def _train(self, ds):
# local binding
method = self.__method
# two major modes
if method == 'union':
# slice mask default: take none
mask = np.zeros(ds.shape[1], dtype=np.bool)
# method: OR
cfunc = np.logical_or
elif method == 'intersection':
# slice mask default: take all
mask = np.ones(ds.shape[1], dtype=np.bool)
# method: AND
cfunc = np.logical_and
else:
raise ValueError("Unknown combining method '%s'" % method)
for fs in self.__selectors:
# first: train all embedded selections
fs.train(ds)
# now get boolean mask of selections
fsmask = np.zeros(mask.shape, dtype=np.bool)
# use slicearg to select features
fsmask[fs._slicearg] = True
# merge with current global mask
mask = cfunc(mask, fsmask)
# turn the derived boolean mask into a slice if possible
slicearg = mask2slice(mask)
# and assign to baseclass, done
self._safe_assign_slicearg(slicearg)
def _call(self, ds):
# attribute to detect boundaries
battr = ds.sa[self.get_space()].value
# filter which samples to keep
filter_ = np.ones(battr.shape, dtype='bool')
# determine boundary indices -- shift by one to have the new value
# as the boundary
bindices = (battr[:-1] != battr[1:]).nonzero()[0] + 1
# for all boundaries
for b in bindices:
lower = b - self._prestrip
upper = b + self._poststrip
filter_[lower:upper] = False
filter_ = mask2slice(filter_)
return ds[filter_]
def _call(self, ds):
# attribute to detect boundaries
battr = ds.sa[self.get_space()].value
# filter which samples to keep
filter_ = np.ones(battr.shape, dtype='bool')
# determine boundary indices -- shift by one to have the new value
# as the boundary
bindices = (battr[:-1] != battr[1:]).nonzero()[0] + 1
# for all boundaries
for b in bindices:
lower = b - self._prestrip
upper = b + self._poststrip
filter_[lower:upper] = False
filter_ = mask2slice(filter_)
return ds[filter_]
