def makeWeights(options, X, Y, GX, GY):
# hack to compute scores using normalized projections
if options.normalize_projections == 1:
X = common.normalize_rows(X, 2) # note that normalize_rows works on arrays, not matrices.
Y = common.normalize_rows(Y, 2)
X = np.mat(X)
Y = np.mat(Y)
if GX is not None or GY is not None:
GX = np.mat(GX)
GY = np.mat(GY)
Z = np.mat(0)
if options.weight_type == 'inner':
U = X*Y.T # linear kernel
if options.alpha > 0: # TODO: add higher order graphs
Z = GX * U * GY.T
W = (1-options.alpha)*U + options.alpha*Z
else:
W = U
W = np.max(W) - W
elif options.weight_type == 'dist':
U = common.dist(X, Y)
if options.alpha > 0: # TODO: add higher order graph
Z = common.dist(GX * X, GY * Y)
W = (1-options.alpha)*U + options.alpha*Z
else:
W = U
elif options.weight_type == 'sqrdist':
U = common.dist(X, Y, metric='sqeuclidean')
if options.alpha > 0: # TODO: add higher order graph
Z = common.dist(GX * X, GY * Y, metric='sqeuclidean')
W = (1-options.alpha)*U + options.alpha*Z
else:
W = U
elif options.weight_type == 'graph_min_dist':
U = common.dist(X, Y)
GX = np.array(GX)
GY = np.array(GY)
if options.alpha > 0: # TODO: add higher order graph
Z, IX, IY = cy_getGraphMinDist(GX, GY, U)
W = (1-options.alpha)*U + options.alpha*Z
else:
W = []
saveWUZ(U, W, Z, options)
#print 'norm(U) = ', np.linalg.norm(U, 2), '| norm(Z) = ', np.linalg.norm(Z, 2)
return W, U, Z
def setupFeatures(self, options=None):
#logFr = np.log(X.freq)
# L = strings.strlen(self.words)
# normalize the features
if self.isPickled():
(orthoDD, orthoFeatures) = strings.to_ngram_dictionary(self.words, affix=True)
if options.log_features == 1:
self.orthoMSK = MSK(orthoDD, self.words, orthoFeatures).log(offset=1).normalize(norm='l2')
self.contextMSK = MSK(self.repr, self.words, self.featureNames).log(offset=1).normalize(norm='l2')
else:
self.orthoMSK = MSK(orthoDD, self.words, orthoFeatures).normalize(norm='l2')
self.contextMSK = MSK(self.repr, self.words, self.featureNames).normalize(norm='l2')
else:
self.features = common.normalize_rows(self.features)
