R.append(all_data['R'])
n_voxel, n_tr = data[0].shape
# Run HTFA with downloaded data
from brainiak.factoranalysis.htfa import HTFA
# uncomment below line to get help message on HTFA
#help(HTFA)
K = 5
htfa = HTFA(K=K,
n_subj=n_subj,
max_global_iter=5,
max_local_iter=2,
voxel_ratio=0.5,
tr_ratio=0.5,
max_voxel=n_voxel,
max_tr=n_tr,
verbose=True)
htfa.fit(data, R)
if rank == 0:
print("\n centers of global latent factors are:")
print(htfa.get_centers(htfa.global_posterior_))
print("\n widths of global latent factors are:")
widths = htfa.get_widths(htfa.global_posterior_)
print(widths)
print("\n stds of global latent RBF factors are:")
rbf_std = np.sqrt(widths/(2.0))
print(rbf_std)
def main():
parser = argparse.ArgumentParser(
description='Input: cluster info, Output: brain images')
parser.add_argument(
"json_file",
help='cluster information: key=cluster_number, value=name of images')
parser.add_argument("--K",
type=int,
default=5,
help='number of points on the brain')
parser.add_argument("--n",
type=int,
default=0,
help='number of combinations')
parser.add_argument("--voxel")
parser.add_argument("--tfa")
args = parser.parse_args()
json_file = args.json_file
out_dir = json_file.strip('.yml')
if not os.path.exists(out_dir):
os.makedirs(out_dir)
K = args.K
n = args.n
voxel = args.voxel
# read in image name to fmri data mapping info
onlyfiles = [
join('/Users/hyundonglee/Desktop/Order/', f)
for f in listdir('/Users/hyundonglee/Desktop/Order')
if isfile(join('/Users/hyundonglee/Desktop/Order/', f)) and '.txt' in f
]
fmri = defaultdict(list)
for f in onlyfiles:
with open(f, 'rb') as fh:
name = f.split('.')[0]
i = 3
for line in fh:
line = line.strip()
fn = name.split('/')[-1].split(
'sess')[0] + 'ses-' + name.split('/')[-1].split(
'sess')[1].split('run')[0] + 'run-' + name.split(
'/')[-1].split('sess')[1].split('run')[1]
if 'CSI_' in fn:
fn = fn.split('CSI_')[0] + 'CSI1_' + fn.split('CSI_')[1]
fmri[line].append(('sub-' + fn + '_bold_stnd.nii', i, i + 3,
'sub-' + fn + '_brainmask.nii'))
i += 5
with open(json_file, 'r') as fh:
#cluster = json.load(fh)
cluster = yaml.load(fh)
i = 0
for pair in list(combinations(cluster.keys(), 2)):
cluster_x = pair[0]
cluster_y = pair[1]
# pick 1 image from each cluster
# retrieve relevant fmri data for x,y
for i in range(10):
random.shuffle(cluster[cluster_x])
for i in range(5):
random.shuffle(cluster[cluster_y])
for x in cluster[cluster_x]:
fmri_x = fmri[bytes(
x, encoding='utf-8'
)] # (*_bold_stnd.nii, slice_start, slice_end, *_brainmask.nii)
if len(fmri_x) > 0:
x_n = x.split('.')[0]
break
for y in cluster[cluster_y]:
fmri_y = fmri[bytes(
y, encoding='utf-8'
)] # (*_bold_stnd.nii, slice_start, slice_end, *_brainmask.nii)
if len(fmri_y) > 0:
if fmri_y[0][0] == fmri_x[0][0]:
continue
y_n = y.split('.')[0]
break
#fmri_x = fmri[bytes('rep_homeoffice9.jpg', encoding='utf-8')][0]
#fmri_y = fmri[bytes('rep_homeoffice9.jpg', encoding='utf-8')][1]
#x_n = y_n = 'rep_homeoffice9'
fmri_x = fmri_x[random.randint(0, len(fmri_x) - 1)]
fmri_y = fmri_y[random.randint(0, len(fmri_y) - 1)]
subj_x = fmri_x[0].split('_bold')[0].lstrip('sub-') + '_'
subj_y = fmri_y[0].split('_bold')[0].lstrip('sub-') + '_'
x_nii = image.smooth_img('/Users/hyundonglee/Desktop/Nifti/' +
fmri_x[0],
fwhm=7)
y_nii = image.smooth_img('/Users/hyundonglee/Desktop/Nifti/' +
fmri_y[0],
fwhm=7)
cmu_data = list(
map(lambda n: nii2cmu(n, fmri_x[3]),
[x_nii.slicer[:, :, :, fmri_x[1]:fmri_x[2]]]))
cmu_data.extend(
list(
map(lambda n: nii2cmu(n, fmri_y[3]),
[y_nii.slicer[:, :, :, fmri_y[1]:fmri_y[2]]])))
if args.voxel:
print("saving voxel_locations")
hyp.plot(cmu_data[0]['R'],
'k.',
save_path=out_dir + '/voxel_locations_' + subj_x + x_n +
'.png')
hyp.plot(cmu_data[1]['R'],
'k.',
save_path=out_dir + '/voxel_locations_' + subj_y + y_n +
'.png')
print("save voxel_locations complete")
# Convert between (timepoints by voxels) and (voxels by timepoints) data matrices
htfa_data = list(map(lambda x: {'R': x['R'], 'Z': x['Y'].T}, cmu_data))
nvoxels, ntimepoints = htfa_data[0]['Z'].shape
if args.tfa:
# Use TFA to find network hubs in one subject's data
print("running TFA")
tfa_x = TFA(K=K,
max_num_voxel=int(nvoxels * 0.05),
max_num_tr=int(ntimepoints),
verbose=False)
tfa_x.fit(htfa_data[0]['Z'], htfa_data[0]['R'])
#plot the hubs on a glass brain!
niplot.plot_connectome(np.eye(K),
tfa_x.get_centers(tfa_x.local_posterior_),
node_color='k',
output_file=out_dir + '/network_hubs_' +
subj_x + x_n + '.png')
# Visualizing how the brain images are simplified using TFA
original_image = cmu2nii(htfa_data[0]['Z'][:, 0].T,
htfa_data[0]['R'], x_nii)
niplot.plot_glass_brain(original_image,
plot_abs=False,
output_file=out_dir +
'/simplified_by_TFA_' + subj_x + x_n +
'.png')
connectome = 1 - sd.squareform(sd.pdist(tfa_x.W_), 'correlation')
niplot.plot_connectome(connectome,
tfa_x.get_centers(tfa_x.local_posterior_),
node_color='k',
edge_threshold='75%',
output_file=out_dir + '/connectome' +
subj_x + x_n + '.png')
tfa_y = TFA(K=K,
max_num_voxel=int(nvoxels * 0.05),
max_num_tr=int(ntimepoints),
verbose=False)
tfa_y.fit(htfa_data[1]['Z'], htfa_data[1]['R'])
#plot the hubs on a glass brain!
niplot.plot_connectome(np.eye(K),
tfa_y.get_centers(tfa_y.local_posterior_),
node_color='k',
output_file=out_dir + '/network_hubs_' +
subj_y + y_n + '.png')
# Visualizing how the brain images are simplified using TFA
original_image = cmu2nii(htfa_data[1]['Z'][:, 0].T,
htfa_data[1]['R'], y_nii)
niplot.plot_glass_brain(original_image,
plot_abs=False,
output_file=out_dir +
'/simplified_by_TFA_' + subj_y + y_n +
'.png')
connectome = 1 - sd.squareform(sd.pdist(tfa_y.W_), 'correlation')
niplot.plot_connectome(connectome,
tfa_y.get_centers(tfa_y.local_posterior_),
node_color='k',
edge_threshold='75%',
output_file=out_dir + '/connectome' +
subj_y + y_n + '.png')
print("TFA complete")
print("running HTFA")
htfa = HTFA(K=K,
n_subj=len(htfa_data),
max_global_iter=5,
max_local_iter=2,
voxel_ratio=0.5,
tr_ratio=0.5,
max_voxel=int(nvoxels * 0.05),
max_tr=int(ntimepoints))
htfa.fit(list(map(lambda x: x['Z'], htfa_data)),
list(map(lambda x: x['R'], htfa_data)))
#set the node display properties
colors = np.repeat(np.vstack([[0, 0, 0],
sns.color_palette(
"Spectral", htfa.n_subj)]),
K,
axis=0)
colors = list(
map(lambda x: x[0],
np.array_split(colors, colors.shape[0],
axis=0))) #make colors into a list
sizes = np.repeat(
np.hstack([np.array(50),
np.array(htfa.n_subj * [20])]), K)
#extract the node locations from the fitted htfa model
global_centers = htfa.get_centers(htfa.global_posterior_)
local_centers = list(
map(htfa.get_centers,
np.array_split(htfa.local_posterior_, htfa.n_subj)))
centers = np.vstack([global_centers, np.vstack(local_centers)])
#make the plot
niplot.plot_connectome(np.eye(K * (1 + htfa.n_subj)),
centers,
node_color=colors,
node_size=sizes,
output_file=out_dir + '/htfa_' + subj_x + x_n +
'_' + subj_y + y_n + '.png')
n_timepoints = list(
map(lambda x: x['Z'].shape[1],
htfa_data)) #number of timepoints for each person
inds = np.hstack([0, np.cumsum(np.multiply(K, n_timepoints))])
W = list(
map(
lambda i: htfa.local_weights_[inds[i]:inds[i + 1]].reshape(
[K, n_timepoints[i]]).T, np.arange(htfa.n_subj)))
static_isfc = dynamic_ISFC(W)
niplot.plot_connectome(sd.squareform(static_isfc[0, :]),
global_centers,
node_color='k',
edge_threshold='75%',
output_file=out_dir + '/static_isfc_' + subj_x +
x_n + '_' + subj_y + y_n + '.png')
print("HTFA complete")
i += 1
if n != 0 and i == n:
break
R.append(all_data['R'])
n_voxel, n_tr = data[0].shape
# Run HTFA with downloaded data
from brainiak.factoranalysis.htfa import HTFA
# uncomment below line to get help message on HTFA
#help(HTFA)
K = 5
htfa = HTFA(K=K,
n_subj=n_subj,
max_global_iter=5,
max_local_iter=2,
voxel_ratio=0.5,
tr_ratio=0.5,
max_voxel=n_voxel,
max_tr=n_tr,
verbose=True)
htfa.fit(data, R)
if rank == 0:
print("\n centers of global latent factors are:")
print(htfa.get_centers(htfa.global_posterior_))
print("\n widths of global latent factors are:")
widths = htfa.get_widths(htfa.global_posterior_)
print(widths)
print("\n stds of global latent RBF factors are:")
rbf_std = np.sqrt(widths / (2.0))
print(rbf_std)
tr_solver=tr_solver,
upper_ratio=upper_ratio,
lower_ratio=lower_ratio,
max_tr=max_sample_tr,
max_voxel=max_sample_voxel,
comm=htfa_comm,
verbose=True)
htfa.fit(train_data, R)
for s in range(n_local_subj):
#get posterior for each subject
subj_idx = mapping[str(s)]
start_idx = s * htfa.prior_size
end_idx = (s + 1) * htfa.prior_size
local_posteiror = htfa.local_posterior_[start_idx:end_idx]
local_centers = htfa.get_centers(local_posteiror)
local_widths = htfa.get_widths(local_posteiror)
htfa.n_dim = n_dim
htfa.cov_vec_size = np.sum(np.arange(htfa.n_dim) + 1)
htfa.map_offset = htfa.get_map_offset()
#training happens on all voxels, but part of TRs
unique_R_all, inds_all = htfa.get_unique_R(R[s])
train_F = htfa.get_factors(unique_R_all, inds_all,
local_centers, local_widths)
#calculate train_recon_err
tmp_train_recon_errs[subj_idx, p, idx] = get_train_err(
htfa, train_data[s], train_F)
#calculate weights on test_weight_data, test_recon_err on test_recon_data
tr_solver=tr_solver,
upper_ratio=upper_ratio,
lower_ratio=lower_ratio,
max_tr=max_sample_tr,
max_voxel=max_sample_voxel,
comm=htfa_comm,
verbose=True)
htfa.fit(train_data, R)
for s in range(n_local_subj):
#get posterior for each subject
subj_idx = mapping[str(s)]
start_idx = s * htfa.prior_size
end_idx = (s + 1) * htfa.prior_size
local_posteiror = htfa.local_posterior_[start_idx:end_idx]
local_centers = htfa.get_centers(local_posteiror)
local_widths = htfa.get_widths(local_posteiror)
htfa.n_dim = n_dim
htfa.cov_vec_size = np.sum(np.arange(htfa.n_dim) + 1)
htfa.map_offset = htfa.get_map_offset()
#training happens on all voxels, but part of TRs
unique_R_all, inds_all = htfa.get_unique_R(R[s])
train_F = htfa.get_factors(unique_R_all,
inds_all,
local_centers,
local_widths)
#calculate train_recon_err
tmp_train_recon_errs[subj_idx, p,idx] = get_train_err(htfa,
train_data[s],
