def fit_srm(data_train, data_test, n_components):
"""Fit the shared response model
Parameters
----------
n_components: k
data_train: 3d array (n_subj, n_features, n_examples/tps)
data_test: 3d array (n_subj, n_features, n_examples/tps)
Returns
-------
data_train_sr: 3d array (n_subj, n_components, n_examples/tps)
the transformed training set
data_test_sr: 3d array (n_subj, n_components, n_examples/tps)
the transformed test set
srm: the fitted model
"""
assert len(data_train) == len(data_test)
n_subjects = len(data_train)
# fit SRM on the training set
srm = SRM(features=n_components)
data_train_sr = srm.fit_transform(data_train)
# transform the hidden activity (on the test set) to the shared space
data_test_sr = srm.transform(data_test)
# calculate variance explained
var_exp_train = calc_srm_var_exp(data_train, data_train_sr, srm.w_)
return data_train_sr, data_test_sr, srm, var_exp_train
def SRM_V3(run1, run2, srm_k, n_iter):
# initialize model
print('Building Models')
n_iter = n_iter
srm_k = srm_k
srm_train = SRM(n_iter=n_iter, features=srm_k)
# concatenate run1 and run2 within subject before fitting SRM
runs = []
for i in range(len(run1)):
runs.append(np.concatenate((run1[i], run2[i]), axis=1))
# fit model to training data
print('Training Models')
srm_train.fit(runs)
print('Testing Models')
shared_data_run1 = srm_train.transform(run1)
shared_data_run2 = srm_train.transform(run2)
# average test data across subjects
run1 = sum(shared_data_run1) / len(shared_data_run1)
run2 = sum(shared_data_run2) / len(shared_data_run2)
return run1, run2
def srm(run1, run2):
# initialize model
print('Building Models')
n_iter = 50
srm_k = 30
srm_train_run1 = SRM(n_iter=n_iter, features=srm_k)
srm_train_run2 = SRM(n_iter=n_iter, features=srm_k)
# fit model to training data
print('Training Models')
srm_train_run1.fit(run1)
srm_train_run2.fit(run2)
print('Testing Models')
shared_data_run1 = stats.zscore(np.dstack(srm_train_run2.transform(run1)),
axis=1,
ddof=1)
shared_data_run2 = stats.zscore(np.dstack(srm_train_run1.transform(run2)),
axis=1,
ddof=1)
# average test data across subjects
run1 = np.mean(shared_data_run1, axis=2)
run2 = np.mean(shared_data_run2, axis=2)
return run1, run2
def sfn(l, msk, myrad, bcast_var):
# Arguments:
# l -- a list of 4D arrays, containing data from a single searchlight
# msk -- a 3D binary array, mask of this searchlight
# myrad -- an integer, sl_rad
# bcast_var -- whatever is broadcasted
# extract training and testing data
train_data = []
test_data = []
d1,d2,d3,ntr = l[0].shape
nvx = d1*d2*d3
for s in l:
train_data.append(np.reshape(s[:,:,:,:int(ntr/2)],(nvx,int(ntr/2))))
test_data.append(np.reshape(s[:,:,:,int(ntr/2):],(nvx,ntr-int(ntr/2))))
# train an srm model
srm = SRM(bcast_var[0],bcast_var[1])
srm.fit(train_data)
# transform test data
shared_data = srm.transform(test_data)
for s in range(len(l)):
shared_data[s] = np.nan_to_num(stats.zscore(shared_data[s],axis=1,ddof=1))
# run experiment
accu = time_segment_matching_accuracy(shared_data)
# return: can also return several values. In that case, the final output will be
# a 3D array of tuples
return np.mean(accu)
def corr2_coeff(A,msk,myrad,bcast_var):
#if not np.all(msk):
# return None
print('Assigning Masked Data')
run1 = [A[i][msk==1] for i in np.arange(0, int(len(A)/2))]
run2 = [A[i][msk==1] for i in np.arange(int(len(A)/2), len(A))]
print('Building Model')
srm = SRM(bcast_var[0],bcast_var[1])
print('Training Model')
srm.fit(run1)
print('Testing Model')
shared_data = srm.transform(run2)
shared_data = stats.zscore(np.dstack(shared_data),axis=1,ddof=1)
others = np.mean(shared_data[:,:,np.arange(shared_data.shape[-1]) != loo_idx],axis=2)
loo = shared_data[:,:,loo_idx]
corrAB = np.corrcoef(loo.T,others.T)[16:,:16]
corr_eye = np.identity(16)
same_songs = corrAB[corr_eye == 1]
diff_songs = corrAB[corr_eye == 0]
avg_same_songs = np.mean(same_songs)
avg_diff_songs = np.mean(diff_songs)
same_song_minus_diff_song = avg_same_songs - avg_diff_songs
# Compute difference score for permuted matrices
np.random.seed(0)
diff_perm_holder = np.zeros((1000,1))
for i in range(1000):
corrAB_perm = corrAB[np.random.permutation(16),:]
same_songs_perm = corrAB_perm[corr_eye == 1]
diff_songs_perm = corrAB_perm[corr_eye == 0]
diff_perm_holder[i] = np.mean(same_songs_perm) - np.mean(diff_songs_perm)
z = (same_song_minus_diff_song - np.mean(diff_perm_holder))/np.std(diff_perm_holder)
return z
def HMM(X,K,loo_idx,song_idx,song_bounds):
"""fit hidden markov model
Fit HMM to average data and cross-validate with leftout subject using within song and between song average correlations
Parameters
----------
A: voxel by time ndarray (2D)
B: voxel by time ndarray (2D)
C: voxel by time ndarray (2D)
D: voxel by time ndarray (2D)
K: # of events for HMM (scalar)
Returns
-------
z: z-score after performing permuted cross-validation analysis
"""
w = 6
srm_k = 45
nPerm = 1000
within_across = np.zeros(nPerm+1)
run1 = [X[i] for i in np.arange(0, int(len(X)/2))]
run2 = [X[i] for i in np.arange(int(len(X)/2), len(X))]
print('Building Model')
srm = SRM(n_iter=10, features=srm_k)
print('Training Model')
srm.fit(run1)
print('Testing Model')
shared_data = srm.transform(run2)
shared_data = stats.zscore(np.dstack(shared_data),axis=1,ddof=1)
others = np.mean(shared_data[:,:,np.arange(shared_data.shape[-1]) != loo_idx],axis=2)
loo = shared_data[:,song_bounds[song_idx]:song_bounds[song_idx + 1],loo_idx]
nTR = loo.shape[1]
# Fit to all but one subject
ev = brainiak.eventseg.event.EventSegment(K)
ev.fit(others[:,song_bounds[song_idx]:song_bounds[song_idx + 1]].T)
events = np.argmax(ev.segments_[0],axis=1)
# Compute correlations separated by w in time
corrs = np.zeros(nTR-w)
for t in range(nTR-w):
corrs[t] = pearsonr(loo[:,t],loo[:,t+w])[0]
# Compute within vs across boundary correlations, for real and permuted bounds
for p in range(nPerm+1):
within = corrs[events[:-w] == events[w:]].mean()
across = corrs[events[:-w] != events[w:]].mean()
within_across[p] = within - across
np.random.seed(p)
events = np.zeros(nTR, dtype=np.int)
events[np.random.choice(nTR,K-1,replace=False)] = 1
events = np.cumsum(events)
return within_across
def HMM(X, human_bounds, song_idx, song_bounds, srm_k, hrf):
"""fit hidden markov model
Fit HMM to average data and cross-validate with leftout subject using within song and between song average correlations
Parameters
----------
A: voxel by time ndarray (2D)
B: voxel by time ndarray (2D)
C: voxel by time ndarray (2D)
D: voxel by time ndarray (2D)
K: # of events for HMM (scalar)
Returns
-------
z: z-score after performing permuted cross-validation analysis
"""
w = 3
nPerm = 1000
run1 = [X[i] for i in np.arange(0, int(len(X) / 2))]
run2 = [X[i] for i in np.arange(int(len(X) / 2), len(X))]
print('Building Model')
srm = SRM(n_iter=10, features=srm_k)
print('Training Model')
srm.fit(run2)
print('Testing Model')
shared_data = srm.transform(run1)
shared_data = stats.zscore(np.dstack(shared_data), axis=1, ddof=1)
data = np.mean(shared_data[:, song_bounds[song_idx]:song_bounds[song_idx +
1]],
axis=2)
nTR = data.shape[1]
# Fit to all but one subject
K = len(human_bounds) + 1
ev = brainiak.eventseg.event.EventSegment(K)
ev.fit(data.T)
bounds = np.where(np.diff(np.argmax(ev.segments_[0], axis=1)))[0] + 1
match = np.zeros(nPerm + 1)
events = np.argmax(ev.segments_[0], axis=1)
_, event_lengths = np.unique(events, return_counts=True)
perm_bounds = bounds.copy()
for p in range(nPerm + 1):
for hb in human_bounds:
if np.any(np.abs(perm_bounds - hb) <= w):
match[p] += 1
match[p] /= len(human_bounds)
np.random.seed(p)
perm_bounds = np.cumsum(np.random.permutation(event_lengths))[:-1]
return match
def SRM_V1(train, test, srm_k, n_iter):
# initialize model
print('Building Models')
srm_train_data = SRM(n_iter=n_iter, features=srm_k)
# fit model to training data
print('Training Models')
srm_train_data.fit(train)
print('Testing Models')
shared_data = srm_train_data.transform(test)
return shared_data
def srm(train_data, test_data, n_features):
# Z-score the data
n = len(train_data)
for subject in range(n):
train_data[subject] = stats.zscore(train_data[subject], axis=1, ddof=1)
test_data[subject] = stats.zscore(test_data[subject], axis=1, ddof=1)
model = SRM(n_iter=10, features=n_features)
model.fit(train_data)
projected_data = model.transform(test_data)
for subject in range(n):
projected_data[subject] = stats.zscore(projected_data[subject], axis=1, ddof=1)
return projected_data
def train_srm(training_data,
use_robust_srm=True,
n_comps=50,
n_iters=20,
printruntime=True):
"""
Fit srm on training data
"""
# TODO: delete after test
from mpi4py import MPI
mpicomm = MPI.COMM_WORLD
if use_robust_srm:
srm = RSRM(n_iter=n_iters, features=n_comps, comm=mpicomm)
else:
srm = SRM(n_iter=n_iters, features=n_comps, comm=mpicomm)
# fit
if printruntime:
start = time.time()
srm.fit(training_data)
if printruntime:
elapsed = start - time.time()
print('fitting srm took: ', elapsed)
return srm
def srm(train_data, test_data, n_features):
# Z-score the data
n = len(train_data)
for subject in range(n):
train_data[subject] = stats.zscore(train_data[subject], axis=1, ddof=1)
test_data = stats.zscore(test_data, axis=1, ddof=1)
model = SRM(n_iter=10, features=n_features)
model.fit(train_data)
# now we do the mstep of srm to get its W for the new X
a_new = test_data.dot(model.s_.T)
perturbation = np.zeros(a_new.shape)
np.fill_diagonal(perturbation, 0.001)
u, s, v = np.linalg.svd(a_new + perturbation, full_matrices=False)
w_new = u.dot(v)
return rmse(w_new.dot(model.s_), test_data), relative_mse(w_new.dot(model.s_), test_data)
def isc_srm(X,set_srm):
"""perform isc on srm searchlights
Parameters
----------
X: list of searchlights where searchlights are voxels by time
Returns
-------
r: correlations for each searchlights timecourse correlated with all other timecourses
"""
run1 = [X[i] for i in np.arange(0, int(len(X)/2))]
run2 = [X[i] for i in np.arange(int(len(X)/2), len(X))]
data = np.zeros((run1[0].shape[0],run1[0].shape[1]*2,len(run1)))
if set_srm == 0:
for i in range(data.shape[2]):
data[:,:,i] = np.hstack((run1[i],run2[i]))
# run isc
isc_output = brainiak.isfc.isfc(data)
elif set_srm == 1:
# train on run 1 and test on run 2
print('Building Model')
srm = SRM(n_iter=10, features=5)
print('Training Model')
srm.fit(run1)
print('Testing Model')
shared_data = srm.transform(run2)
shared_data = stats.zscore(np.dstack(shared_data),axis=1,ddof=1)
# run isc
isc_output = brainiak.isfc.isfc(shared_data)
print(isc_output)
# average isc results
mean_isc = np.mean(isc_output)
return mean_isc,isc_output
def SRM_V2(run1, run2, srm_k, n_iter):
# initialize model
print('Building Models')
n_iter = n_iter
srm_k = srm_k
srm_train_run1 = SRM(n_iter=n_iter, features=srm_k)
srm_train_run2 = SRM(n_iter=n_iter, features=srm_k)
# fit model to training data
print('Training Models')
srm_train_run1.fit(run1)
srm_train_run2.fit(run2)
print('Testing Models')
shared_data_run1 = srm_train_run2.transform(run1)
shared_data_run2 = srm_train_run1.transform(run2)
# average test data across subjects
run1 = sum(shared_data_run1) / len(shared_data_run1)
run2 = sum(shared_data_run2) / len(shared_data_run2)
return run1, run2
def sfn(l, msk, myrad, bcast_var):
# extract training and testing data
train_data = []
test_data = []
d1, d2, d3, ntr = l[0].shape
nvx = d1 * d2 * d3
for s in l:
train_data.append(
np.reshape(s[:, :, :, :int(ntr / 2)], (nvx, int(ntr / 2))))
test_data.append(
np.reshape(s[:, :, :, int(ntr / 2):], (nvx, ntr - int(ntr / 2))))
# train an srm model
srm = SRM(bcast_var[0], bcast_var[1])
srm.fit(train_data)
# transform test data
shared_data = srm.transform(test_data)
for s in range(len(l)):
shared_data[s] = np.nan_to_num(
stats.zscore(shared_data[s], axis=1, ddof=1))
# experiment
accu = timesegmentmatching_accuracy(shared_data, 6)
return np.mean(accu), stats.sem(
accu) # multiple outputs will be saved as tuples
def srm(train_data, n_features):
# Z-score the data
n = len(train_data)
for subject in range(n):
train_data[subject] = stats.zscore(train_data[subject], axis=1, ddof=1)
all_subj = np.arange(n)
np.random.shuffle(all_subj)
model1 = SRM(n_iter=10, features=n_features)
model2 = SRM(n_iter=10, features=n_features)
dset1 = train_data[all_subj[:n // 2], :, :]
dset2 = train_data[all_subj[n // 2:], :, :]
model1.fit(dset1)
model2.fit(dset2)
return norm_from_ortho(model1.s_, model2.s_)
def train_srm(training_data,
use_robust_srm=True,
n_comps=10,
n_iters=4,
printruntime=True):
"""
Fit srm on training data
"""
if use_robust_srm:
srm = RSRM(n_iter=n_iters, features=n_comps)
else:
srm = SRM(n_iter=n_iters, features=n_comps)
# fit
if printruntime:
start = time.time()
srm.fit(training_data)
if printruntime:
elapsed = time.time() - start
print('fitting srm took: ', elapsed)
return srm
test = np.nan_to_num(
stats.zscore(np.load(datadir +
'angular_gyrus_raw_hmm_searchlight_run2_n25.npy'),
axis=1,
ddof=1))
# Convert data into lists where each element is voxels by samples
train_list = []
test_list = []
for i in range(0, train.shape[2]):
train_list.append(train[:, :, i])
test_list.append(test[:, :, i])
# Initialize model
print('Building Model')
srm = SRM(n_iter=10, features=25)
# Fit model to training data (run 1)
print('Training Model')
srm.fit(train_list)
# Test model on testing data to produce shared response
print('Testing Model')
shared_data = srm.transform(test_list)
avg_response = sum(shared_data) / len(shared_data)
human_bounds = np.cumsum(np.load(datadir + 'songs2Dur.npy'))[:-1]
human_bounds2 = np.array([
0, 90, 270, 449, 538, 672, 851, 1031, 1255, 1480, 1614, 1704, 1839, 2063,
2288, 2377, 2511
])
train_list_A1 = []
test_list_A1 = []
train_list_vmPFC = []
test_list_vmPFC = []
for i in range(0, train_A1.shape[2]):
train_list_A1.append(train_A1[:, :, i])
test_list_A1.append(test_A1[:, :, i])
for i in range(0, train_vmPFC.shape[2]):
train_list_vmPFC.append(train_vmPFC[:, :, i])
test_list_vmPFC.append(test_vmPFC[:, :, i])
# Initialize models
print('Building Model')
srm_A1 = SRM(n_iter=10, features=50)
srm_vmPFC = SRM(n_iter=10, features=10)
# Fit model to training data (run 1)
print('Training Model')
srm_A1.fit(train_list_A1)
srm_vmPFC.fit(train_list_vmPFC)
# Test model on testing data to produce shared response
print('Testing Model')
shared_data_A1 = srm_A1.transform(test_list_A1)
shared_data_vmPFC = srm_vmPFC.transform(test_list_vmPFC)
avg_response_A1 = sum(shared_data_A1) / len(shared_data_A1)
avg_response_vmPFC = sum(shared_data_vmPFC) / len(shared_data_vmPFC)
# extract max target activation as the metric for recall
tma[cond] = np.array([
np.max(ma_dlist[cond][i_s]['targ'], axis=-1) for i_s in range(n_subjs)
]).transpose((0, 2, 1))
print(f'np.shape(tma[{cond}]) = {np.shape(tma[cond])}')
# organize brain activity
CMs_darray, DAs_darray = {}, {}
for cond in all_conds:
CMs_darray[cond] = np.array(CMs_dlist[cond]).transpose((0, 3, 2, 1))
DAs_darray[cond] = np.array(DAs_dlist[cond]).transpose((0, 3, 2, 1))
print(f'np.shape(CMs_darray[{cond}]) = {np.shape(CMs_darray[cond])}')
print(f'np.shape(DAs_darray[{cond}]) = {np.shape(DAs_darray[cond])}')
'''isc'''
dim_srm = 16
srm = SRM(features=dim_srm)
test_prop = .5
n_examples_tr = int(n_examples * (1 - test_prop))
n_examples_te = n_examples - n_examples_tr
data = DAs_darray
_, nH, _, _ = np.shape(data[cond])
# split data
data_tr = {cond: [] for cond in all_conds}
data_te = {cond: [] for cond in all_conds}
for cond in all_conds:
data_tr_cond = data[cond][:, :, :, :n_examples_tr]
data_te_cond = data[cond][:, :, :, n_examples_tr:]
for story in train_stories:
roi_data = roi_datasets[story]['data']
if story in test_stories:
roi_trs = roi_data.shape[0]
train_roi[story] = np.nan_to_num(
zscore(roi_data[:roi_trs // 2, ...], axis=0))
test_roi[story] = np.nan_to_num(
zscore(roi_data[roi_trs // 2:, ...], axis=0))
else:
train_roi[story] = np.nan_to_num(zscore(roi_data, axis=0))
# Time-series SRM on ROI data
if timeseries_srm:
test_shared = {}
for story in stories:
srm = SRM(n_iter=n_iter, features=n_features)
# Change subjects to list for SRM
train_list = [train.T for train in np.moveaxis(train_roi[story], 2, 0)]
test_list = [test.T for test in np.moveaxis(test_roi[story], 2, 0)]
# Train SRM and apply
srm.fit(train_list)
test_transformed = srm.transform(test_list)
test_shared[story] = np.dstack([test.T for test in test_transformed])
# Load in the whole-brain surface data
train_target, test_target = {}, {}
for story in train_stories:
train_mean = np.mean(train, axis=axis)
train_std = np.std(train, axis=axis)
train_z = (train - train_mean) / train_std
test_z = (test - train_mean) / train_std
return train_z, test_z
# Initialize 10-fold cross-validation splitter
cv = KFold(n_splits=10)
# Initialize SRM with fixed number of features
adaptive_k = False
n_features = 50
srm = SRM(n_iter=10, features=n_features)
# Loop through all pairs of embedding types and get cross-validated SRM
srm_pairs = {}
for pair in emb_pairs:
# Cross-validation split indices
srm_tests = {pair[0]: [], pair[1]: []}
for cv_i, (train_ids, test_ids) in enumerate(cv.split(np.arange(n_embs))):
# Split and z-score embeddings
emb0_train, emb0_test = zscore_cv(embeddings[pair[0]][train_ids],
embeddings[pair[0]][test_ids])
emb1_train, emb1_test = zscore_cv(embeddings[pair[1]][train_ids],
embeddings[pair[1]][test_ids])
include_list.append(ints)
include_list = np.hstack(include_list)
# grab first 4 songs from run 1 to train which is equal to first 628 seconds
for i in subjs:
run1 = datadir + 'subjects/' + i + '/analysis/run1.feat/trans_filtered_func_data.nii'
run2 = datadir + 'subjects/' + i + '/analysis/run2.feat/trans_filtered_func_data.nii'
run1 = nib.load(run1).get_data()[:, :, :, 0:2511]
run2 = nib.load(run2).get_data()[:, :, :, 0:2511]
run2 = run2[mask == 1, :]
train.append(run1[mask == 1, 0:628])
test.append(run2[:, omit_mask == 1])
print('Building Model')
srm = SRM(n_iter=10, features=50)
print('Training Model')
srm.fit(train)
print('Testing Model')
shared_data = srm.transform(test)
shared_data = stats.zscore(np.dstack(shared_data), axis=1, ddof=1)
corrAB_holder = []
correct_songs = np.zeros((25, 12))
for i in range(len(subjs)):
loo = shared_data[:, :, i]
others = np.mean(shared_data[:, :,
np.arange(shared_data.shape[-1]) != i],
axis=2)
corrAB = np.corrcoef(loo.T, others.T)[12:, :12]
def srm_fit(target_fcs, stories=None, subjects=None,
hemisphere=None, k=360, n_iter=10,
half=1, save_prefix=None):
# By default grab all stories
stories = check_keys(target_fcs, keys=stories)
# Recompile FCs accounting for repeat subjects
subject_fcs = {}
for story in stories:
# By default just grab all subjects
subject_list = check_keys(target_fcs[story], keys=subjects,
subkey=story)
for subject in subject_list:
# For simplicity we just assume same hemis across stories/subjects
hemis = check_keys(target_fcs[story][subject],
keys=hemisphere)
for hemi in hemis:
# If subject is not already there, make new dict for them
if subject not in subject_fcs:
subject_fcs[subject] = {}
# If hemispheres aren't in there, add them
if hemi not in subject_fcs[subject]:
subject_fcs[subject][hemi] = []
# Finally, make list of connectivity matrices per subject
subject_fcs[subject][hemi].append(
target_fcs[story][subject][hemi])
# Stack FCs in connectivity space (for all subjects across stories!)
all_subjects = list(subject_fcs.keys())
for subject in all_subjects:
for hemi in hemis:
# If more than one connectivity per subject, take average
if len(subject_fcs[subject][hemi]) > 1:
subject_fcs[subject][hemi] = np.mean(subject_fcs[subject][hemi],
axis=0)
else:
subject_fcs[subject][hemi] = subject_fcs[subject][hemi][0]
# Convert FCs to list for SRM (grab the shared space too)
transforms, shared_space, srms = {}, {}, {}
for hemi in hemis:
# Declare SRM for this hemi
srm = SRM(n_iter=n_iter, features=k)
subject_ids, subject_stack = [], []
for subject in all_subjects:
subject_ids.append(subject)
subject_stack.append(subject_fcs[subject][hemi])
if subject not in transforms:
transforms[subject] = {}
# Train SRM and apply
start = time()
srm.fit(subject_stack)
print(f"Finished fitting SRM after {time() - start:.1f} seconds")
for subject_id, transform in zip(subject_ids, srm.w_):
transforms[subject_id][hemi] = transform
shared_space[hemi] = srm.s_
srms[hemi] = srm
if save_prefix:
np.save(f'data/half-{half}_{save_prefix}_w.npy', transforms)
np.save(f'data/half-{half}_{save_prefix}_s.npy', shared_space)
return transforms, srms
test_list_roi_1 = []
train_list_roi_2 = []
test_list_roi_2 = []
for i in range(0,train_roi_1.shape[2]):
train_list_roi_1.append(train_roi_1[:,:,i])
test_list_roi_1.append(test_roi_1[:,:,i])
for i in range(0,train_roi_2.shape[2]):
train_list_roi_2.append(train_roi_2[:,:,i])
test_list_roi_2.append(test_roi_2[:,:,i])
# Initialize models
print('Building Model')
srm_roi_1 = SRM(n_iter=50, features=5)
srm_roi_2 = SRM(n_iter=50, features=5)
# Fit model to training data (run 1)
print('Training Model')
srm_roi_1.fit(train_list_roi_1)
srm_roi_2.fit(train_list_roi_2)
# Test model on testing data to produce shared response
print('Testing Model')
shared_data_roi_1 = srm_roi_1.transform(test_list_roi_1)
shared_data_roi_2 = srm_roi_2.transform(test_list_roi_2)
avg_response_roi_1 = sum(shared_data_roi_1)/len(shared_data_roi_1)
avg_response_roi_2 = sum(shared_data_roi_2)/len(shared_data_roi_2)
def parcel_srm(data,
atlas,
k=3,
parcel_labels=None,
stories=None,
subjects=None):
# By default grab all stories
stories = check_keys(data, keys=stories)
# Firsts compute mean time-series for all target parcels
targets = parcel_means(data,
atlas,
parcel_labels=parcel_labels,
stories=stories,
subjects=subjects)
# Compute ISFCs with targets for all vertices
target_fcs = target_isfc(data, targets, stories=stories, subjects=subjects)
parcels = {}
for story in stories:
parcels[story] = {}
# By default just grab all subjects
subject_list = check_keys(data[story], keys=subjects, subkey=story)
# Loop through both hemispheres
hemi_stack = []
for hemi in ['lh', 'rh']:
# Loop through parcels
parcel_tss = []
for parcel_label in parcel_labels[hemi]:
# Resort parcel FCs into list of subject parcels
fc_stack = []
ts_stack = []
for subject in subject_list:
# Grab the connectivities for this parcel
parcel_fcs = target_fcs[story][subject][hemi][
atlas[hemi] == parcel_label, :]
fc_stack.append(parcel_fcs)
ts_stack.append(data[story][subject][hemi]
[:, atlas[hemi] == parcel_label])
# Set up fresh SRM
srm = SRM(features=k)
# Train SRM on parcel connectivities
srm.fit(np.nan_to_num(fc_stack))
# Apply transformations to time series
transformed_stack = [
ts.dot(w) for ts, w in zip(ts_stack, srm.w_)
]
transformed_stack = np.dstack(transformed_stack)
parcel_tss.append(transformed_stack)
print(f"Finished SRM for {hemi} parcel "
f"{parcel_label} in '{story}'")
# Stack parcel means
parcel_tss = np.hstack(parcel_tss)
hemi_stack.append(parcel_tss)
# Stack hemispheres
hemi_stack = np.hstack(hemi_stack)
assert hemi_stack.shape[1] == (len(parcel_labels['lh']) +
len(parcel_labels['rh'])) * k
assert hemi_stack.shape[2] == len(subject_list)
# Unstack subjects
hemi_stack = np.dsplit(hemi_stack, hemi_stack.shape[2])
for subject, ts in zip(subject_list, hemi_stack):
parcels[story][subject] = np.squeeze(ts)
print(f"Finished applying cSRM to parcels for '{story}'")
return parcels
test_list.append(test[:,:,i])
songs = ['St_Pauls_Suite', 'I_Love_Music', 'Moonlight_Sonata', 'Change_of_the_Guard','Waltz_of_Flowers','The_Bird', 'Island', 'Allegro_Moderato', 'Finlandia', 'Early_Summer', 'Capriccio_Espagnole', 'Symphony_Fantastique', 'Boogie_Stop_Shuffle', 'My_Favorite_Things', 'Blue_Monk','All_Blues']
durs = np.array([90,180,180,90,135,180,180,225,225,135,90,135,225,225,90,135])
song_bounds = np.array([0,90,270,449,538,672,851,1031,1255,1480,1614,1704,1839,2063,2288,2377,2511])
features = np.array([5,10,15,20,25,30,35,40,45,50])
n_iter = 50
hrf = 5
nPerm = 1000
w = 3
z_scores = np.zeros((len(features),len(songs)))
t_scores = np.zeros((len(features)))
for k in range(len(features)):
srm = SRM(n_iter=n_iter, features=features[k])
# Fit model to training data (run 1)
srm.fit(train_list)
# Test model on testing data to produce shared response
shared_data = srm.transform(test_list)
avg_response = sum(shared_data)/len(shared_data)
for s in range(len(songs)):
# Get start and end of chosen song
start = song_bounds[s] + hrf
end = song_bounds[s+1] + hrf
nTR = shared_data[0][:,start:end].shape[1]
np.random.seed(bootNum)
for b in range(nboot):
resamp_subjs = np.random.choice(nSubj, size=nSubj, replace=True)
run1_list = []
run2_list = []
for r in range(len(resamp_subjs)):
run1_list.append(run1_list_orig[resamp_subjs[r]])
run2_list.append(run2_list_orig[resamp_subjs[r]])
n_iter = 50
features = 10
# Initialize model
print('Building Model')
srm_train_run1 = SRM(n_iter=n_iter, features=features)
srm_train_run2 = SRM(n_iter=n_iter, features=features)
# Fit model to training data
print('Training Model')
srm_train_run1.fit(run1_list)
srm_train_run2.fit(run2_list)
# Test model on testing data to produce shared response
print('Testing Model')
shared_data_train_run1 = srm_train_run1.transform(run2_list)
shared_data_train_run2 = srm_train_run2.transform(run1_list)
avg_response_train_run1 = sum(shared_data_train_run1) / len(
shared_data_train_run1)
avg_response_train_run2 = sum(shared_data_train_run2) / len(
subjects = movie_data.shape[0]
nVoxel, nTR = movie_data[0].shape
train_data = []
test_data = []
for s in range(subjects):
train_data.append(movie_data[s, :, :nTR // 2])
test_data.append(movie_data[s, :, nTR // 2:])
for subject in range(subjects):
train_data[subject] = stats.zscore(train_data[subject], axis=1, ddof=1)
for subject in range(subjects):
test_data[subject] = stats.zscore(test_data[subject], axis=1, ddof=1)
srm = SRM(n_iter=10, features=50)
srm.fit(train_data)
data_shared = srm.transform(test_data)
timesegmentmatching_accuracy_evaluation_loo_cv(data_shared,
win_size=6,
method="SRM")
dpsrm = DPSRM(n_features=50)
dpsrm.fit(np.array(train_data))
data_shared_dpsrm = dpsrm.transform(test_data)
timesegmentmatching_accuracy_evaluation_loo_cv(data_shared_dpsrm,
win_size=6,
method="DPSRM")
def HMM(X, human_bounds, song_idx, song_bounds, hrf, srm_k):
"""fit hidden markov model
Fit HMM to average data and cross-validate with leftout subjects using within song and between song average correlations
Parameters
----------
A: list of 50 (contains 2 runs per subject) 2D (voxels x full time course) arrays
B: # of events for HMM (scalar)
song_idx: song index (scalar)
C: voxel by time ndarray (2D)
D: array of song boundaries (1D)
Returns
-------
wVa score: final score after performing cross-validation of leftout subjects
"""
w = 6
nPerm = 1000
within_across = np.zeros(nPerm + 1)
run1 = [X[i] for i in np.arange(0, int(len(X) / 2))]
run2 = [X[i] for i in np.arange(int(len(X) / 2), len(X))]
print('Building Model')
srm = SRM(n_iter=10, features=srm_k)
print('Training Model')
srm.fit(run1)
print('Testing Model')
shared_data = srm.transform(run2)
shared_data = stats.zscore(np.dstack(shared_data), axis=1, ddof=1)
others = np.mean(
shared_data[:, song_bounds[song_idx]:song_bounds[song_idx + 1], :13],
axis=2)
loo = np.mean(shared_data[:,
song_bounds[song_idx]:song_bounds[song_idx + 1],
13:],
axis=2)
nTR = loo.shape[1]
# Fit to all but one subject
K = len(human_bounds) + 1
ev = brainiak.eventseg.event.EventSegment(K)
ev.fit(others.T)
events = np.argmax(ev.segments_[0], axis=1)
# Compute correlations separated by w in time
corrs = np.zeros(nTR - w)
for t in range(nTR - w):
corrs[t] = pearsonr(loo[:, t], loo[:, t + w])[0]
# Compute within vs across boundary correlations, for real and permuted bounds
for p in range(nPerm + 1):
within = corrs[events[:-w] == events[w:]].mean()
across = corrs[events[:-w] != events[w:]].mean()
within_across[p] = within - across
np.random.seed(p)
events = np.zeros(nTR, dtype=np.int)
events[np.random.choice(nTR, K - 1, replace=False)] = 1
events = np.cumsum(events)
print((within_across[0] - np.mean(within_across[1:])) /
np.std(within_across[1:]))
return within_across
def srm_fit(target_fcs, stories=None, subjects=None,
hemisphere=None, k=360, n_iter=10):
# By default grab all stories
stories = check_keys(target_fcs, keys=stories)
# Recompile FCs accounting for repeat subjects
subject_fcs = {}
for story in stories:
# By default just grab all subjects
subject_list = check_keys(target_fcs[story], keys=subjects,
subkey=story)
for subject in subject_list:
# For simplicity we just assume same hemis across stories/subjects
hemis = check_keys(target_fcs[story][subject],
keys=hemisphere)
for hemi in hemis:
if subject not in subject_fcs:
subject_fcs[subject] = {}
if hemi not in subject_fcs[subject]:
subject_fcs[subject][hemi] = []
subject_fcs[subject][hemi].append(
target_fcs[story][subject][hemi])
# Stack FCs across stories in connectivity space
for subject in subject_list:
for hemi in hemis:
if len(subject_fcs[subject][hemi]) > 1:
subject_fcs[subject][hemi] = np.mean(subject_fcs[subject][hemi],
axis=0)
else:
subject_fcs[subject][hemi] = subject_fcs[subject][hemi][0]
# Convert FCs to list for SRM
transforms = {}
for hemi in hemis:
# Declare SRM for this hemi
srm = SRM(n_iter=n_iter, features=k)
subject_ids, subject_stack = [], []
for subject in subject_list:
subject_ids.append(subject)
subject_stack.append(subject_fcs[subject][hemi])
if subject not in transforms:
transforms[subject] = {}
# Train SRM and apply
start = time()
srm.fit(subject_stack)
print(f"Finished fitting SRM after {time() - start:.1f} seconds")
for subject_id, transform in zip(subject_ids, srm.w_):
transforms[subject_id][hemi] = transform
return transforms
