if site_correction == 'MGLM':
# pull out site variable and convert to one-hot
connectivity_data = data
enc = OneHotEncoder(sparse=False)
enc.fit(site.reshape(-1, 1))
site_one_hot = enc.transform(site.reshape(-1, 1))
site_one_hot = np.transpose(site_one_hot)
# transpose and reshape the connectivity data
data_perm = np.reshape(connectivity_data,
(n_subjects, n_regions, n_regions))
data_perm = np.transpose(data_perm, (1, 2, 0))
# generate MGLM regressed data
p, V, E, Y_hat, gnorm = mglm_spd(site_one_hot, data_perm, 100)
# calculate matrix 'residuals'
for i in range(n_subjects):
mglm_corrected_matrix = Y_hat[:, :, i]
mglm_residuals_matrix = logmap_spd(mglm_corrected_matrix,
data_perm[:, :, i])
mglm_residuals_vector = np.reshape(mglm_residuals_matrix,
(n_regions * n_regions, ))
connectivity_data[i, :] = mglm_residuals_vector
# set up MCCV
n_repeats = 200
test_fraction = 0.2
train_inds_all, test_inds_all, test_size = ShuffleSplitFixed(
# join metadata and data
connectivity_metadata_data = pd.merge(definitive_subject_metadata,
connectivity_subjects_data,
left_on=0,
right_on='subject')
# get number of subjects
n_subjects = len(connectivity_metadata_data)
# pull out connectivity data
connectivity_data = connectivity_metadata_data.iloc[:, 8:].as_matrix()
# get demographics
demographics = connectivity_metadata_data.iloc[:,
2:4].as_matrix().astype(float)
demographics = np.flip(demographics, 1)
# normalise predictors - optional
if normalise_demographics:
demographics = demographics - np.min(demographics, axis=0)
demographics = demographics / np.max(demographics, axis=0)
# transpose and reshape for mglm
demographics = np.transpose(demographics)
connectivity_data = np.transpose(connectivity_data)
connectivity_data = np.reshape(connectivity_data, (290, 290, 108))
# try to run MGLM...
p, V, E, Y_hat, gnorm = mglm_spd(demographics, connectivity_data, 10)