def convolve_column(df, col):
categorical = df[col].dtype.name == 'category'
arr = np.array(df[col])
time = np.array(df.timestamp)
time_cur = 0.
start = 0
out = []
for i in range(len(df)):
if i > 0 and time[i] < time[i-1]:
start = i
time_cur = 0.
if time[i] > time_cur:
time_cur += step
if categorical:
if i > start:
out.append(arr[i-1])
else:
## Take the initial category value from the first row in the series
out.append(arr[i])
else:
convolve = np.vectorize(lambda x: hrf(time_cur - x))
if i > start:
t_conv = convolve(time[start:i])
p_conv = (t_conv*np.nan_to_num(arr[start:i])).sum()
else:
p_conv = 0
out.append(p_conv)
if categorical:
out = pd.Series(out, dtype='category')
else:
out = pd.Series(np.array(out), dtype='float')
return out
def convolve(X, y, first_obs, last_obs, columns):
"""
Convolve predictors in **X** into timestamps in **y** using canonical HRF.
:param X: ``pandas`` ``DataFrame``; impulse (predictor) data.
:param y: ``pandas`` ``DataFrame``; response data.
:param y: ``numpy`` ``array``; integer indices of first observation in X from time series in y.
:param y: ``numpy`` ``array``; integer indices of last observation in X preceding each y in the same time series..
:param series_ids: ``list`` of ``str``; column names whose jointly unique values define unique time series.
:param columns: ``list`` of ``str``; column names to convolve.
:return: ``pandas`` ``DataFrame``; matrix of convolved predictors
"""
time_X = np.array(X.time)
time_y = np.array(y.time)
X = np.array(X[columns])
X_conv = np.zeros((y.shape[0], X.shape[1]))
for i in range(len(y)):
s, e = first_obs[i], last_obs[i]
hrf_weights = hrf(time_y[i] - time_X[s:e])[..., None]
X_conv[i] = (X[s:e] * hrf_weights).sum(axis=0)
X_conv = pd.DataFrame(X_conv, columns=columns, index=y.index)
y = pd.concat([y, X_conv], axis=1)
return y
def convolve_column(df, col):
categorical = df[col].dtype.name == 'category'
arr = np.array(df[col])
time = np.array(df.timestamp)
time_cur = 0.
start = 0
out = []
for i in range(len(df)):
if i > 0 and time[i] < time[i - 1]:
start = i
time_cur = 0.
if time[i] > time_cur:
time_cur += step
if categorical:
if i > start:
out.append(arr[i - 1])
else:
## Take the initial category value from the first row in the series
out.append(arr[i])
else:
convolve = np.vectorize(lambda x: hrf(time_cur - x))
if i > start:
t_conv = convolve(time[start:i])
p_conv = (t_conv * np.nan_to_num(arr[start:i])).sum()
else:
p_conv = 0
out.append(p_conv)
if categorical:
out = pd.Series(out, dtype='category')
else:
out = pd.Series(np.array(out), dtype='float')
return out
def convolve(X, y, first_obs, last_obs, columns):
time_X = np.array(X.time)
time_y = np.array(y.time)
X = np.array(X[columns])
X_conv = np.zeros((y.shape[0], X.shape[1]))
for i in range(len(y)):
if i == 0 or i % 1000 == 999 or i == len(y) - 1:
sys.stderr.write('\r%d/%d' % (i+1, len(y)))
sys.stderr.flush()
s, e = first_obs[i], last_obs[i]
hrf_weights = hrf(time_y[i] - time_X[s:e])[..., None]
X_conv[i] = (X[s:e] * hrf_weights).sum(axis=0)
X_conv = pd.DataFrame(X_conv, columns=columns, index=y.index)
y = pd.concat([y, X_conv], axis=1)
sys.stderr.write('\n')
return y
for i in range(len(series)):
df_cur = series[i]
if 'duration' in df_cur.columns:
duration = df_cur['duration']
else:
duration = None
X = df_cur[cols]
impulse_times = df_cur.time.values
max_response_time = int(np.ceil(df_cur.time.max()))
if max_response_time % 2 != 0:
max_response_time += 1
tr = np.arange(0, max_response_time // args.step)
response_times = tr * args.step + args.start
D = response_times[..., None] - impulse_times[None, ...]
G_mask = D >= 0
G = hrf(D)
G = np.where(G_mask, G, 0)
if duration is not None:
X = X.multiply(duration, axis=0)
X_conv = np.dot(G, X)
X_conv = pd.DataFrame(X_conv, columns=cols)
X_conv['time'] = response_times
X_conv['tr'] = tr
for col, val in zip(args.grouping_columns, series_names[i]):
X_conv[col] = val
out.append(X_conv)
out = pd.concat(out, axis=0)
out.reset_index(drop=True, inplace=True)
out['sampleid'] = 1
out.sampleid = out.groupby(args.grouping_columns).sampleid.cumsum().apply(
t = np.array(df.time)[..., None]
sys.stderr.write('Convolving sound power with canonical HRF...\n')
soundPowerHRF = []
for i in range(0, len(t), step):
sys.stderr.write('\rRows completed: %d/%d' % (i, len(t)))
sys.stderr.flush()
doc_ix_cur = docid_ix[i:i + step]
soundPowerHRF_cur = np.zeros((doc_ix_cur.shape[0], ))
impulse = power_padded[doc_ix_cur]
tau = tau_padded[doc_ix_cur]
t_cur = t[i:i + step]
valid = np.where(tau.sum(axis=1) > 0)
impulse = impulse[valid]
tau = tau[valid]
t_cur = t_cur[valid]
soundPowerHRF_cur[valid] = np.nan_to_num(
impulse * hrf(t_cur - tau)).sum(axis=1) / 441.
soundPowerHRF.append(soundPowerHRF_cur)
soundPowerHRF = np.concatenate(soundPowerHRF, axis=0)
sys.stderr.write('\n')
df['soundPowerHRF'] = soundPowerHRF
df.to_csv(sys.stdout, sep=' ', na_rep='NaN', index=False)
import sys, argparse, pandas as pd, numpy as np
from mvpa2.misc.data_generators import double_gamma_hrf as hrf
argparser = argparse.ArgumentParser(description='Convolve data table using HRF')
argparser.add_argument('data', type=str, help='Path to data table')
argparser.add_argument('-s', '--step', type=float, default=2.0, help='Step size (in seconds) between fMRI samples')
argparser.add_argument('-d', '--doc_names', nargs='+', default=['Boar', 'Aqua', 'MatchstickSeller', 'KingOfBirds', 'Elvis', 'MrSticky', 'HighSchool', 'Roswell', 'Tulips', 'Tourettes'], help='List of document names in input data')
args, unknown = argparser.parse_known_args()
step = float(args.step)
convolve = np.vectorize(lambda x: hrf(time_cur + step - x))
doc_names = args.doc_names
def get_docid(timeseries):
timeseries = np.array(timeseries)
docid = []
docix = 0
for i in range(len(timeseries)):
if i > 0 and timeseries[i] < timeseries[i-1]:
docix += 1
docid.append(doc_names[docix])
return pd.Series(docid).astype('category')
def convolve_column(df, col):
categorical = df[col].dtype.name == 'category'
arr = np.array(df[col])
time = np.array(df.timestamp)
time_cur = 0.
start = 0
default=2.0,
help='Step size (in seconds) between fMRI samples')
argparser.add_argument('-d',
'--doc_names',
nargs='+',
default=[
'Boar', 'Aqua', 'MatchstickSeller', 'KingOfBirds',
'Elvis', 'MrSticky', 'HighSchool', 'Roswell',
'Tulips', 'Tourettes'
],
help='List of document names in input data')
args, unknown = argparser.parse_known_args()
step = float(args.step)
convolve = np.vectorize(lambda x: hrf(time_cur + step - x))
doc_names = args.doc_names
def get_docid(timeseries):
timeseries = np.array(timeseries)
docid = []
docix = 0
for i in range(len(timeseries)):
if i > 0 and timeseries[i] < timeseries[i - 1]:
docix += 1
docid.append(doc_names[docix])
return pd.Series(docid).astype('category')
t = np.array(df.time)[..., None]
sys.stderr.write('Convolving sound power with canonical HRF...\n')
soundPowerHRF = []
for i in range(0, len(t), step):
sys.stderr.write('\rRows completed: %d/%d' %(i, len(t)))
sys.stderr.flush()
doc_ix_cur = docid_ix[i:i+step]
soundPowerHRF_cur = np.zeros((doc_ix_cur.shape[0],))
impulse = power_padded[doc_ix_cur]
tau = tau_padded[doc_ix_cur]
t_cur = t[i:i+step]
valid = np.where(tau.sum(axis=1) > 0)
impulse = impulse[valid]
tau = tau[valid]
t_cur = t_cur[valid]
soundPowerHRF_cur[valid] = np.nan_to_num(impulse * hrf(t_cur - tau)).sum(axis=1) / 441.
soundPowerHRF.append(soundPowerHRF_cur)
soundPowerHRF = np.concatenate(soundPowerHRF, axis=0)
sys.stderr.write('\n')
df['soundPowerHRF'] = soundPowerHRF
df.to_csv(sys.stdout, sep=' ', na_rep='nan', index=False)
