, int(total_duration / TR), axis=0)
np.savetxt('/data/odc/derivatives/prf/dm.txt', dm.reshape(-1, n_pixels**2))
np.savetxt('/data/odc/derivatives/prf/bold_dm.txt',
bold_dm.reshape(-1, n_pixels**2))
np.save('/data/odc/derivatives/prf/dm.npy', dm)
# remove fixation point
dm[49, 88, :] = 0
dm[49, 89, :] = 0
dm[50, 88, :] = 0
dm[50, 89, :] = 0
#revert y axis
dm = dm[::-1, :, :] # this is how popeye wants y dim (0 point is top of dm)
########################################################################################
# setup popeye filtered css model
########################################################################################
stimulus = VisualStimulus(stim_arr=dm,
viewing_distance=225,
screen_width=69.84,
scale_factor=1,
tr_length=TR,
dtype=np.short)
model_func = CompressiveSpatialSummationModelFiltered(
stimulus=stimulus,
hrf_model=utils.spm_hrf,
sg_filter_window_length=120,
sg_filter_order=3,
tr=TR)
model_func.hrf_delay = hrf_delay
print 'loading data'
# load data
def test_dog():
# stimulus features
viewing_distance = 31
screen_width = 41
thetas = np.arange(0, 360, 90)
# thetas = np.insert(thetas,0,-1)
# thetas = np.append(thetas,-1)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
auto_fit = True
verbose = 0
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = dog.DifferenceOfGaussiansModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0
model.mask_size = 20
# set the pRF params
x = 2.2
y = 2.5
sigma = 0.90
sigma_ratio = 1.5
volume_ratio = 0.5
beta = 0.25
baseline = -0.10
# create "data"
data = model.generate_prediction(x, y, sigma, sigma_ratio, volume_ratio,
beta, baseline)
# set up the grids
x_grid = utils.grid_slice(-5, 5, 4)
y_grid = utils.grid_slice(-5, 5, 4)
s_grid = utils.grid_slice(1 / stimulus.ppd0 * 1.10, 3.5, 4)
sr_grid = utils.grid_slice(1.0, 2.0, 4)
vr_grid = utils.grid_slice(0.10, 0.90, 4)
grids = (
x_grid,
y_grid,
s_grid,
sr_grid,
vr_grid,
)
# set up the bounds
x_bound = (-ecc, ecc)
y_bound = (-ecc, ecc)
s_bound = (1 / stimulus.ppd, 5)
sr_bound = (1.0, None)
vr_bound = (1e-8, 1.0)
bounds = (
x_bound,
y_bound,
s_bound,
sr_bound,
vr_bound,
)
# fit it
fit = dog.DifferenceOfGaussiansFit(model, data, grids, bounds)
# coarse fit
ballpark = [
1.666666666666667, 1.666666666666667, 2.8243187483428391,
1.9999999999999998, 0.10000000000000001
]
npt.assert_almost_equal((fit.x0, fit.y0, fit.s0, fit.sr0, fit.vr0),
ballpark)
# the baseline/beta should be 0/1 when regressed data vs. estimate
(m, b) = np.polyfit(fit.scaled_ballpark_prediction, data, 1)
npt.assert_almost_equal(m, 1.0)
npt.assert_almost_equal(b, 0.0)
# fine fit
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 2)
npt.assert_almost_equal(fit.sigma_ratio, sigma_ratio, 1)
npt.assert_almost_equal(fit.volume_ratio, volume_ratio, 1)
# test the RF
rf = fit.model.receptive_field(*fit.estimate[0:-2])
est = fit.estimate[0:-2].copy()
rf_new = fit.model.receptive_field(*est)
value_1 = np.sqrt(simps(simps(rf)))
value_2 = np.sqrt(simps(simps(rf_new)))
nt.assert_almost_equal(value_1, value_2)
# polar coordinates
npt.assert_almost_equal(
[fit.theta, fit.rho],
[np.arctan2(y, x), np.sqrt(x**2 + y**2)], 4)
def test_strf_css_fit():
viewing_distance = 38
screen_width = 25
thetas = np.tile(np.arange(0,360,90),2)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1,2,3)
auto_fit = True
verbose = 1
projector_hz = 480
tau = 0.00875
mask_size = 5
hrf = 0.25
# create the sweeping bar stimulus in memory
stim1 = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc, clip=0.33)
# create the sweeping bar stimulus in memory
stim2 = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc, clip=0.0001)
stim = np.concatenate((stim1,stim2),-1)
# create an instance of the Stimulus class
stimulus = VisualStimulus(stim, viewing_distance, screen_width, scale_factor, tr_length, dtype)
stimulus.fps = projector_hz
flicker_vec = np.zeros_like(stim1[0,0,:]).astype('uint8')
flicker_vec[1*20:5*20] = 1
flicker_vec[5*20:9*20] = 2
flicker_vec = np.tile(flicker_vec,2)
stimulus.flicker_vec = flicker_vec
stimulus.flicker_hz = [10,20,10,20]
# initialize the gaussian model
model = strf.SpatioTemporalModel(stimulus, utils.spm_hrf)
model.tau = tau
model.hrf_delay = hrf
model.mask_size = mask_size
# generate a random pRF estimate
x = -2.24
y = 1.58
sigma = 1.23
n = 0.90
weight = 0.95
beta = 0.88
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, n, weight, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-8.0,7.0,4)
y_grid = utils.grid_slice(-8.0,7.0,4)
s_grid = utils.grid_slice(0.75,3.0,4)
n_grid = utils.grid_slice(0.25,0.95,4)
w_grid = utils.grid_slice(0.25,0.95,4)
# set search bounds
x_bound = (-10,10)
y_bound = (-10,10)
s_bound = (1/stimulus.ppd,10)
n_bound = (1e-8,1.0-1e-8)
w_bound = (1e-8,1.0-1e-8)
b_bound = (1e-8,1e5)
u_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (x_grid, y_grid, s_grid, n_grid, w_grid,)
bounds = (x_bound, y_bound, s_bound, n_bound, w_bound, b_bound, u_bound)
# fit the response
fit = strf.SpatioTemporalFit(model, data, grids, bounds)
# coarse fit
ballpark = [-3.0,
2.0,
1.5,
0.95,
0.95,
0.88574075,
-0.25]
npt.assert_almost_equal((fit.x0,fit.y0,fit.sigma0, fit.n0, fit.weight0,fit.beta0,fit.baseline0),ballpark)
# fine fit
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 1)
npt.assert_almost_equal(fit.n, n, 2)
npt.assert_almost_equal(fit.weight, weight, 2)
npt.assert_almost_equal(fit.beta, beta, 2)
npt.assert_almost_equal(fit.baseline, baseline, 2)
# overloaded
npt.assert_almost_equal(fit.overloaded_estimate,[2.5266437, 2.7390143, 1.3014282, 0.9004958, 0.9499708, 0.8801774], 2)
# rfs
m_rf = fit.model.m_rf(fit.model.tau)
p_rf = fit.model.p_rf(fit.model.tau)
npt.assert_almost_equal(simps(np.abs(m_rf)),simps(p_rf),5)
# responses
m_resp = fit.model.generate_m_resp(fit.model.tau)
p_resp = fit.model.generate_p_resp(fit.model.tau)
npt.assert_(np.max(m_resp,0)[0]<np.max(m_resp,0)[1])
npt.assert_(np.max(p_resp,0)[0]>np.max(p_resp,0)[1])
# amps
npt.assert_(fit.model.m_amp[0]<fit.model.m_amp[1])
npt.assert_(fit.model.p_amp[0]>fit.model.p_amp[1])
# receptive field
npt.assert_almost_equal(4.0, fit.receptive_field.sum())
def test_css_fit():
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 90)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 50
pixels_across = 50
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = css.CompressiveSpatialSummationModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0.2
# generate a random pRF estimate
x = -2.24
y = 1.58
sigma = 1.23
n = 0.90
beta = 1.0
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, n, beta, baseline)
# set search grid
x_grid = (-3, 2)
y_grid = (-3, 2)
s_grid = (1 / stimulus.ppd, 2.75)
n_grid = (0.1, 0.90)
# set search bounds
x_bound = (-10, 10)
y_bound = (-10, 10)
s_bound = (1 / stimulus.ppd, 10)
n_bound = (1e-8, 1.0)
b_bound = (1e-8, 1e5)
h_bound = (-3.0, 3.0)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
n_grid,
)
bounds = (
x_bound,
y_bound,
s_bound,
n_bound,
b_bound,
)
# fit the response
fit = css.CompressiveSpatialSummationFit(model, data, grids, bounds, Ns=Ns)
# coarse fit
observed = [fit.x0, fit.y0, fit.s0, fit.n0, fit.beta0, fit.baseline0]
expected = [-3., 2., 0.72833938, 0.5, 0.9358213, -0.24999999999999997]
npt.assert_almost_equal(observed, expected)
# fine fit
npt.assert_almost_equal(fit.x, x)
npt.assert_almost_equal(fit.y, y)
npt.assert_almost_equal(fit.sigma, sigma)
npt.assert_almost_equal(fit.n, n)
npt.assert_almost_equal(fit.beta, beta)
npt.assert_almost_equal(fit.beta, beta)
# overloaded
npt.assert_almost_equal(fit.overloaded_estimate, [
2.5272803327893043, 2.7411676344215277, 1.2965338406691291,
0.90000000000036384, 0.99999999999999067, -0.25000000000200889
])
def test_bounded_amplitude_failure():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 90)
thetas = np.insert(thetas, 0, -1)
thetas = np.append(thetas, -1)
num_blank_steps = 30
num_bar_steps = 30
ecc = 12
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 1.0
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.double_gamma_hrf,
utils.percent_change)
model.hrf_delay = 0
model.mask_size = 6
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
beta = -0.25
baseline = 0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-10, 10, 5)
y_grid = utils.grid_slice(-10, 10, 5)
s_grid = utils.grid_slice(0.25, 5.25, 5)
# set search bounds
x_bound = (-12.0, 12.0)
y_bound = (-12.0, 12.0)
s_bound = (0.001, 12.0)
b_bound = (1e-8, None)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# fit the response
fit = og.GaussianFit(model, data, grids, bounds)
nt.assert_true(fit.model.bounded_amplitude)
nt.assert_true(fit.slope > 0)
nt.assert_true(fit.beta0 > 0)
nt.assert_true(fit.beta > 0)
nt.assert_true(fit.beta != beta)
def test_recast_estimation_results():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 45)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.10
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
beta = 2.5
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-5, 4, 5)
y_grid = utils.grid_slice(-5, 7, 5)
s_grid = utils.grid_slice(1 / stimulus.ppd, 5.25, 5)
b_grid = utils.grid_slice(0.1, 4.0, 5)
# set search bounds
x_bound = (-12.0, 12.0)
y_bound = (-12.0, 12.0)
s_bound = (1 / stimulus.ppd, 12.0)
b_bound = (1e-8, 1e2)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# create 3 voxels of data
all_data = np.array([data, data, data])
indices = [(0, 0, 0), (0, 0, 1), (0, 0, 2)]
# bundle the voxels
bundle = utils.multiprocess_bundle(og.GaussianFit, model, all_data, grids,
bounds, indices)
# run analysis
with sharedmem.Pool(np=3) as pool:
output = pool.map(utils.parallel_fit, bundle)
# create grid parent
arr = np.zeros((1, 1, 3))
grid_parent = nibabel.Nifti1Image(arr, np.eye(4, 4))
# recast the estimation results
nif = utils.recast_estimation_results(output, grid_parent)
dat = nif.get_data()
# assert equivalence
npt.assert_almost_equal(np.mean(dat[..., 0]), x)
npt.assert_almost_equal(np.mean(dat[..., 1]), y)
npt.assert_almost_equal(np.mean(dat[..., 2]), sigma)
npt.assert_almost_equal(np.mean(dat[..., 3]), beta)
npt.assert_almost_equal(np.mean(dat[..., 4]), baseline)
# recast the estimation results - OVERLOADED
nif = utils.recast_estimation_results(output, grid_parent, True)
dat = nif.get_data()
# assert equivalence
npt.assert_almost_equal(np.mean(dat[..., 0]), np.arctan2(y, x), 2)
npt.assert_almost_equal(np.mean(dat[..., 1]), np.sqrt(x**2 + y**2), 2)
npt.assert_almost_equal(np.mean(dat[..., 2]), sigma)
npt.assert_almost_equal(np.mean(dat[..., 3]), beta)
npt.assert_almost_equal(np.mean(dat[..., 4]), baseline)
rsq_mask_crossv = np.mean(prf_data[:, :, -1], axis=1) > rsq_threshold
############################################################################################################################################
# setting up prf timecourses - NOTE, this is for the 'all' situation, so should be really done on a run-by-run basis using a run's
# loo data and prf parameters. A test set would then be taken from the single_run data as this hasn't been used for that run's fit.
############################################################################################################################################
# set up model with hrf etc.
def my_spmt(delay, tr):
return spmt(np.arange(0, 33, tr))
# we're going to use these popeye convenience functions
# because they are fast, and because they were used in the fitting procedure
stimulus = VisualStimulus(dm_crossv, screen_distance, screen_width, 1.0 / 3.0,
TR, ctypes.c_int16)
css_model = CompressiveSpatialSummationModelFiltered(stimulus, my_spmt)
css_model.hrf_delay = 0
############################################################################################################################################
# setting up prf spatial profiles for subsequent covariances, now some per-run stuff was done
############################################################################################################################################
#use this number to choose which section to use in crossvalidate setup
crossv = 0
deg_x, deg_y = np.meshgrid(
np.linspace(extent[0], extent[1], n_pix, endpoint=True),
np.linspace(extent[0], extent[1], n_pix, endpoint=True))
rfs = generate_og_receptive_fields(prf_data[rsq_mask_crossv, crossv, 0],
prf_data[rsq_mask_crossv, crossv, 1],
def test_strf_hrf_fit():
viewing_distance = 38
screen_width = 25
thetas = np.tile(np.arange(0, 360, 90), 2)
thetas = np.insert(thetas, 0, -1)
thetas = np.append(thetas, -1)
num_blank_steps = 20
num_bar_steps = 20
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 200
pixels_across = 200
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
projector_hz = 480
tau = 0.00875
mask_size = 5
# create the sweeping bar stimulus in memory
stim = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(stim, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
stimulus.fps = projector_hz
flicker_vec = np.zeros_like(stim[0, 0, :]).astype('uint8')
flicker_vec[1 * 20:5 * 20] = 1
flicker_vec[5 * 20:9 * 20] = 2
stimulus.flicker_vec = flicker_vec
stimulus.flicker_hz = [10, 20]
# initialize the gaussian model
model = strf.SpatioTemporalModel(stimulus, utils.double_gamma_hrf)
model.tau = tau
model.mask_size = mask_size
# generate a random pRF estimate
x = -2.24
y = 1.58
sigma = 1.23
weight = 0.90
hrf_delay = -0.13
beta = 1.0
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, weight, hrf_delay, beta,
baseline)
# set search grid
x_grid = utils.grid_slice(-8.0, 7.0, 3)
y_grid = utils.grid_slice(-8.0, 7.0, 3)
s_grid = utils.grid_slice(0.75, 3.0, 3)
w_grid = utils.grid_slice(0.05, 0.95, 3)
h_grid = utils.grid_slice(-0.25, 0.25, 3)
# set search bounds
x_bound = (-10, 10)
y_bound = (-10, 10)
s_bound = (1 / stimulus.ppd, 10)
w_bound = (1e-8, 1.0)
b_bound = (1e-8, 1e5)
u_bound = (None, None)
h_bound = (-2.0, 2.0)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
w_grid,
h_grid,
)
bounds = (x_bound, y_bound, s_bound, w_bound, h_bound, b_bound, u_bound)
# fit the response
fit = strf.SpatioTemporalFit(model, data, grids, bounds)
# coarse fit
npt.assert_almost_equal((fit.x0, fit.y0, fit.sigma0, fit.weight0, fit.hrf0,
fit.beta0, fit.baseline0),
[-0.5, -0.5, 3., 0.95, -0.25, 1., 0.02], 2)
# fine fit
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 2)
npt.assert_almost_equal(fit.weight, weight, 2)
npt.assert_almost_equal(fit.beta, beta, 2)
npt.assert_almost_equal(fit.baseline, baseline, 2)
# overloaded
npt.assert_almost_equal(fit.overloaded_estimate,
[2.53, 2.74, 1.23, 0.9, 5.87, 1., -0.25], 2)
m_rf = fit.model.m_rf(fit.model.tau)
p_rf = fit.model.p_rf(fit.model.tau)
npt.assert_almost_equal(simps(np.abs(m_rf)), simps(p_rf), 5)
# responses
m_resp = fit.model.generate_m_resp(fit.model.tau)
p_resp = fit.model.generate_p_resp(fit.model.tau)
npt.assert_(np.max(m_resp, 0)[0] < np.max(m_resp, 0)[1])
npt.assert_(np.max(p_resp, 0)[0] > np.max(p_resp, 0)[1])
# amps
npt.assert_(fit.model.m_amp[0] < fit.model.m_amp[1])
npt.assert_(fit.model.p_amp[0] > fit.model.p_amp[1])
# receptive field
npt.assert_almost_equal(4.0, fit.receptive_field.sum())
def test_strf_css_fit():
viewing_distance = 38
screen_width = 25
thetas = np.tile(np.arange(0,360,90),2)
thetas = np.insert(thetas,0,-1)
thetas = np.append(thetas,-1)
num_blank_steps = 20
num_bar_steps = 20
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
Ns = 3
voxel_index = (1,2,3)
auto_fit = True
verbose = 1
projector_hz = 480
tau = 0.00875
mask_size = 5
hrf = 0.25
# create the sweeping bar stimulus in memory
stim1 = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc, clip=0.33)
# create the sweeping bar stimulus in memory
stim2 = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps, num_blank_steps, ecc, clip=0.0001)
stim = np.concatenate((stim1,stim2),-1)
# create an instance of the Stimulus class
stimulus = VisualStimulus(stim, viewing_distance, screen_width, scale_factor, tr_length, dtype)
stimulus.fps = projector_hz
flicker_vec = np.zeros_like(stim1[0,0,:]).astype('uint8')
flicker_vec[1*20:5*20] = 1
flicker_vec[5*20:9*20] = 2
flicker_vec = np.tile(flicker_vec,2)
stimulus.flicker_vec = flicker_vec
stimulus.flicker_hz = [10,20,10,20]
# initialize the gaussian model
model = strf.SpatioTemporalModel(stimulus, utils.double_gamma_hrf)
model.tau = tau
model.hrf_delay = hrf
model.mask_size = mask_size
# generate a random pRF estimate
x = -2.24
y = 1.58
sigma = 1.23
n = 0.90
weight = 0.95
beta = 1.0
baseline = 0
# create the "data"
data = model.generate_prediction(x, y, sigma, n, weight, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-8.0,7.0,3)
y_grid = utils.grid_slice(-8.0,7.0,3)
s_grid = utils.grid_slice(0.75,3.0,3)
n_grid = utils.grid_slice(0.25,0.95,3)
w_grid = utils.grid_slice(0.25,0.95,3)
# set search bounds
x_bound = (-10,10)
y_bound = (-10,10)
s_bound = (1/stimulus.ppd,10)
n_bound = (1e-8,1.0-1e-8)
w_bound = (1e-8,1.0-1e-8)
b_bound = (1e-8,1e5)
u_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (x_grid, y_grid, s_grid, n_grid, w_grid,)
bounds = (x_bound, y_bound, s_bound, n_bound, w_bound, b_bound, u_bound)
# fit the response
fit = strf.SpatioTemporalFit(model, data, grids, bounds)
# coarse fit
npt.assert_almost_equal((fit.x0,fit.y0,fit.sigma0, fit.n0, fit.weight0,fit.beta0,fit.baseline0),[-0.5 , -0.5 , 1.875, 0.95 , 0.95 , 1. , 0. ])
# fine fit
npt.assert_almost_equal(fit.x, x, 1)
npt.assert_almost_equal(fit.y, y, 1)
npt.assert_almost_equal(fit.sigma, sigma, 1)
npt.assert_almost_equal(fit.n, n, 1)
npt.assert_almost_equal(fit.weight, weight, 1)
npt.assert_almost_equal(fit.beta, beta, 1)
npt.assert_almost_equal(fit.baseline, baseline, 1)
# overloaded
npt.assert_almost_equal(fit.overloaded_estimate, [2.5259863707822303,
2.7330681871539069,
1.3062396482386418,
0.9011492100931614,
0.94990930073215352,
1.0005707740082497],4)
# rfs
m_rf = fit.model.m_rf(fit.model.tau)
p_rf = fit.model.p_rf(fit.model.tau)
npt.assert_almost_equal(simps(np.abs(m_rf)),simps(p_rf),5)
# responses
m_resp = fit.model.generate_m_resp(fit.model.tau)
p_resp = fit.model.generate_p_resp(fit.model.tau)
npt.assert_(np.max(m_resp,0)[0]<np.max(m_resp,0)[1])
npt.assert_(np.max(p_resp,0)[0]>np.max(p_resp,0)[1])
# amps
npt.assert_(fit.model.m_amp[0]<fit.model.m_amp[1])
npt.assert_(fit.model.p_amp[0]>fit.model.p_amp[1])
# receptive field
npt.assert_almost_equal(4.0, fit.receptive_field.sum())
def test_resurrect_model():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0,360,90)
thetas = np.insert(thetas,0,-1)
thetas = np.append(thetas,-1)
num_blank_steps = 20
num_bar_steps = 20
ecc = 10
tr_length = 1.5
frames_per_tr = 1.0
scale_factor = 1.0
pixels_across = 100
pixels_down = 100
dtype = ctypes.c_int16
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance, screen_width,
thetas, num_bar_steps, num_blank_steps, ecc, clip=0.01)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width, scale_factor, tr_length, dtype)
# set cache grids
x_grid = utils.grid_slice(-10, 10, 5)
y_grid = utils.grid_slice(-10, 10, 5)
s_grid = utils.grid_slice(0.55,5.25, 5)
grids = (x_grid, y_grid, s_grid,)
# set search bounds
x_bound = (-12.0,12.0)
y_bound = (-12.0,12.0)
s_bound = (0.001,12.0)
b_bound = (1e-8,None)
m_bound = (None,None)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.double_gamma_hrf)
model.hrf_delay = 0
model.mask_size = 5
cache = model.cache_model(grids, ncpus=3)
# seed rng
np.random.seed(4932)
# pluck an estimate and create timeseries
x, y, sigma = cache[51][1]
beta = 1.25
baseline = 0.25
# create "data"
data = cache[51][0]
# fit it
fit = og.GaussianFit(model, data, grids, bounds, verbose=0)
# assert
npt.assert_equal(fit.estimate,fit.ballpark)
# create "data"
data = model.generate_prediction(x,y,sigma,beta,baseline)
# fit it
fit = og.GaussianFit(model, data, grids, bounds, verbose=0)
# assert
npt.assert_almost_equal(np.sum(fit.scaled_ballpark_prediction-fit.data)**2,0)
def test_bootstrap():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.array([-1, 0, 90, 180, 270, -1])
num_blank_steps = 30
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.10
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
# rng
np.random.seed(2764932)
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
beta = 2.5
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-10, 10, 5)
y_grid = utils.grid_slice(-10, 10, 5)
s_grid = utils.grid_slice(0.5, 3.25, 5)
# set search bounds
x_bound = (-12.0, 12.0)
y_bound = (-12.0, 12.0)
s_bound = (0.001, 12.0)
b_bound = (1e-8, None)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# pack multiple "runs"
data = np.vstack((data, data, data))
# make it a singular "voxel"
data = np.reshape(data, (1, data.shape[0], data.shape[1]))
# set bootstraps and resamples
bootstraps = 2
resamples = np.array((2, ))
# make fodder
bundle = utils.bootstrap_bundle(bootstraps, resamples, og.GaussianFit,
model, data, grids, bounds,
np.tile((1, 2, 3), (bootstraps, 1)))
# test
for b in bundle:
fit = utils.parallel_bootstrap(b)
npt.assert_almost_equal(fit.rss, 0)
npt.assert_equal(fit.n_resamples, resamples[0])
npt.assert_equal(np.sum(fit.resamples),
np.sum(np.arange(resamples[0])))
def test_dog():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 90)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.50
pixels_down = 50
pixels_across = 50
dtype = ctypes.c_int16
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = dog.DifferenceOfGaussiansModel(stimulus, utils.spm_hrf)
model.hrf_delay = 0.2
# set the pRF params
x = -1.4
y = 1.5
sigma = 1.0
sigma_ratio = 2.0
volume_ratio = 0.5
hrf_delay = -0.2
# create "data"
data = model.generate_prediction(
x,
y,
sigma,
sigma_ratio,
volume_ratio,
)
# set up the grids
x_grid = slice(-10, 10, 3)
y_grid = slice(-10, 10, 3)
s_grid = slice(1 / stimulus.ppd, 5, 3)
sr_grid = slice(1.0, 5.0, 3)
vr_grid = slice(0.01, 0.99, 3)
grids = (
x_grid,
y_grid,
s_grid,
sr_grid,
vr_grid,
)
# set up the bounds
x_bound = (-ecc, ecc)
y_bound = (-ecc, ecc)
s_bound = (1 / stimulus.ppd, 5)
sr_bound = (1.0, None)
vr_bound = (1e-8, 1.0)
bounds = (
x_bound,
y_bound,
s_bound,
sr_bound,
vr_bound,
)
# fit it
fit = dog.DifferenceOfGaussiansFit(model, data, grids, bounds, voxel_index)
# coarse fit
nt.assert_almost_equal((fit.x0, fit.y0, fit.s0, fit.sr0, fit.vr0),
(-1.0, 2.0, 0.72833937882323319, 1.0, 0.01))
# fine fit
nt.assert_almost_equal(fit.x, x)
nt.assert_almost_equal(fit.y, y)
nt.assert_almost_equal(fit.sigma, sigma)
nt.assert_almost_equal(fit.sigma_ratio, sigma_ratio)
nt.assert_almost_equal(fit.volume_ratio, volume_ratio)
# test the RF
rf = fit.model.receptive_field(*fit.estimate)
est = fit.estimate.copy()
est[2] *= 2
rf_new = fit.model.receptive_field(*est)
value_1 = np.sqrt(simps(simps(rf)))
value_2 = np.sqrt(simps(simps(rf_new)))
nt.assert_almost_equal(value_2 / value_1, sigma_ratio, 1)
# polar coordinates
npt.assert_almost_equal(
[fit.theta, fit.rho],
[np.arctan2(y, x), np.sqrt(x**2 + y**2)])
def test_parallel_fit_manual_grids():
# stimulus features
viewing_distance = 38
screen_width = 25
thetas = np.arange(0, 360, 45)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1.0
frames_per_tr = 1.0
scale_factor = 0.10
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
# create the sweeping bar stimulus in memory
bar = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(bar, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
# initialize the gaussian model
model = og.GaussianModel(stimulus, utils.double_gamma_hrf)
model.hrf_delay = 0
# generate a random pRF estimate
x = -5.24
y = 2.58
sigma = 1.24
beta = 2.5
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, beta, baseline)
# set search grid
x_grid = slice(-5, 4, 5)
y_grid = slice(-5, 7, 5)
s_grid = slice(1 / stimulus.ppd, 5.25, 5)
b_grid = slice(0.1, 4.0, 5)
# set search bounds
x_bound = (-12.0, 12.0)
y_bound = (-12.0, 12.0)
s_bound = (1 / stimulus.ppd, 12.0)
b_bound = (1e-8, 1e2)
m_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
)
bounds = (x_bound, y_bound, s_bound, b_bound, m_bound)
# make 3 voxels
all_data = np.array([data, data, data])
num_voxels = data.shape[0]
indices = [(1, 2, 3)] * 3
# bundle the voxels
bundle = utils.multiprocess_bundle(og.GaussianFit, model, all_data, grids,
bounds, indices)
# run analysis
with sharedmem.Pool(np=3) as pool:
output = pool.map(utils.parallel_fit, bundle)
# assert equivalence
for fit in output:
npt.assert_almost_equal(fit.x, x, 2)
npt.assert_almost_equal(fit.y, y, 2)
npt.assert_almost_equal(fit.sigma, sigma, 2)
npt.assert_almost_equal(fit.beta, beta, 2)
npt.assert_almost_equal(fit.baseline, baseline, 2)
import numpy as np import sharedmem import popeye.og_hrf as og import popeye.utilities as utils from popeye.visual_stimulus import VisualStimulus, simulate_bar_stimulus # seed random number generator so we get the same answers ... np.random.seed(2764932) ### STIMULUS ## create sweeping bar stimulus sweeps = np.array([-1, 0, 90, 180, 270, -1]) # in degrees, -1 is blank bar = simulate_bar_stimulus(100, 100, 40, 20, sweeps, 30, 30, 10) ## create an instance of the Stimulus class stimulus = VisualStimulus(bar, 50, 25, 0.50, 1.0, ctypes.c_int16) ### MODEL ## initialize the gaussian model model = og.GaussianModel(stimulus, utils.double_gamma_hrf) ## generate a random pRF estimate x = -5.24 y = 2.58 sigma = 1.24 hrf_delay = -0.25 beta = 0.55 baseline = -0.88 x = 6 y = 6
def test_strf_2dcos_fit():
viewing_distance = 38
screen_width = 25
thetas = np.tile(np.arange(0, 360, 90), 2)
thetas = np.insert(thetas, 0, -1)
thetas = np.append(thetas, -1)
num_blank_steps = 0
num_bar_steps = 30
ecc = 10
tr_length = 1
frames_per_tr = 1
scale_factor = 1.0
pixels_down = 100
pixels_across = 100
dtype = ctypes.c_int16
Ns = 5
voxel_index = (1, 2, 3)
auto_fit = True
verbose = 1
projector_hz = 480
tau = 0.00875
mask_size = 5
hrf = 0.25
# create the sweeping bar stimulus in memory
stim = simulate_bar_stimulus(pixels_across, pixels_down, viewing_distance,
screen_width, thetas, num_bar_steps,
num_blank_steps, ecc)
# create an instance of the Stimulus class
stimulus = VisualStimulus(stim, viewing_distance, screen_width,
scale_factor, tr_length, dtype)
stimulus.fps = projector_hz
flicker_vec = np.zeros_like(stim[0, 0, :]).astype('uint8')
flicker_vec[1 * 20:5 * 20] = 1
flicker_vec[5 * 20:9 * 20] = 2
stimulus.flicker_vec = flicker_vec
stimulus.flicker_hz = [10, 20]
# initialize the gaussian model
model = strf.SpatioTemporalModel(stimulus, utils.spm_hrf)
model.tau = tau
model.hrf_delay = hrf
model.mask_size = mask_size
model.power = 0.7
# generate a random pRF estimate
x = -2.24
y = 1.58
sigma = 1.23
weight = 0.90
beta = 1.0
baseline = -0.25
# create the "data"
data = model.generate_prediction(x, y, sigma, weight, beta, baseline)
# set search grid
x_grid = utils.grid_slice(-8.0, 7.0, 5)
y_grid = utils.grid_slice(-8.0, 7.0, 5)
s_grid = utils.grid_slice(0.75, 3.0, 5)
w_grid = utils.grid_slice(0.05, 0.95, 5)
# set search bounds
x_bound = (-10, 10)
y_bound = (-10, 10)
s_bound = (1 / stimulus.ppd, 10)
w_bound = (1e-8, 1.0)
b_bound = (1e-8, 1e5)
u_bound = (None, None)
# loop over each voxel and set up a GaussianFit object
grids = (
x_grid,
y_grid,
s_grid,
w_grid,
)
bounds = (x_bound, y_bound, s_bound, w_bound, b_bound, u_bound)
# fit the response
fit = strf.SpatioTemporalFit(model, data, grids, bounds)
# coarse fit
ballpark = [
-0.5, 3.25, 3.0, 0.72499999999999998, 0.858317, -0.25000000000000011
]
npt.assert_almost_equal(
(fit.x0, fit.y0, fit.sigma0, fit.weight0, fit.beta0, fit.baseline0),
ballpark)
# fine fit
npt.assert_almost_equal(fit.x, x)
npt.assert_almost_equal(fit.y, y)
npt.assert_almost_equal(fit.sigma, sigma)
npt.assert_almost_equal(fit.weight, weight)
npt.assert_almost_equal(fit.beta, beta)
npt.assert_almost_equal(fit.baseline, baseline)
# overloaded
npt.assert_almost_equal(fit.overloaded_estimate, [
2.5272803327887128, 2.7411676344185993, 1.2300000000008835,
0.89999999999333258, 1.0000000000005003, -0.25000000000063088
])
m_rf = fit.model.m_rf(fit.model.tau)
p_rf = fit.model.p_rf(fit.model.tau)
npt.assert_almost_equal(simps(np.abs(m_rf)), simps(p_rf), 5)
# responses
m_resp = fit.model.generate_m_resp(fit.model.tau)
p_resp = fit.model.generate_p_resp(fit.model.tau)
npt.assert_(np.max(m_resp, 0)[0] < np.max(m_resp, 0)[1])
npt.assert_(np.max(p_resp, 0)[0] > np.max(p_resp, 0)[1])
# amps
npt.assert_(fit.model.m_amp[0] < fit.model.m_amp[1])
npt.assert_(fit.model.p_amp[0] > fit.model.p_amp[1])
# receptive field
rf = generate_2dcos_receptive_field(x, y, sigma, fit.model.power,
fit.model.stimulus.deg_x,
fit.model.stimulus.deg_y)
rf /= (2 * np.pi * sigma**2) * 1 / np.diff(model.stimulus.deg_x[0, 0:2])**2
npt.assert_almost_equal(np.round(rf.sum()),
np.round(fit.receptive_field.sum()))
# test model == fit RF
npt.assert_almost_equal(
np.round(fit.model.generate_receptive_field(x, y, sigma).sum()),
np.round(fit.receptive_field.sum()))
# visual_dm = np.vstack(visual_dm).transpose((1,2,0))
# the code above recreates Kendrick's design matrices, we'll now just load them.
visual_dm = []
# for i in [1,3,5]:
for i in [5]:
file = tables.open_file(
os.path.join(base_dir, 'retinotopysmall{i}.mat'.format(i=i)))
visual_dm.append(file.get_node('/stim')[:])
file.close()
visual_dm = np.vstack(visual_dm).transpose((1, 2, 0))
stimulus = VisualStimulus(stim_arr=visual_dm,
viewing_distance=analysis_info["screen_distance"],
screen_width=analysis_info["screen_width"],
scale_factor=0.05,
tr_length=analysis_info["TR"],
dtype=np.short)
############################################################################################################################################
#
# load cii data, timepoints by grayordinates
#
############################################################################################################################################
# averaged_runs = ['CCW','EXP','RETBAR1']
averaged_runs = ['RETBAR1']
gii_files = [
glob.glob(
os.path.join(
subject_folder,
