def computeCorr(X, Y, options):
X = np.mat(X)
Y = np.mat(Y)
(NX, Dx) = X.shape
(NY, Dy) = Y.shape
assert NX == NY
N = NX
tau = options.tau
covar_type = options.covar_type
## process weights
if options.cca_weights is None:
weights = common.asVector(np.ones((N, 1)))
else:
print 'CCA Weights:', options.cca_weights
sum_weights = np.sum(options.cca_weights)
weights = common.asVector(N*options.cca_weights / sum_weights)
weights = common.asVector(weights)
print 'weights:', weights
assert np.all(weights > 0)
assert len(weights) == N
if covar_type == 'outer':
W = np.diag(weights)
Cxx = X.T*W*X/N # C[:Dx, :Dx]
Cxy = X.T*W*Y/N # C[:Dx, Dx:]
Cyy = Y.T*W*Y/N # C[Dx:, Dx:]
else:
W = np.diag(np.sqrt(weights))
Z = np.c_[W*X, W*Y] # stack X Y by rows
C = np.cov(Z.T)
Cxx = C[:Dx, :Dx]
Cxy = C[:Dx, Dx:]
Cyy = C[Dx:, Dx:]
Cxx = np.mat((1-tau)*Cxx + tau*np.eye(Dx))
Cxy = np.mat(Cxy)
Cyy = np.mat((1-tau)*Cyy + tau*np.eye(Dy))
return Cxx, Cxy, Cyy
def output_edges(M, L, reverseLookup):
L = np.array(L)
print >> sys.stderr, 'class:', L.__class__
# output top K values.
Nw, Nf = M.M.shape
for iw in xrange(Nw):
word = M.reverseStrings[iw] # get words
vw = L[iw, :]
J = common.asVector(np.argsort(vw))
count = 0
if iw % 100 == 0:
print >> sys.stderr, "word i", iw, '=', word
for jf in reversed(J):
feature = reverseLookup[jf]
if feature in M.strings and L[iw, jf] > 0:
print "%s,%s,%f" % (word, feature, L[iw, jf])
#count += 1
if count == options.K:
break
def hist(v, bins=10):
v = common.asVector(v)
H = plt.hist(v, bins=bins)
show()