def get_bold():
# TODO add second model
hrf_x = np.linspace(0, 25, 250)
hrf = double_gamma_hrf(hrf_x) - single_gamma_hrf(hrf_x, 0.8, 1, 0.05)
samples = 1200
exp_time = np.linspace(0, 120, samples)
fast_er_onsets = np.array([50, 240, 340, 590, 640, 940, 960])
fast_er = np.zeros(samples)
fast_er[fast_er_onsets] = 1
model_hr = np.convolve(fast_er, hrf)[:samples]
tr = 2.0
model_lr = signal.resample(model_hr, int(samples / tr / 10), window='ham')
## moderate noise level
baseline = 800
wsignal = baseline + 8.0 \
* model_lr + np.random.randn(int(samples / tr / 10)) * 4.0
nsignal = baseline \
+ np.random.randn(int(samples / tr / 10)) * 4.0
ds = Dataset(samples=np.array([wsignal, nsignal]).T,
sa={'model': model_lr})
return ds
def simple_hrf_dataset(events=None,
hrf_gen=lambda t: double_gamma_hrf(t) -
single_gamma_hrf(t, 0.8, 1, 0.05),
fir_length=15,
nsamples=None,
tr=2.0,
tres=1,
baseline=800.0,
signal_level=1,
noise='normal',
noise_level=1,
resampling='scipy'):
"""
events: list of Events or ndarray of onsets for simple(r) designs
"""
if events is None:
events = [1, 20, 25, 50, 60, 90, 92, 140]
if isinstance(events, np.ndarray) or not isinstance(events[0], dict):
events = [Event(onset=o) for o in events]
else:
assert (isinstance(events, list))
for e in events:
assert (isinstance(e, dict))
# play fmri
# full-blown HRF with initial dip and undershoot ;-)
hrf_x = np.arange(0, float(fir_length) * tres, tres)
if isinstance(hrf_gen, np.ndarray):
# just accept provided HRF and only verify size match
assert (len(hrf_x) == len(hrf_gen))
hrf = hrf_gen
else:
# actually generate it
hrf = hrf_gen(hrf_x)
if not nsamples:
# estimate number of samples needed if not provided
max_onset = max([e['onset'] for e in events])
nsamples = int(max_onset / tres + len(hrf_x) * 1.5)
# come up with an experimental design
fast_er = np.zeros(nsamples)
for e in events:
on = int(e['onset'] / float(tres))
off = int((e['onset'] + e.get('duration', 1.)) / float(tres))
if off == on:
off += 1 # so we have at least 1 point
assert (range(on, off))
fast_er[on:off] = e.get('intensity', 1)
# high resolution model of the convolved regressor
model_hr = np.convolve(fast_er, hrf)[:nsamples]
# downsample the regressor to fMRI resolution
if resampling == 'scipy':
from scipy import signal
model_lr = signal.resample(model_hr,
int(tres * nsamples / tr),
window='ham')
elif resampling == 'naive':
if tr % tres != 0.0:
raise ValueError("You must use resample='scipy' since your TR=%.2g"
" is not multiple of tres=%.2g" % (tr, tres))
if tr < tres:
raise ValueError("You must use resample='scipy' since your TR=%.2g"
" is less than tres=%.2g" % (tr, tres))
step = int(tr // tres)
model_lr = model_hr[::step]
else:
raise ValueError("resampling can only be 'scipy' or 'naive'. Got %r" %
resampling)
# generate artifical fMRI data: two voxels one is noise, one has
# something
wsignal = baseline + model_lr * signal_level
nsignal = np.ones(wsignal.shape) * baseline
# build design matrix: bold-regressor and constant
design = np.array([model_lr, np.repeat(1, len(model_lr))]).T
# two 'voxel' dataset
ds = dataset_wizard(samples=np.array((wsignal, nsignal)).T, targets=1)
ds.a['baseline'] = baseline
ds.a['tr'] = tr
ds.sa['design'] = design
ds.fa['signal_level'] = [signal_level, False]
if noise == 'autocorrelated':
# this one seems to be quite unstable and can provide really
# funky noise at times
noise = autocorrelated_noise(ds,
1 / tr,
1 / (2 * tr),
lfnl=noise_level,
hfnl=noise_level,
add_baseline=False)
elif noise == 'normal':
noise = np.random.randn(*ds.shape) * noise_level
else:
raise ValueError(noise)
ds.sa['noise'] = noise
ds.samples += noise
return ds
def simple_hrf_dataset(events=[1, 20, 25, 50, 60, 90, 92, 140],
hrf_gen=lambda t:double_gamma_hrf(t) - single_gamma_hrf(t, 0.8, 1, 0.05),
fir_length=15,
nsamples=None,
tr=2.0,
tres=1,
baseline=800.0,
signal_level=1,
noise='normal',
noise_level=1,
resampling='scipy',
):
"""
events: list of Events or ndarray of onsets for simple(r) designs
"""
if isinstance(events, np.ndarray) or not isinstance(events[0], dict):
events = [Event(onset=o) for o in events]
else:
assert(isinstance(events, list))
for e in events:
assert(isinstance(e, dict))
# play fmri
# full-blown HRF with initial dip and undershoot ;-)
hrf_x = np.arange(0, float(fir_length)*tres, tres)
if isinstance(hrf_gen, np.ndarray):
# just accept provided HRF and only verify size match
assert(len(hrf_x) == len(hrf_gen))
hrf = hrf_gen
else:
# actually generate it
hrf = hrf_gen(hrf_x)
if not nsamples:
# estimate number of samples needed if not provided
max_onset = max([e['onset'] for e in events])
nsamples = int(max_onset/tres + len(hrf_x)*1.5)
# come up with an experimental design
fast_er = np.zeros(nsamples)
for e in events:
on = int(e['onset'] / float(tres))
off = int((e['onset'] + e.get('duration', 1.)) / float(tres))
if off == on:
off += 1 # so we have at least 1 point
assert(range(on, off))
fast_er[on:off] = e.get('intensity', 1)
# high resolution model of the convolved regressor
model_hr = np.convolve(fast_er, hrf)[:nsamples]
# downsample the regressor to fMRI resolution
if resampling == 'scipy':
from scipy import signal
model_lr = signal.resample(model_hr,
int(tres * nsamples / tr),
window='ham')
elif resampling == 'naive':
if tr % tres != 0.0:
raise ValueError("You must use resample='scipy' since your TR=%.2g"
" is not multiple of tres=%.2g" % (tr, tres))
if tr < tres:
raise ValueError("You must use resample='scipy' since your TR=%.2g"
" is less than tres=%.2g" % (tr, tres))
step = int(tr // tres)
model_lr = model_hr[::step]
else:
raise ValueError("resampling can only be 'scipy' or 'naive'. Got %r"
% resampling)
# generate artifical fMRI data: two voxels one is noise, one has
# something
wsignal = baseline + model_lr*signal_level
nsignal = np.ones(wsignal.shape) * baseline
# build design matrix: bold-regressor and constant
design = np.array([model_lr, np.repeat(1, len(model_lr))]).T
# two 'voxel' dataset
ds = dataset_wizard(samples=np.array((wsignal, nsignal)).T, targets=1)
ds.a['baseline'] = baseline
ds.a['tr'] = tr
ds.sa['design'] = design
ds.fa['signal_level'] = [signal_level, False]
if noise == 'autocorrelated':
# this one seems to be quite unstable and can provide really
# funky noise at times
noise = autocorrelated_noise(ds, 1/tr, 1/(2*tr),
lfnl=noise_level, hfnl=noise_level,
add_baseline=False)
elif noise == 'normal':
noise = np.random.randn(*ds.shape) * noise_level
else:
raise ValueError(noise)
ds.sa['noise'] = noise
ds.samples += noise
return ds