def localscale_neighbors_opposing_r_x(self, x, y, k, N, rho):
"""
nearest neighbors from opposing classes with local scale computed
based on D and B
(class other than that of the example considered)
x - data points
y - class membership
k - nearest neighbors to be considered from opposing class for
computing local scale
rho - List of multipliers to be used when computing the graph with
local scale. For a value r from this, an edge will be created when
d(x_i, x_j) < r*local_scale_i*local_scale_j and i, j belong to
opposing classes
N - max number of nearest neighbors from opposing classes (user parameter)
2 hop neighbors are then computed to complete the triangles and form
simplices
SIMILAR to localscale_neighbors_opposing_r but uses Cython for speedup
This involves restricting the maximum number of neighbors for each point to N
"""
numFilt = len(rho)
nTriv = np.zeros(numFilt)
m = x.shape[0]
rows = np.ceil(float(numFilt) / 5)
cols = 5
# Compute opposing neighbors
S, DN = self.opposing_neighbors(x, y, N)
# Sigma array
sigarr = np.sqrt(DN[:, k - 1].ravel())
# Other inits
C = -1 * np.ones((m, N + 1), dtype=np.int32)
Pupd = np.zeros((m, m), dtype=np.int32)
P = np.zeros((m, m), dtype=np.int32)
inds = np.zeros(m, dtype=np.int32)
#t1 = time()
for j, r in enumerate(rho):
# Re-initialize some arrays
if j > 0:
C.fill(-1)
inds.fill(0)
Pupd.fill(0)
# Call Cython functions for updating P
init_C_inds(C, inds)
compute_C_closest(C, S, inds, DN, r, sigarr)
update_P(C, inds, Pupd, 1)
np.maximum(Pupd, Pupd.transpose(), out=Pupd)
np.add(P, Pupd, out=P)
nTriv[j] = np.sum(inds == 1)
#sum(np.sum(C, axis = 0) == 1)
if self.showComplexes:
f = plt.figure(num=1000, figsize=(rows * 5, cols * 5))
titl = "r = " + str(r)
self.show_simplicial_complex(x, y, Pupd.astype(np.double), f,
rows, cols, j + 1, titl)
if self.saveComplexes:
titl = "r = " + str(r)
self.save_simplicial_complex(x, y, Pupd.astype(np.double),
rows, cols, j + 1, titl)
P = numFilt - P
return P, nTriv, numFilt, C
def localscale_neighbors_opposing_x(self, x, y, neighborsv, N):
"""
k nearest neighbors from opposing classes
(class other than that of the example considered)
x - data points
y - class membership
neighborsv - nearest neighbors to be considered from opposing class for
computing local scale
N - max number of nearest neighbors from opposing classes (user parameter)
2 hop neighbors are then computed to complete the triangles and form
simplices
SIMILAR to localscale_neighbors_opposing but uses Cython for speedup
This involves restricting the maximum number of neighbors for each point to N
"""
numFilt = len(neighborsv)
nTriv = np.zeros(numFilt)
m = x.shape[0]
rows = np.ceil(float(numFilt) / 5)
cols = 5
# Adjust N depending on max(neighborsv)
N = int(np.maximum(N, np.max(neighborsv)))
print("N adjusted to %d based on requested max nearest neighbors" %
(N))
# Compute opposing neighbors
S, DN = self.opposing_neighbors(x, y, N)
# Other inits
C = -1 * np.ones((m, N + 1), dtype=np.int32)
Pupd = np.zeros((m, m), dtype=np.int32)
P = np.zeros((m, m), dtype=np.int32)
inds = np.zeros(m, dtype=np.int32)
for j, k in enumerate(neighborsv):
# Sigma array
sigarr = np.sqrt(DN[:, int(k) - 1].ravel())
# Re-initialize some arrays
if j > 0:
C.fill(-1)
inds.fill(0)
Pupd.fill(0)
# Call Cython functions for updating P
init_C_inds(C, inds)
compute_C_closest(C, S, inds, DN, 1.0, sigarr)
update_P(C, inds, Pupd, 1)
np.maximum(Pupd, Pupd.transpose(), out=Pupd)
np.add(P, Pupd, out=P)
nTriv[j] = np.sum(inds == 1)
if self.showComplexes:
f = plt.figure(num=1000, figsize=(rows * 5, cols * 5))
titl = "k = " + str(neighborsv[j])
self.show_simplicial_complex(x, y, Pupd.astype(np.double), f,
rows, cols, j + 1, titl)
if self.saveComplexes:
titl = "r = " + str(neighborsv[j])
self.save_simplicial_complex(x, y, Pupd.astype(np.double),
rows, cols, j + 1, titl)
P = numFilt - P
return P, nTriv, numFilt
