def data_compression(fmri_masked, mask_img, mask_np, output_size):
"""
data : array_like
A matrix of shape (`V`, `N`) with `V` voxels `N` timepoints
The functional dataset that needs to be reduced
mask : a numpy array of the mask
output_size : integer
The number of elements that the data should be reduced to
"""
## Transform nifti files to a data matrix with the NiftiMasker
import time
from nilearn import input_data
datacompressiontime = time.time()
nifti_masker = input_data.NiftiMasker(mask_img=mask_img,
memory='nilearn_cache',
mask_strategy='background',
memory_level=1,
standardize=False)
ward = []
# Perform Ward clustering
from sklearn.feature_extraction import image
shape = mask_np.shape
connectivity = image.grid_to_graph(n_x=shape[0],
n_y=shape[1],
n_z=shape[2],
mask=mask_np)
#import pdb;pdb.set_trace()
from sklearn.cluster import FeatureAgglomeration
start = time.time()
ward = FeatureAgglomeration(n_clusters=output_size,
connectivity=connectivity,
linkage='ward')
ward.fit(fmri_masked)
#print("Ward agglomeration compressing voxels into clusters: %.2fs" % (time.time() - start))
labels = ward.labels_
#print ('Extracting reduced Dimension Data')
data_reduced = ward.transform(fmri_masked)
fmri_masked = []
#print('Data compression took ', (time.time()- datacompressiontime), ' seconds')
return {'data': data_reduced, 'labels': labels}
def data_compression(fmri_masked, mask_img, mask_np, compression_dim):
# TODO @AKI update doc
"""
Perform...
Parameters
----------
fmri_masked : np.ndarray[ndim=2]
A matrix of shape (`V`, `N`) with `V` voxels `N` timepoints
The functional dataset that needs to be reduced
mask_img : an nibabel img object of the mask
mask_np : a numpy array of the mask
compression_dim : integer
The number of elements that the data should be reduced to
Returns
-------
A dictionaty ...
"""
from sklearn.feature_extraction import image
from sklearn.cluster import FeatureAgglomeration
# Perform Ward clustering
shape = mask_np.shape
connectivity = image.grid_to_graph(n_x=shape[0],
n_y=shape[1],
n_z=shape[2],
mask=mask_np)
ward = FeatureAgglomeration(n_clusters=compression_dim,
connectivity=connectivity,
linkage='ward')
ward.fit(fmri_masked)
labels = ward.labels_
data_reduced = ward.transform(fmri_masked)
return {
'compressor': ward,
'compressed': data_reduced,
'labels': labels,
}
def cross_cluster_timeseries(data1, data2, roi_mask_nparray, n_clusters, similarity_metric, affinity_threshold, cluster_method = 'ward'):
"""
Cluster a timeseries dataset based on its relationship to a second timeseries dataset
Parameters
----------
data1 : array_like
A matrix of shape (`N`, `M`) with `N1` samples and `M1` dimensions.
This is the matrix to receive cluster assignment
data2 : array_like
A matrix of shape (`N`, `M`) with `N2` samples and `M2` dimensions.
This is the matrix with which distances will be calculated to assign clusters to data1
n_clusters : integer
Number of clusters
similarity_metric : {'euclidean', 'correlation', 'minkowski', 'cityblock', 'seuclidean'}
Type of similarity measure for distance matrix. The pairwise similarity measure
specifies the edges of the similarity graph. 'data' option assumes X as the similarity
matrix and hence must be symmetric. Default is kneighbors_graph [1]_ (forced to be
symmetric)
affinity_threshold : float
Threshold of similarity metric when 'correlation' similarity metric is used.
Returns
-------
y_pred : array_like
Predicted cluster labels
Examples
--------
np.random.seed(30)
offset = np.random.randn(30)
x1 = np.random.randn(200,30) + 2*offset
x2 = np.random.randn(100,30) + 44*np.random.randn(30)
x3 = np.random.randn(400,30)
sampledata1 = np.vstack((x1,x2,x3))
np.random.seed(99)
offset = np.random.randn(30)
x1 = np.random.randn(200,30) + 2*offset
x2 = np.random.randn(100,30) + 44*np.random.randn(30)
x3 = np.random.randn(400,30)
sampledata2 = np.vstack((x1,x2,x3))
cross_cluster(sampledata1, sampledata2, 3, 'euclidean')
References
----------
https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.distance.cdist.html#scipy.spatial.distance.cdist
http://scikit-learn.org/stable/modules/clustering.html#spectral-clustering
"""
import scipy as sp
import time
import sklearn as sk
from sklearn import cluster, datasets, preprocessing
from sklearn.cluster import FeatureAgglomeration
from sklearn.feature_extraction import image
print("Calculating Cross-clustering")
print("Calculating pairwise distances between areas")
dist_btwn_data_1_2 = np.array(sp.spatial.distance.cdist(data1.T, data2.T, metric = similarity_metric))
sim_btwn_data_1_2=1-dist_btwn_data_1_2
sim_btwn_data_1_2[np.isnan(sim_btwn_data_1_2)]=0
sim_btwn_data_1_2[sim_btwn_data_1_2<affinity_threshold]=0
print("Calculating pairwise distances between voxels in ROI 1 ")
dist_of_1 = sp.spatial.distance.pdist(sim_btwn_data_1_2, metric = 'euclidean')
dist_matrix = sp.spatial.distance.squareform(dist_of_1)
sim_matrix=1-sk.preprocessing.normalize(dist_matrix, norm='max')
sim_matrix[sim_matrix<affinity_threshold]=0
if cluster_method == 'ward':
# ## BEGIN WARD CLUSTERING CODE
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
if roi_mask_nparray!='empty':
#import pdb; pdb.set_trace()
shape = roi_mask_nparray.shape
connectivity = image.grid_to_graph(n_x=shape[0], n_y=shape[1],
n_z=shape[2], mask=roi_mask_nparray)
ward = FeatureAgglomeration(n_clusters=n_clusters, connectivity=connectivity,
linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
else:
print("Calculating Hierarchical Cross-clustering")
ward = FeatureAgglomeration(n_clusters=n_clusters, affinity='euclidean', linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
# # END WARD CLUSTERING CODE
else:
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
#cluster_method== 'spectral':
#Spectral method
spectral = cluster.SpectralClustering(n_clusters, eigen_solver='arpack', random_state = 5, affinity="precomputed", assign_labels='discretize')
spectral.fit(sim_matrix)
y_pred = spectral.labels_.astype(np.int)
#
# BEGIN SPECTRAL CLUSTERING CODE
# END SPECTRAL CLUSTERING CODE
# sim_matrix[np.isnan((sim_matrix))]=0
# sim_matrix[sim_matrix<0]=0
# sim_matrix[sim_matrix>1]=1
## BEGIN WARD CLUSTERING CODE
# print("Calculating Hierarchical Cross-clustering")
# ward = FeatureAgglomeration(n_clusters=n_clusters, affinity='euclidean', linkage='ward')
# ward.fit(sim_matrix)
# y_pred = ward.labels_.astype(np.int)
#
## END WARD CLUSTERING CODE
# # BEGIN SPECTRAL CLUSTERING CODE
# spectral = cluster.SpectralClustering(n_clusters, eigen_solver='arpack', random_state = 5, affinity="precomputed", assign_labels='discretize')
# spectral.fit(sim_matrix)
# y_pred = spectral.labels_.astype(np.int)
# # END SPECTRAL CLUSTERING CODE
return y_pred
def cluster_timeseries(X, roi_mask_nparray, n_clusters, similarity_metric, affinity_threshold, cluster_method = 'ward'):
"""
Cluster a given timeseries
Parameters
----------
X : array_like
A matrix of shape (`N`, `M`) with `N` samples and `M` dimensions
n_clusters : integer
Number of clusters
similarity_metric : {'k_neighbors', 'correlation', 'data'}
Type of similarity measure for spectral clustering. The pairwise similarity measure
specifies the edges of the similarity graph. 'data' option assumes X as the similarity
matrix and hence must be symmetric. Default is kneighbors_graph [1]_ (forced to be
symmetric)
affinity_threshold : float
Threshold of similarity metric when 'correlation' similarity metric is used.
Returns
-------
y_pred : array_like
Predicted cluster labels
Examples
--------
References
----------
.. [1] http://scikit-learn.org/dev/modules/generated/sklearn.neighbors.kneighbors_graph.html
if similarity_metric == 'correlation':
# Calculate empirical correlation matrix between samples
Xn = X - X.mean(1)[:,np.newaxis]
Xn = Xn/np.sqrt( (Xn**2.).sum(1)[:,np.newaxis] )
C_X = np.dot(Xn, Xn.T)
C_X[C_X < affinity_threshold] = 0
from scipy.sparse import lil_matrix
C_X = lil_matrix(C_X)
elif similarity_metric == 'data':
C_X = X
elif similarity_metric == 'k_neighbors':
from sklearn.neighbors import kneighbors_graph
C_X = kneighbors_graph(X, n_neighbors=neighbors)
C_X = 0.5 * (C_X + C_X.T)
else:
raise ValueError("Unknown value for similarity_metric: '%s'." % similarity_metric)
#sklearn code is not stable for bad clusters which using correlation as a stability metric
#tends to give for more info see:
#http://scikit-learn.org/dev/modules/clustering.html#spectral-clustering warning
#from sklearn import cluster
#algorithm = cluster.SpectralClustering(k=n_clusters, mode='arpack')
#algorithm.fit(C_X)
#y_pred = algorithm.labels_.astype(np.int)
from python_ncut_lib import ncut, discretisation
eigen_val, eigen_vec = ncut(C_X, n_clusters)
eigen_discrete = discretisation(eigen_vec)
#np.arange(n_clusters)+1 isn't really necessary since the first cluster can be determined
#by the fact that the each cluster is a disjoint set
y_pred = np.dot(eigen_discrete.toarray(), np.diag(np.arange(n_clusters))).sum(1)
"""
import sklearn as sk
from sklearn import cluster, datasets, preprocessing
import scipy as sp
import time
from sklearn.cluster import FeatureAgglomeration
from sklearn.feature_extraction import image
print('Beginning Calculating pairwise distances between voxels')
X = np.array(X)
X_dist = sp.spatial.distance.pdist(X.T, metric = similarity_metric)
temp=X_dist
temp[np.isnan(temp)]=0
tempmax=temp.max()
X_dist = sp.spatial.distance.squareform(X_dist)
X_dist[np.isnan(X_dist)]=tempmax
#import pdb;pdb.set_trace()
sim_matrix=1-sk.preprocessing.normalize(X_dist, norm='max')
sim_matrix[sim_matrix<affinity_threshold]=0
#import pdb;pdb.set_trace()
if cluster_method == 'ward':
# ## BEGIN WARD CLUSTERING CODE
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
print("ward")
if roi_mask_nparray!='empty':
#import pdb; pdb.set_trace()
shape = roi_mask_nparray.shape
connectivity = image.grid_to_graph(n_x=shape[0], n_y=shape[1],
n_z=shape[2], mask=roi_mask_nparray)
ward = FeatureAgglomeration(n_clusters=n_clusters, connectivity=connectivity,
linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
else:
print("Calculating Hierarchical Clustering")
ward = FeatureAgglomeration(n_clusters=n_clusters, affinity='euclidean', linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
# # END WARD CLUSTERING CODE
else:
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
print("spectral")
#cluster_method== 'spectral':
#Spectral method
spectral = cluster.SpectralClustering(n_clusters, eigen_solver='arpack', random_state = 5, affinity="precomputed", assign_labels='discretize')
spectral.fit(sim_matrix)
y_pred = spectral.labels_.astype(np.int)
#
# BEGIN SPECTRAL CLUSTERING CODE
# END SPECTRAL CLUSTERING CODE
return y_pred
def cross_cluster_timeseries(data1,
data2,
roi_mask_data,
n_clusters,
similarity_metric,
affinity_threshold,
cluster_method='ward',
random_state=None):
"""
Cluster a timeseries dataset based on its relationship
to a second timeseries dataset
Parameters
----------
data1 : array_like
A matrix of shape (`N`, `M`) with `N1` samples and `M1` dimensions.
This is the matrix to receive cluster assignment
data2 : array_like
A matrix of shape (`N`, `M`) with `N2` samples and `M2` dimensions.
This is the matrix with which distances will be calculated to assign
clusters to data1
roi_mask_data : array_like
An array that contains a binary mask of the region of interest (ROI)
being parcellated.
n_clusters : integer
Number of clusters
similarity_metric : {'euclidean', 'correlation', 'minkowski', 'cityblock',
'seuclidean'}
Type of similarity measure for distance matrix. The pairwise similarity
measure specifies the edges of the similarity graph. 'data' option
assumes X as the similarity matrix and hence must be symmetric.
Default is kneighbors_graph [1]_ (forced to be symmetric)
affinity_threshold : float
Threshold of similarity metric when 'correlation' similarity metric
is used.
cluster_method : {'ward', 'spectral', 'kmeans', 'gaussianmixture'}
A string that says which cluster method to use.
random_state : integer
the random state to seed the bootstrap
Returns
-------
y_pred : array_like
Predicted cluster labels
Examples
--------
np.random.seed(30)
offset = np.random.randn(30)
x1 = np.random.randn(200, 30) + 2 * offset
x2 = np.random.randn(100, 30) + 44 * np.random.randn(30)
x3 = np.random.randn(400, 30)
sampledata1 = np.vstack((x1, x2, x3))
np.random.seed(99)
offset = np.random.randn(30)
x1 = np.random.randn(200, 30) + 2 * offset
x2 = np.random.randn(100, 30) + 44 * np.random.randn(30)
x3 = np.random.randn(400, 30)
sampledata2 = np.vstack((x1, x2, x3))
cross_cluster(sampledata1, sampledata2, 3, 'euclidean')
References
----------
https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.distance.cdist.html#scipy.spatial.distance.cdist
http://scikit-learn.org/stable/modules/clustering.html#spectral-clustering
"""
from scipy.spatial.distance import pdist, cdist, squareform
from sklearn.preprocessing import normalize
from sklearn.feature_extraction import image
from sklearn.cluster import FeatureAgglomeration, KMeans, SpectralClustering
from sklearn.mixture import GaussianMixture
dist_btwn_data_1_2 = np.array(
cdist(data1.T, data2.T, metric=similarity_metric))
max_dist = np.nanmax(dist_btwn_data_1_2)
dist_btwn_data_1_2[np.isnan(dist_btwn_data_1_2)] = max_dist
dist_of_1 = pdist(dist_btwn_data_1_2, metric='euclidean')
dist_matrix = squareform(dist_of_1)
sim_matrix = 1 - normalize(dist_matrix, norm='max')
sim_matrix[sim_matrix < affinity_threshold] = 0
if cluster_method == 'ward':
if roi_mask_data is not None:
shape = roi_mask_data.shape
connectivity = image.grid_to_graph(n_x=shape[0],
n_y=shape[1],
n_z=shape[2],
mask=roi_mask_data)
ward = FeatureAgglomeration(n_clusters=n_clusters,
connectivity=connectivity,
linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
else:
ward = FeatureAgglomeration(n_clusters=n_clusters,
affinity='euclidean',
linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
elif cluster_method == 'spectral':
spectral = SpectralClustering(n_clusters,
eigen_solver='arpack',
affinity="precomputed",
assign_labels='discretize',
random_state=random_state)
spectral.fit(sim_matrix)
y_pred = spectral.labels_.astype(np.int)
elif cluster_method == 'kmeans':
kmeans = KMeans(n_clusters=n_clusters,
init='k-means++',
n_init=10,
random_state=random_state)
kmeans.fit(sim_matrix)
y_pred = kmeans.labels_.astype(np.int)
elif cluster_method == 'gaussianmixture':
gaussianmixture = GaussianMixture(n_components=n_clusters,
init_params='kmeans',
random_state=random_state)
y_pred = gaussianmixture.fit_predict(sim_matrix)
return y_pred
def cluster_timeseries(X,
roi_mask_data,
n_clusters,
similarity_metric,
affinity_threshold,
cluster_method='ward',
random_state=None):
"""
Cluster a given timeseries
Parameters
----------
X : array_like
A matrix of shape (`N`, `M`) with `N` samples and `M` dimensions
roi_mask_data : array_like
An array that contains a binary mask of the region of interest (ROI)
being parcellated.
n_clusters : integer
Number of clusters
similarity_metric : {'k_neighbors', 'correlation', 'data'}
Type of similarity measure for spectral clustering. The pairwise
similarity measure specifies the edges of the similarity graph.
'data' option assumes X as the similarity matrix and hence must be
symmetric. Default is kneighbors_graph [1]_ (forced to be symmetric)
affinity_threshold : float
Threshold of similarity metric when 'correlation' similarity
metric is used.
cluster_method : {'ward', 'spectral', 'kmeans', 'gaussianmixture'}
A string that says which cluster method to use.
random_state : integer
the random state to seed the bootstrap
Returns
-------
y_pred : array_like
Predicted cluster labels
Examples
--------
References
----------
.. [1] http://scikit-learn.org/dev/modules/generated/sklearn.neighbors.kneighbors_graph.html
"""
import numpy as np
import scipy as sp
import sklearn as sk
from sklearn.feature_extraction import image
from sklearn.cluster import FeatureAgglomeration, SpectralClustering, KMeans
from sklearn.mixture import GaussianMixture
X = np.array(X)
X_dist = sp.spatial.distance.pdist(X.T, metric=similarity_metric)
max_dist = np.nanmax(X_dist)
X_dist = sp.spatial.distance.squareform(X_dist)
X_dist[np.isnan(X_dist)] = max_dist
sim_matrix = 1 - sk.preprocessing.normalize(X_dist, norm='max')
sim_matrix[sim_matrix < affinity_threshold] = 0
print("Calculating Hierarchical Clustering")
cluster_method = cluster_method.lower()
if cluster_method == 'ward':
if roi_mask_data is not None:
shape = roi_mask_data.shape
connectivity = image.grid_to_graph(n_x=shape[0],
n_y=shape[1],
n_z=shape[2],
mask=roi_mask_data)
ward = FeatureAgglomeration(n_clusters=n_clusters,
connectivity=connectivity,
linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
else:
ward = FeatureAgglomeration(n_clusters=n_clusters,
affinity='euclidean',
linkage='ward')
ward.fit(sim_matrix)
y_pred = ward.labels_.astype(np.int)
elif cluster_method == 'spectral':
spectral = SpectralClustering(n_clusters,
eigen_solver='arpack',
affinity="precomputed",
assign_labels='discretize',
random_state=random_state)
spectral.fit(sim_matrix)
y_pred = spectral.labels_.astype(np.int)
elif cluster_method == 'kmeans':
kmeans = KMeans(n_clusters=n_clusters,
init='k-means++',
n_init=10,
random_state=random_state)
kmeans.fit(sim_matrix)
y_pred = kmeans.labels_.astype(np.int)
elif cluster_method == 'gaussianmixture':
gaussianmixture = GaussianMixture(n_components=n_clusters,
init_params='kmeans',
random_state=random_state)
y_pred = gaussianmixture.fit_predict(sim_matrix)
return y_pred
