skoudoro on master
NF: add unbiased groupwise bund… TEST: add tests for groupwise s… DOC: add groupwise slr example and 12 more (compare)
mapping.forward
ormapping.backward
.class mam(Metric):
""" Computes the mam distance between two streamlines. """
def init(self):
# For simplicity, features will be the vector between endpoints of a streamline.
super(mam, self).__init__(feature=VectorOfEndpointsFeature())
def are_compatible(self, shape1, shape2):
""" Checks if two features are vectors of same dimension.
Basically this method exists so we don't have to do this check
inside the `dist` method (speedup).
"""
return shape1 == shape2 and shape1[0] == 1
def dist(self, v1, v2):
track1 = np.ascontiguousarray(v1, dtype=np.float32)
t1_len = track1.shape[0]
track2 = np.ascontiguousarray(v2, dtype=np.float32)
t2_len = track2.shape[0]
# preallocate buffer array for track distance calculations
#distances_buffer = np.zeros((t1_len ,), dtype=np.float32)
min_t2t1 = np.zeros((t2_len ,), dtype=np.float32)
min_t1t2 = np.zeros((t1_len ,), dtype=np.float32)
for t2_pi in range(0,t2_len):
min_t2t1[t2_pi] = np.inf
for t1_pi in range(0,t1_len):
min_t1t2[t1_pi] = np.inf
# pointer to current point in track 1
t1_pt = track1
t2_pt = track2
# calculate min squared distance between each point in the two
# lines. Squared distance to delay doing the sqrt until after this
# speed-critical loop
for t1_pi in range(0,t1_len):
# pointer to current point in track 2
for t2_pi in range(0,t2_len):
d0 = t1_pt[t1_pi][0] - t2_pt[t2_pi][0]
d1 = t1_pt[t1_pi][1] - t2_pt[t2_pi][1]
delta2 = d0*d0 + d1*d1 #+ d2*d2
if delta2 < min_t1t2[t1_pi]:
min_t1t2[t1_pi]=delta2
for t2_pi in range(0,t2_len):
# pointer to current point in track 2
for t1_pi in range(0,t1_len):
d0 = t1_pt[t1_pi][0] - t2_pt[t2_pi][0]
d1 = t1_pt[t1_pi][1] - t2_pt[t2_pi][1]
delta2 = d0*d0 + d1*d1 #+ d2*d2
if delta2 < min_t2t1[t2_pi]:
min_t2t1[t2_pi]=delta2
# sqrt to get Euclidean distance from squared distance
for t1_pi in range(0,t1_len):
min_t1t2[t1_pi]=math.sqrt(min_t1t2[t1_pi])
for t2_pi in range(0,t2_len):
min_t2t1[t2_pi]=math.sqrt(min_t2t1[t2_pi])
mean_t2t1 = 0
mean_t1t2 = 0
for t1_pi in range(0, t1_len):
mean_t1t2+=min_t1t2[t1_pi]
mean_t1t2=mean_t1t2 / t1_len
for t2_pi in range(0, t2_len):
mean_t2t1+=min_t2t1[t2_pi]
mean_t2t1=mean_t2t1 / t2_len
return np.min((mean_t2t1,mean_t1t2))
metric = mam()
qb2 = QuickBundles(threshold=0.15, metric=metric)
clus = qb2.cluster(streamlines)
FetchFlow
load_tractogram
using the b0 image as a reference and plotting them with spherical ROIs using sphere_actor
and stream_actor
, the two are in very different spaces. It looks like the origin is shifted and the spatial scaling is different. Is there some transformation happening under the hood in load_tractogram
somewhere? If so, I'm guessing applying the same transformation to my coordinates will solve the issue.
IdentityFeature
instead of VectorOfEndpointsFeature()
and then I will remove shape1[0] == 1
on your function are_compatible
. You do not need vector in your case for MAM. Let me know if it works, otherwise create an issue on DIPY, it will be easier to help you