def _get_gabor_filters(self, params):
""" Return a Gabor filterbank (generate it if needed)
Inputs:
params -- filters parameters (dict)
Outputs:
filt_l -- filterbank (list)
"""
if self._filt_l is None:
# -- get parameters
fh, fw = params['kshape']
orients = params['orients']
freqs = params['freqs']
phases = params['phases']
nf = len(orients) * len(freqs) * len(phases)
fbshape = nf, fh, fw
gsw = fw/5.
gsh = fw/5.
xc = fw/2
yc = fh/2
filt_l = []
# -- build the filterbank
for freq in freqs:
for orient in orients:
for phase in phases:
# create 2d gabor
filt = v1s_math.gabor2d(gsw,gsh,
xc,yc,
freq,orient,phase,
(fw,fh))
# vectors for separable convolution
U,S,V = v1s_math.fastsvd(filt)
tot = 0
vectors = []
idx = 0
S **= 2.
while tot <= params['sep_threshold']:
row = (U[:,idx]*S[idx])[:, newaxis]
col = (V[idx,:])[newaxis, :]
vectors += [(row,col)]
tot += S[idx]/S.sum()
idx += 1
filt_l += [vectors]
self._filt_l = filt_l
return self._filt_l
def _dimr(self,fvectors,pca_threshold):
# -- Reduce dimensionality using a pca / eigen subspace projection
nvectors, vsize = fvectors.shape
if nvectors < vsize:
print "pca...",
print fvectors.shape, "=>",
U,S,V = v1s_math.fastsvd(fvectors)
eigvectors = V.T
i = tot = 0
S **= 2.
while (tot <= pca_threshold) and (i < S.size):
tot += S[i]/S.sum()
i += 1
eigvectors = eigvectors[:, :i+1]
fvectors = scipy.dot(fvectors, eigvectors)
return fvectors, eigvectors
