import sys as _sys
import mdp
from mdp import Node,NodeException,numx,numx_rand
from mdp.nodes import WhiteningNode
from mdp.utils import (DelayCovarianceMatrix,MultipleCovarianceMatrices
rotate, mult)
# rename often used functions
sum, cos, sin, PI = numx.sum, numx.cos, numx.sin, numx.pi
SQRT_EPS_D = numx.sqrt(numx.finfo('d').eps)
def _triu(m, k=0):
""" returns the elements on and above the k-th diagonal of m. k=0 is the
main diagonal, k > 0 is above and k < 0 is below the main diagonal."""
N = m.shape[0]
M = m.shape[1]
x = numx.greater_equal(numx.subtract.outer(numx.arange(N),
numx.arange(M)),1-k)
out = (1-x)*m
return out
#############
class ISFANode(Node):
"""
Perform Independent Slow Feature Analysis on the input data.
Internal variables of interest:
self.RP -- the global rotation-permutation matrix. This is the filter
applied on input_data to get output_data
self.RPC -- the *complete* global rotation-permutation matrix. This
is a matrix of dimension input_dim x input_dim
(the 'outer space' is retained)
self.covs -- A mdp.utils.MultipleCovarianceMatrices instance containing
the current time-delayed covariance matrices of the
input_data. After convergence the uppermost
output_dim x output_dim submatrices should be almost
diagonal.
self.covs[n-1] is the covariance matrix relative to
the n-th time-lag
Note: they are not cleared after convergence. If you
need to free some memory, you can safely delete them with
>>> del self.covs
self.initial_contrast -- a dictionary with the starting contrast
and the SFA and ICA parts of it.
self.final_contrast -- like the above but after convergence.
Note: If you intend to use this node for large datasets please have
a look at the stop_training method documentation for
speeding things up.
References:
Blaschke, T. , Zito, T., and Wiskott, L. (2007).
Independent Slow Feature Analysis and Nonlinear Blind Source Separation.
Neural Computation 19(4):994-1021 (2007)
http://itb.biologie.hu-berlin.de/~wiskott/Publications/BlasZitoWisk2007-ISFA-NeurComp.pdf
"""
def __init__(self, lags=1, sfa_ica_coeff=(1., 1.), icaweights=None,
sfaweights=None, whitened=False, white_comp = None,
white_parm = None, eps_contrast=1e-6, max_iter=10000,
RP=None, verbose=False, input_dim=None, output_dim=None,
dtype=None):
"""
Perform Independent Slow Feature Analysis.
The notation is the same used in the paper by Blaschke et al. Please
refer to the paper for more information.
Keyword arguments:
lags -- list of time-lags to generate the time-delayed covariance
matrices (in the paper this is the set of \tau). If
lags is an integer, time-lags 1,2,...,'lags' are used.
Note that time-lag == 0 (instantaneous correlation) is
always implicitly used.
sfa_ica_coeff -- a list of float with two entries, which defines the
weights of the SFA and ICA part of the
objective function. They are called b_{SFA} and
b_{ICA} in the paper.
sfaweights -- weighting factors for the covariance matrices relative
to the SFA part of the objective function (called
\kappa_{SFA}^{\tau} in the paper). Default is
[1., 0., ..., 0.]
For possible values see the description of icaweights.
icaweights -- weighting factors for the cov matrices relative
to the ICA part of the objective function (called
\kappa_{ICA}^{\tau} in the paper). Default is 1.
Possible values are:
an integer n: all matrices are weighted the same
(note that it does not make sense to
have n != 1)
a list or array of floats of len == len(lags): each
element of the list is used for
weighting the corresponding matrix
None: use the default values.
whitened -- True if input data is already white, False otherwise (the
data will be whitened internally).
white_comp -- If whitened is False, you can set 'white_comp' to the
number of whitened components to keep during the
calculation (i.e., the input dimensions are reduced to
white_comp by keeping the components of largest variance).
white_parm -- a dictionary with additional parameters for whitening.
It is passed directly to the WhiteningNode constructor.
Ex: white_parm = { 'svd' : True }
eps_contrast -- Convergence is achieved when the relative
improvement in the contrast is below this threshold.
Values in the range [1E-4, 1E-10] are usually
reasonable.
max_iter -- If the algorithms does not achieve convergence within
max_iter iterations raise an Exception. Should be
larger than 100.
RP -- Starting rotation-permutation matrix. It is an
input_dim x input_dim matrix used to initially rotate the
input components. If not set, the identity matrix is used.
In the paper this is used to start the algorithm at the
SFA solution (which is often quite near to the optimum).
verbose -- print progress information during convergence. This can
slow down the algorithm, but it's the only way to see
the rate of improvement and immediately spot if something
is going wrong.
output_dim -- sets the number of independent components that have to
be extracted. Note that if this is not smaller than
input_dim, the problem is solved linearly and SFA
would give the same solution only much faster.
"""
# check that the "lags" argument has some meaningful value
if isinstance(lags, (int, long)):
lags = range(1, lags+1)
elif isinstance(lags, (list, tuple)):
lags = numx.array(lags, "i")
elif isinstance(lags, numx.ndarray):
if not (lags.dtype.char in ['i', 'l']):
err_str = "lags must be integer!"
raise NodeException(err_str)
else:
pass
else:
err_str = ("Lags must be int, list or array. Found "
"%s!" % (type(lags).__name__))
raise NodeException(err_str)
self.lags = lags
# sanity checks for weights
if icaweights is None:
self.icaweights = 1.
else:
if (len(icaweights) != len(lags)):
err = ("icaweights vector length is %d, "
"should be %d" % (str(len(icaweights)), str(len(lags))))
raise NodeException(err)
self.icaweights = icaweights
if sfaweights is None:
self.sfaweights = [0]*len(lags)
self.sfaweights[0] = 1.
else:
if (len(sfaweights) != len(lags)):
err = ("sfaweights vector length is %d, "
"should be %d" % (str(len(sfaweights)), str(len(lags))))
raise NodeException(err)
self.sfaweights = sfaweights
# store attributes
self.sfa_ica_coeff = sfa_ica_coeff
self.max_iter = max_iter
self.verbose = verbose
self.eps_contrast = eps_contrast
# if input is not white, insert a WhiteningNode
self.whitened = whitened
if not whitened:
if white_parm is None:
white_parm = {}
if output_dim is not None:
white_comp = output_dim
elif white_comp is not None:
output_dim = white_comp
self.white = WhiteningNode(input_dim=input_dim,
output_dim=white_comp,
dtype=dtype, **white_parm)
# initialize covariance matrices
self.covs = [ DelayCovarianceMatrix(dt, dtype=dtype) for dt in lags ]
# initialize the global rotation-permutation matrix
# if not set that we'll eventually be an identity matrix
self.RP = RP
# initialize verbose structure to print nice and useful progress info
if verbose:
info = { 'sweep' : max(len(str(self.max_iter)), 5),
'perturbe': max(len(str(self.max_iter)), 5),
'float' : 5+8,
'fmt' : "%.5e",
'sep' : " | "}
f1 = "Sweep".center(info['sweep'])
f1_2 = "Pertb". center(info['perturbe'])
f2 = "SFA part".center(info['float'])
f3 = "ICA part".center(info['float'])
f4 = "Contrast".center(info['float'])
header = info['sep'].join([f1, f1_2, f2, f3, f4])
info['header'] = header+'\n'
info['line'] = len(header)*"-"
self._info = info
# finally call base class constructor
super(ISFANode, self).__init__(input_dim, output_dim, dtype)
def _get_supported_dtypes(self):
"""Return the list of dtypes supported by this node."""
return ['float64']
def _set_dtype(self, dtype):
# when typecode is set, we set the whitening node if needed and
# the SFA and ICA weights
self._dtype = dtype
if not self.whitened and self.white.dtype is None:
self.white.dtype = dtype
self.icaweights = numx.array(self.icaweights, dtype)
self.sfaweights = numx.array(self.sfaweights, dtype)
def _set_input_dim(self, n):
self._input_dim = n
if not self.whitened and self.white.output_dim is not None:
self._effective_input_dim = self.white.output_dim
else:
self._effective_input_dim = n
def _train(self, x):
# train the whitening node if needed
if not self.whitened:
self.white.train(x)
# update the covariance matrices
[self.covs[i].update(x) for i in range(len(self.lags))]
def _execute(self, x):
# filter through whitening node if needed
if not self.whitened:
x = self.white.execute(x)
# rotate input
return mult(x, self.RP)
def _inverse(self, y):
# counter-rotate input
x = mult(y, self.RP.T)
# invert whitening node if needed
if not self.whitened:
x = self.white.inverse(x)
return x
def _fmt_prog_info(self, sweep, pert, contrast, sfa = None, ica = None):
# for internal use only!
# format the progress information
# don't try to understand this code: it Just Works (TM)
fmt = self._info
sweep_str = str(sweep).rjust(fmt['sweep'])
pert_str = str(pert).rjust(fmt['perturbe'])
if sfa is None:
sfa_str = fmt['float']*' '
else:
sfa_str = (fmt['fmt']%(sfa)).rjust(fmt['float'])
if ica is None:
ica_str = fmt['float']*' '
else:
ica_str = (fmt['fmt'] % (ica)).rjust(fmt['float'])
contrast_str = (fmt['fmt'] % (contrast)).rjust(fmt['float'])
table_entry = fmt['sep'].join([sweep_str,
pert_str,
sfa_str,
ica_str,
contrast_str])
return table_entry
def _get_eye(self):
# return an identity matrix with the right dimensions and type
return numx.eye(self._effective_input_dim, dtype=self.dtype)
def _get_rnd_rotation(self, dim):
# return a random rot matrix with the right dimensions and type
return mdp.utils.random_rot(dim, self.dtype)
def _get_rnd_permutation(self, dim):
# return a random permut matrix with the right dimensions and type
zero = numx.zeros((dim, dim), dtype=self.dtype)
row = numx_rand.permutation(dim)
for col in range(dim):
zero[row[col], col] = 1.
return zero
def _givens_angle(self, i, j, covs, bica_bsfa=None, complete=0):
# Return the Givens rotation angle for which the contrast function
# is minimal
if bica_bsfa is None:
bica_bsfa = self._bica_bsfa
if j < self.output_dim:
return self._givens_angle_case1(i, j, covs,
bica_bsfa, complete=complete)
else:
return self._givens_angle_case2(i, j, covs,
bica_bsfa, complete=complete)
def _givens_angle_case2(self, m, n, covs, bica_bsfa, complete=0):
# This function makes use of the constants computed in the paper
#
# R -> R
# m -> \mu
# n -> \nu
#
# Note that the minus sign before the angle phi is there because
# in the paper the rotation convention is the opposite of ours.
ncovs = covs.ncovs
covs = covs.covs
icaweights = self.icaweights
sfaweights = self.sfaweights
R = self.output_dim
bica, bsfa = bica_bsfa
Cmm, Cmn, Cnn = covs[m, m, :], covs[m, n, :], covs[n, n, :]
d0 = (sfaweights * Cmm*Cmm).sum()
d1 = 4*(sfaweights * Cmn*Cmm).sum()
d2 = 2*(sfaweights * (2*Cmn*Cmn + Cmm*Cnn)).sum()
d3 = 4*(sfaweights * Cmn*Cnn).sum()
d4 = (sfaweights * Cnn*Cnn).sum()
e0 = 2*(icaweights * ((covs[:R, m, :]*covs[:R, m, :]).sum(axis=0)
- Cmm*Cmm)).sum()
e1 = 4*(icaweights * ((covs[:R, m, :]*covs[:R, n, :]).sum(axis=0)
- Cmm*Cmn)).sum()
e2 = 2*(icaweights * ((covs[:R, n, :]*covs[:R, n, :]).sum(axis=0)
- Cmn*Cmn)).sum()
s22 = 0.25 * bsfa*(d1+d3) + 0.5* bica*(e1)
c22 = 0.5 * bsfa*(d0-d4) + 0.5* bica*(e0-e2)
s24 = 0.125* bsfa*(d1-d3)
c24 = 0.125* bsfa*(d0-d2+d4)
# Compute the contrast function in a grid of angles to find a
# first approximation for the minimum. Repeat two times
# (effectively doubling the resolution). Note that we can do
# that because we know we have a single minimum.
#
# npoints should not be too large otherwise the contrast
# funtion appears to be constant. This is because we hit the
# maximum resolution for the cosine function (ca. 1e-15)
npoints = 100
left = -PI/2 - PI/(npoints+1)
right = PI/2 + PI/(npoints+1)
for iter in (1, 2):
phi = numx.linspace(left, right, npoints+3)
contrast = c22*cos(-2*phi)+s22*sin(-2*phi)+\
c24*cos(-4*phi)+s24*sin(-4*phi)
minidx = contrast.argmin()
left = phi[max(minidx-1, 0)]
right = phi[min(minidx+1, len(phi)-1)]
# The contrast is almost a parabola around the minimum.
# To find the minimum we can therefore compute the derivative
# (which should be a line) and calculate its root.
# This step helps to overcome the resolution limit of the
# cosine function and clearly improve the final result.
der_left = 2*c22*sin(-2*left)- 2*s22*cos(-2*left)+\
4*c24*sin(-4*left)- 4*s24*cos(-4*left)
der_right = 2*c22*sin(-2*right)-2*s22*cos(-2*right)+\
4*c24*sin(-4*right)-4*s24*cos(-4*right)
if abs(der_left - der_right) < SQRT_EPS_D:
minimum = phi[minidx]
else:
minimum = right - der_right*(right-left)/(der_right-der_left)
dc = numx.zeros((ncovs,), dtype = self.dtype)
for t in range(ncovs):
dg = covs[:R, :R, t].diagonal()
dc[t] = (dg*dg).sum(axis=0)
dc = ((dc-Cmm*Cmm)*sfaweights).sum()
ec = numx.zeros((ncovs, ), dtype = self.dtype)
for t in range(ncovs):
ec[t] = sum([covs[i, j, t]*covs[i, j, t] for i in range(R-1)
for j in range(i+1, R) if i != m and j != m])
ec = 2*(ec*icaweights).sum()
a20 = 0.125*bsfa*(3*d0+d2+3*d4+8*dc)+0.5*bica*(e0+e2+2*ec)
minimum_contrast = a20+c22*cos(-2*minimum)+s22*sin(-2*minimum)+\
c24*cos(-4*minimum)+s24*sin(-4*minimum)
if complete:
# Compute the contrast between -pi/2 and pi/2
# (useful for testing purposes)
npoints = 1000
phi = numx.linspace(-PI/2, PI/2, npoints+1)
contrast = a20 + c22*cos(-2*phi) + s22*sin(-2*phi) +\
c24*cos(-4*phi) + s24*sin(-4*phi)
return phi, contrast, minimum, minimum_contrast
else:
return minimum, minimum_contrast
def _givens_angle_case1(self, m, n, covs, bica_bsfa, complete=0):
# This function makes use of the constants computed in the paper
#
# R -> R
# m -> \mu
# n -> \nu
#
# Note that the minus sign before the angle phi is there because
# in the paper the rotation convention is the opposite of ours.
ncovs = covs.ncovs
covs = covs.covs
icaweights = self.icaweights
sfaweights = self.sfaweights
bica, bsfa = bica_bsfa
Cmm, Cmn, Cnn = covs[m, m, :], covs[m, n, :], covs[n, n, :]
d0 = (sfaweights * (Cmm*Cmm+Cnn*Cnn)).sum()
d1 = 4*(sfaweights * (Cmm*Cmn-Cmn*Cnn)).sum()
d2 = 2*(sfaweights * (2*Cmn*Cmn+Cmm*Cnn)).sum()
e0 = 2*(icaweights * Cmn*Cmn).sum()
e1 = 4*(icaweights * (Cmn*Cnn-Cmm*Cmn)).sum()
e2 = (icaweights * ((Cmm-Cnn)*(Cmm-Cnn)-2*Cmn*Cmn)).sum()
s24 = 0.25* (bsfa * d1 + bica * e1)
c24 = 0.25* (bsfa *(d0-d2)+ bica *(e0-e2))
# compute the exact minimum
# Note that 'arctan' finds always the first maximum
# because s24sin(4p)+c24cos(4p)=const*cos(4p-arctan)
# the minimum lies +pi/4 apart (period = pi/2).
# In other words we want that: abs(minimum) < pi/4
phi4 = numx.arctan2(s24, c24)
# use if-structure until bug in numx.sign is solved
if phi4 >= 0:
minimum = -0.25*(phi4-PI)
else:
minimum = -0.25*(phi4+PI)
# compute all constants:
R = self.output_dim
dc = numx.zeros((ncovs, ), dtype = self.dtype)
for t in range(ncovs):
dg = covs[:R, :R, t].diagonal()
dc[t] = (dg*dg).sum(axis=0)
dc = ((dc-Cnn*Cnn-Cmm*Cmm)*sfaweights).sum()
ec = numx.zeros((ncovs, ), dtype = self.dtype)
for t in range(ncovs):
triu_covs = _triu(covs[:R, :R, t], 1).ravel()
ec[t] = ((triu_covs*triu_covs).sum() - covs[m, n, t]*covs[m, n, t])
ec = 2*(icaweights*ec).sum()
a20 = 0.25*(bsfa*(4*dc+d2+3*d0)+bica*(4*ec+e2+3*e0))
minimum_contrast = a20+c24*cos(-4*minimum)+s24*sin(-4*minimum)
npoints = 1000
if complete == 1:
# Compute the contrast between -pi/2 and pi/2
# (useful for testing purposes)
phi = numx.linspace(-PI/2, PI/2, npoints+1)
contrast = a20 + c24*cos(-4*phi) + s24*sin(-4*phi)
return phi, contrast, minimum, minimum_contrast
elif complete == 2:
phi = numx.linspace(-PI/4, PI/4, npoints+1)
contrast = a20 + c24*cos(-4*phi) + s24*sin(-4*phi)
return phi, contrast, minimum, minimum_contrast
else:
return minimum, minimum_contrast
def _get_contrast(self, covs, bica_bsfa = None):
if bica_bsfa is None:
bica_bsfa = self._bica_bsfa
# return current value of the contrast
R = self.output_dim
ncovs = covs.ncovs
covs = covs.covs
icaweights = self.icaweights
sfaweights = self.sfaweights
# unpack the bsfa and bica coefficients
bica, bsfa = bica_bsfa
sfa = numx.zeros((ncovs, ), dtype=self.dtype)
ica = numx.zeros((ncovs, ), dtype=self.dtype)
for t in range(ncovs):
sq_corr = covs[:R, :R, t]*covs[:R, :R, t]
sfa[t] = sq_corr.trace()
ica[t] = 2*_triu(sq_corr, 1).ravel().sum()
return (bsfa*sfaweights*sfa).sum(), (bica*icaweights*ica).sum()
def _adjust_ica_sfa_coeff(self):
# adjust sfa/ica ratio. ica and sfa term are scaled
# differently because sfa accounts for the diagonal terms
# whereas ica accounts for the off-diagonal terms
ncomp = self.output_dim
if ncomp > 1:
bica = self.sfa_ica_coeff[1]/(ncomp*(ncomp-1))
bsfa = -self.sfa_ica_coeff[0]/ncomp
else:
bica = 0.#self.sfa_ica_coeff[1]
bsfa = -self.sfa_ica_coeff[0]
self._bica_bsfa = [bica, bsfa]
def _fix_covs(self, covs=None):
# fiv covariance matrices
if covs is None:
covs = self.covs
if not self.whitened:
white = self.white
white.stop_training()
proj = white.get_projmatrix(transposed=0)
else:
proj = None
# fix and whiten the covariance matrices
for i in range(len(self.lags)):
covs[i], avg, avg_dt, tlen = covs[i].fix(proj)
# send the matrices to the container class
covs = MultipleCovarianceMatrices(covs)
# symmetrize the cov matrices
covs.symmetrize()
self.covs = covs
def _optimize(self):
# optimize contrast function
# save initial contrast
sfa, ica = self._get_contrast(self.covs)
self.initial_contrast = {'SFA': sfa,
'ICA': ica,
'TOT': sfa + ica}
# info headers
if self.verbose:
print self._info['header']+self._info['line']
# initialize control variables
# contrast
contrast = sfa+ica
# local rotation matrix
Q = self._get_eye()
# local copy of correlation matrices
covs = self.covs.copy()
# maximum improvement in the contrast function
max_increase = self.eps_contrast
# Number of sweeps
sweep = 0
# flag for stopping sweeping
sweeping = True
# flag to check if we already perturbed the outer space
# - negative means that we exit from this routine
# because we hit numerical precision or because
# there's no outer space to be perturbed (input_dim == outpu_dim)
# - positive means the number of perturbations done
# before finding no further improvement
perturbed = 0
# size of the perturbation matrix
psize = self._effective_input_dim-self.output_dim
# if there is no outer space don't perturbe
if self._effective_input_dim == self.output_dim:
perturbed = -1
# local eye matrix
eye = self._get_eye()
# main loop
# we'll keep on sweeping until the contrast has improved less
# then self.eps_contrast
part_sweep = 0
while sweeping:
# update number of sweeps
sweep += 1
# perform a single sweep
max_increase, covs, Q, contrast = self._do_sweep(covs, Q, contrast)
if max_increase < 0 or contrast == 0:
# we hit numerical precision, exit!
sweeping = False
if perturbed == 0:
perturbed = -1
else:
perturbed = -perturbed
if (max_increase < self.eps_contrast) and (max_increase) >= 0 :
# rate of change is small for all pairs in a sweep
if perturbed == 0:
# perturbe the outer space one time with a random rotation
perturbed = 1
elif perturbed >= 1 and part_sweep == sweep-1:
# after the last pertubation no useful step has
# been done. exit!
sweeping = False
elif perturbed < 0:
# we can't perturbe anymore
sweeping = False
# keep track of the last sweep we perturbed
part_sweep = sweep
# perform perturbation if needed
if perturbed >= 1 and sweeping is True:
# generate a random rotation matrix for the external subspace
PRT = eye.copy()
rot = self._get_rnd_rotation(psize)
# generate a random permutation matrix for the ext. subspace
perm = self._get_rnd_permutation(psize)
# combine rotation and permutation
rot_perm = mult(rot, perm)
# apply rotation+permutation
PRT[self.output_dim:, self.output_dim:] = rot_perm
covs.transform(PRT)
Q = mult(Q, PRT)
# increment perturbation counter
perturbed += 1
# verbose progress information
if self.verbose:
table_entry = self._fmt_prog_info(sweep,
perturbed,
contrast)
_sys.stdout.write(table_entry+len(table_entry)*'\b')
_sys.stdout.flush()
# if we made too many sweeps exit with error!
if sweep == self.max_iter:
err_str = ("Failed to converge, maximum increase= "
"%.5e" % (max_increase))
raise NodeException(err_str)
# if we land here, we have converged!
# calculate output contrast
sfa, ica = self._get_contrast(covs)
contrast = sfa+ica
# print final information
if self.verbose:
print self._fmt_prog_info(sweep, perturbed,
contrast, sfa, ica)
print self._info['line']
self.final_contrast = {'SFA': sfa,
'ICA': ica,
'TOT': sfa + ica}
# finally return optimal rotation matrix
return Q
def _do_sweep(self, covs, Q, prev_contrast):
# perform a single sweep
# initialize maximal improvement in a single sweep
max_increase = -1
# shuffle rotation order
numx_rand.shuffle(self.rot_axis)
# sweep through all axes combinations
for (i, j) in self.rot_axis:
# get the angle that minimizes the contrast
# and the contrast value
angle, contrast = self._givens_angle(i, j, covs)
if contrast == 0:
# we hit numerical precision in case when b_sfa == 0
# we can only break things from here on, better quit!
max_increase = -1
break
# relative improvement in the contrast function
relative_diff = (prev_contrast-contrast)/abs(prev_contrast)
if relative_diff < 0:
# if rate of change is negative we hit numerical precision
# or we already sit on the optimum for this pair of axis.
# don't rotate anymore and go to the next pair
continue
# update the rotation matrix
rotate(Q, angle, [i, j])
# rotate the covariance matrices
covs.rotate(angle, [i, j])
# store maximum and previous rate of change
max_increase = max(max_increase, relative_diff)
prev_contrast = contrast
return max_increase, covs, Q, contrast
def _stop_training(self, covs=None):
"""Stop the training phase.
If the node is used on large datasets it may be wise to first
learn the covariance matrices, and then tune the parameters
until a suitable parameter set has been found (learning the
covariance matrices is the slowest part in this case). This
could be done for example in the following way (assuming the
data is already white):
covs=[mdp.utils.DelayCovarianceMatrix(dt, dtype=dtype) for dt in lags]
for block in data:
[covs[i].update(block) for i in range(len(lags))]
You can then initialize the ISFANode with the desired parameters,
do a fake training with some random data to set the internal
node structure and then call stop_training with the stored covariance
matrices. For example:
isfa = ISFANode(lags, .....)
x = mdp.numx_rand.random((100, input_dim)).astype(dtype)
isfa.train(x)
isfa.stop_training(covs=covs)
This trick has been used in the paper to apply ISFA to surrogate
matrices, i.e. covariance matrices that were not learnt on a
real dataset.
"""
# fix, whiten, symmetrize and weight the covariance matrices
# the functions sets also the number of input components self.ncomp
self._fix_covs(covs)
# if output_dim were not set, set it to be the number of input comps
if self.output_dim is None:
self.output_dim = self._effective_input_dim
# adjust b_sfa and b_ica
self._adjust_ica_sfa_coeff()
# initialize all possible rotation axes
self.rot_axis = [(i, j) for i in range(0, self.output_dim)
for j in range(i+1, self._effective_input_dim)]
# initialize the global rotation-permutation matrix (RP):
RP = self.RP
if RP is None:
RP = self._get_eye()
else:
# apply the global rotation matrix
self.covs.transform(RP)
# find optimal rotation
Q = self._optimize()
RP = mult(RP, Q)
# rotate and permute the covariance matrices
# we do it here in one step, to avoid the cumulative errors
# of multiple rotations in _optimize
self.covs.transform(Q)
# keep the complete rotation-permutation matrix
self.RPC = RP.copy()
# Reduce dimension to match output_dim#
RP = RP[:, :self.output_dim]
# the variance for the derivative of a whitened signal is
# 0 <= v <= 4, therefore the diagonal elements of the delayed
# covariance matrice with time lag = 1 (covs[:,:,0]) are
# -1 <= v' <= +1
# reorder the components to have them ordered by slowness
d = (self.covs.covs[:self.output_dim, :self.output_dim, 0]).diagonal()
idx = d.argsort()[::-1]
self.RP = RP.take(idx, axis=1)
# we could in principle clean up self.covs, as we do in SFANode or
# PCANode, but this algorithm is not stable enough to rule out
# possible problems. When these occcurs examining the covariance
# matrices is often the only way to debug.
#del self.covs
|