def _update_w(self):
svd_mdl = SVD(self.data)
svd_mdl.factorize()
U, S, V = svd_mdl.U, svd_mdl.S, svd_mdl.V
# The first left singular vector is nonnegative
# (abs is only used as values could be all negative)
self.W[:, 0] = np.sqrt(S[0, 0]) * np.abs(U[:, 0])
#The first right singular vector is nonnegative
self.H[0, :] = np.sqrt(S[0, 0]) * np.abs(V[0, :].T)
for i in range(1, self._num_bases):
# Form the rank one factor
Tmp = np.dot(U[:, i:i + 1] * S[i, i], V[i:i + 1, :])
# zero out the negative elements
Tmp = np.where(Tmp < 0, 0.0, Tmp)
# Apply 2nd SVD
svd_mdl_2 = SVD(Tmp)
svd_mdl_2.factorize()
u, s, v = svd_mdl_2.U, svd_mdl_2.S, svd_mdl_2.V
# The first left singular vector is nonnegative
self.W[:, i] = np.sqrt(s[0, 0]) * np.abs(u[:, 0])
#The first right singular vector is nonnegative
self.H[i, :] = np.sqrt(s[0, 0]) * np.abs(v[0, :].T)
def _update_w(self):
svd_mdl = SVD(self.data)
svd_mdl.factorize()
U, S, V = svd_mdl.U, svd_mdl.S, svd_mdl.V
# The first left singular vector is nonnegative
# (abs is only used as values could be all negative)
self.W[:,0] = np.sqrt(S[0,0]) * np.abs(U[:,0])
#The first right singular vector is nonnegative
self.H[0,:] = np.sqrt(S[0,0]) * np.abs(V[0,:].T)
for i in range(1,self._num_bases):
# Form the rank one factor
Tmp = np.dot(U[:,i:i+1]*S[i,i], V[i:i+1,:])
# zero out the negative elements
Tmp = np.where(Tmp < 0, 0.0, Tmp)
# Apply 2nd SVD
svd_mdl_2 = SVD(Tmp)
svd_mdl_2.factorize()
u, s, v = svd_mdl_2.U, svd_mdl_2.S, svd_mdl_2.V
# The first left singular vector is nonnegative
self.W[:,i] = np.sqrt(s[0,0]) * np.abs(u[:,0])
#The first right singular vector is nonnegative
self.H[i,:] = np.sqrt(s[0,0]) * np.abs(v[0,:].T)
def _update_w(self):
# compute eigenvectors and eigenvalues using SVD
svd_mdl = SVD(self.data)
svd_mdl.factorize()
# argsort sorts in ascending order -> do reverese indexing
# for accesing values in descending order
S = np.diag(svd_mdl.S)
order = np.argsort(S)[::-1]
# select only a few eigenvectors ...
if self._num_bases >0:
order = order[:self._num_bases]
self.W = svd_mdl.U[:,order]
self.eigenvalues = S[order]
def update_w(self):
# compute eigenvectors and eigenvalues using SVD
svd_mdl = SVD(self.data)
svd_mdl.factorize()
# argsort sorts in ascending order -> do reverese indexing
# for accesing values in descending order
S = np.diag(svd_mdl.S)
order = np.argsort(S)[::-1]
# select only a few eigenvectors ...
if self._num_bases >0:
order = order[:self._num_bases]
self.W = svd_mdl.U[:,order]
self.eigenvalues = S[order]
def comp_prob(d, k):
# compute statistical leverage score
c = np.round(k - k/5.0)
svd_mdl = SVD(d, k=c)
svd_mdl.factorize()
if scipy.sparse.issparse(self.data):
A = svd_mdl.V.multiply(svd_mdl.V)
## Rule 1
pcol = np.array(A.sum(axis=0)/k)
else:
A = svd_mdl.V[:k,:]**2.0
## Rule 1
pcol = A.sum(axis=0)/k
#c = k * np.log(k/ (self._eps**2.0))
#pcol = c * pcol.reshape((-1,1))
pcol /= np.sum(pcol)
return pcol
def comp_prob(d, k):
# compute statistical leverage score
c = np.round(k - k / 5.0)
svd_mdl = SVD(d, k=c)
svd_mdl.factorize()
if scipy.sparse.issparse(self.data):
A = svd_mdl.V.multiply(svd_mdl.V)
## Rule 1
pcol = np.array(A.sum(axis=0) / k)
else:
A = svd_mdl.V[:k, :]**2.0
## Rule 1
pcol = A.sum(axis=0) / k
#c = k * np.log(k/ (self._eps**2.0))
#pcol = c * pcol.reshape((-1,1))
pcol /= np.sum(pcol)
return pcol