# Fisher z to r transform on average , now this is back to correlation array
SUM = fisher_z2r(SUM)
# transform correlation array into similarity array
SUM += 1.0
SUM /= 2.0
# get full similartiy matrix of correlations
N = hcp_util.N_original(SUM)
SUM.resize([N,N])
hcp_util.upper_to_down(SUM)
print "SUM.shape", SUM.shape
print "do embed for correlation matrix:", SUM.shape
# Satra's embedding algorithm
embedding, result = embed.compute_diffusion_map(SUM, alpha=0, n_components=20,
diffusion_time=0, skip_checks=True, overwrite=True)
# output prefix
out_prfx=cliarg_out_prfx
# output precision
out_prec="%g"
np.savetxt(out_prfx + "embedding2.csv", embedding, fmt=out_prec, delimiter='\t', newline='\n')
np.savetxt(out_prfx + "lambdas2.csv", result['lambdas'], fmt=out_prec, delimiter='\t', newline='\n')
np.savetxt(out_prfx + "vectors2.csv", result['vectors'], fmt=out_prec, delimiter='\t', newline='\n')
print result['lambdas']
print "loading data - loop done"
# total number of HCP subjects run with corr_top10.py
N = args.cntsubjects
# get mean correlation
SUM = ne.evaluate('SUM / N')
# get similarity matrix (Kernel matrix / Gram matrix)
SUM +=1.0
SUM /=2.0
# set NaN entries to 0
SUM[np.where(np.isnan(SUM) == True)] = 0
# ignore zero entries?
ind = np.where(np.sum(SUM,axis=1) != 1)
print "do embed for corr matrix "
embedding, result = embed.compute_diffusion_map(SUM[ind].T[ind].T,
n_components=10)
print result['lambdas']
print "embedding done!"
print "writing out embedding results..."
h = h5py.File(args.outprfx , 'w')
h.create_dataset('embedding', data=embedding)
h.create_dataset('lambdas', data=result['lambdas'])
h.create_dataset('vectors', data=result['vectors'])
h.close()
corr[i, corr[i, :] < perc[i]] = 0
# Check for minimum & maximum value
print "Minimum value is %f" % corr.min()
print "Maximum value is %f" % corr.max()
## Count negative values per row
#neg_values = np.array([sum(corr[i,:] < 0) for i in range(N_orig)])
#print "Negative values occur in %d rows" % sum(neg_values > 0)
#if sum(neg_values) > 0:
# corr[corr < 0] = 0
##### Step #7: get affinity matrix via cosine similarity ##############
print "calculating affinity matrix..."
aff = 1 - pairwise_distances(corr, metric='cosine')
os.chdir('/data/pt_mar006/subjects_group/')
h = h5py.File('affinity_matrix.h5', 'w')
h.create_dataset("data", data=aff)
h.close()
##### Step #8: get embedding components on affinity matrix ############
print "calculating connectivity components..."
comp = 9
emb, res = embed.compute_diffusion_map(aff, alpha=0.5, n_components=comp)
np.save('embedding_dense_emb.npy', emb)
np.save('embedding_dense_res.npy', res)
# Check for minimum & maximum value
print "Minimum value is %f" % corr.min()
print "Maximum value is %f" % corr.max()
## Count negative values per row
#neg_values = np.array([sum(corr[i,:] < 0) for i in range(N_orig)])
#print "Negative values occur in %d rows" % sum(neg_values > 0)
#if sum(neg_values) > 0:
# corr[corr < 0] = 0
##### Step #7: get affinity matrix via cosine similarity ##############
print "calculating affinity matrix..."
aff = 1 - pairwise_distances(corr, metric = 'cosine')
os.chdir('/data/pt_mar006/subjects_group/')
h = h5py.File('affinity_matrix.h5','w')
h.create_dataset("data", data=aff)
h.close()
##### Step #8: get embedding components on affinity matrix ############
print "calculating connectivity components..."
comp = 9
emb, res = embed.compute_diffusion_map(aff, alpha = 0.5,
n_components = comp)
np.save('embedding_dense_emb.npy', emb)
np.save('embedding_dense_res.npy', res)
def diffusion_embedding(population, workspace, popname):
mask = os.path.join(workspace, 'PA001', 'REGISTRATION/REST_GM_MNI3mm.nii.gz')
corrmat_shape = np.count_nonzero(nb.load(mask).get_data())
corrmat_sum = np.zeros((corrmat_shape, corrmat_shape))
embed_dir = mkdir_path(os.path.join(workspace, 'EMBEDDING'))
if not os.path.isfile(os.path.join(embed_dir, "corrmat_mean_%s.npy"%popname)):
for subject in population:
print 'Creating correlatiom matrix for Subject %s' %subject
#I/O
subject_dir = os.path.join(workspace, subject)
func = os.path.join(subject_dir, 'DENOISE/residuals_compcor/residual_bp_z_fwhm6.nii.gz')
mask = os.path.join(workspace, 'PA001', 'REGISTRATION/REST_GM_MNI3mm.nii.gz')
# Get timeseries
masker = input_data.NiftiMasker(mask)
time_series = masker.fit_transform(func)
print '....', time_series.shape
correlation_measure = connectome.ConnectivityMeasure(kind='correlation')
###########add R to Z
correlation_matrix = correlation_measure.fit_transform([time_series])[0]
print '....', correlation_matrix.shape
corrmat_sum += correlation_matrix
print corrmat_sum.shape
# modify this to save mat files... make sure they are deleted after embedding. will help looping
# embed_dir = mkdir_path(os.path.join(subject_dir, 'EMBED'))
#save matrix as h5py file
# h5 = h5py.File('%s/corrmat_%s.h5'%(embed_dir, subject), 'w')
# h5.create_dataset('data', data=correlation_matrix)
# all_mats.append(correlation_matrix)
print 'Calculating average correlation matrix'
corrmat_mean = corrmat_sum / len(population)
print '....',corrmat_mean.shape
os.chdir(embed_dir)
np.save(os.path.join(embed_dir, "corrmat_mean_%s.npy"%popname), corrmat_mean)
# Nonlinear decomposition with Emebdding
print 'Eembedding'
# add Z to R
corrmat_mean = np.load(os.path.join(embed_dir, "corrmat_mean_%s.npy" % popname))
###### Threshold each row of corr matrix at 90th percentile ##
print "thresholding each row at its 90th percentile..."
percentile90 = np.array([np.percentile(x, 90) for x in corrmat_mean])
for i in range(corrmat_mean.shape[0]):
# print "Row %d" % i
corrmat_mean[i, corrmat_mean[i, :] < percentile90[i]] = 0
##### get affinity matrix via cosine similarity ##############
print "calculating affinity matrix..."
aff = 1 - pairwise_distances(corrmat_mean, metric = 'cosine')
# Check for minimum & maximum value
print "....Minimum value of aff is %f" % aff.min()
print "....Maximum value of aff is %f" % aff.max()
print "saving affinity matrix..."
h = h5py.File(os.path.join(embed_dir, 'affinity_matrix.h5'), 'w')
h.create_dataset("data", data=aff)
h.close()
##### Get embedding components on affinity matrix ############
print "calculating connectivity components..."
emb, res = embed.compute_diffusion_map(aff, alpha=0.5, n_components=10)
np.save(os.path.join(embed_dir,'embedding_dense_emb.npy'), emb)
np.save(os.path.join(embed_dir,'embedding_dense_res.npy'), res)
#### Step #9: projecting components back to MNI space as nifti #######
print "saving components as nifti..."
emb = np.load(os.path.join(embed_dir,'embedding_dense_emb.npy'))
##### get indices of voxels, which are equal to 1 in mask
mask_array = nb.load(mask).get_data()
voxel_x = np.where(mask_array==1)[0]
voxel_y = np.where(mask_array==1)[1]
voxel_z = np.where(mask_array==1)[2]
print "%s voxels are in GM..." % len(voxel_x)
mni_affine = nb.load(mni_brain_3mm).get_affine()
data_temp = np.zeros(nb.load(mni_brain_3mm).get_data().shape)
for j in range(0, 10):
print np.shape(emb[:,j])
data_temp[voxel_x, voxel_y, voxel_z] = emb[:,j]
img_temp = nb.Nifti1Image(data_temp, mni_affine)
name_temp = os.path.join(embed_dir, 'mni3_component_%s.nii.gz' % (j+1))
nb.save(img_temp, name_temp)
