def sample(self, n, seed=900):
pdata = self.pdata
n_sam = pdata.sample_size()
if n > n_sam:
raise ValueError("Cannot subsample %d points from %d points." %
(n, n_sam))
X, Y = pdata.xy()
# permute X, Y. Keep pairs
I = util.subsample_ind(n_sam, n_sam, seed=seed + 3)
X = X[I, :]
Y = Y[I, :]
perm_pivot = self.pivot[I]
list_chosenI = []
for ui, v in enumerate(self._uniques):
Iv = np.nonzero(np.abs(perm_pivot - v) <= 1e-8)
Iv = Iv[0]
niv = self._counts[ui]
# ceil guarantees that at least 1 instance will be chosen
# from each class.
n_class = int(math.ceil(niv / n_sam * n))
chosenI = Iv[:n_class]
list_chosenI.append(chosenI)
final_chosenI = np.hstack(list_chosenI)
reduceI = util.subsample_ind(len(final_chosenI),
min(n, len(final_chosenI)), seed + 5)
final_chosenI = final_chosenI[reduceI]
assert len(final_chosenI) == n, (
"final_chosenI has length %d which is not n=%d" %
(len(final_chosenI), n))
Xsam = X[final_chosenI, :]
Ysam = Y[final_chosenI, :]
new_label = None if pdata.label is None else pdata.label + "_stra"
return PairedData(Xsam, Ysam, label=new_label)
def subsample(self, n, seed=87):
"""Subsample without replacement. Return a new PairedData """
if n > self.X.shape[0] or n > self.Y.shape[0]:
raise ValueError('n should not be larger than sizes of X, Y.')
ind_x = util.subsample_ind(self.X.shape[0], n, seed)
ind_y = util.subsample_ind(self.Y.shape[0], n, seed)
return PairedData(self.X[ind_x, :], self.Y[ind_y, :], self.label)
def sample(self, n, seed=981):
if n > self.pdata.sample_size():
raise ValueError('cannot sample more points than what the original dataset has')
X, Y = self.pdata.xy()
if n == 1:
ind = util.subsample_ind(self.pdata.sample_size(), 2, seed=seed)
nX = X[[ind[0]], :]
nY = Y[[ind[1]], :]
else:
ind = util.subsample_ind(self.pdata.sample_size(), n, seed=seed)
nX = X[ind, :]
ind_shift1 = np.roll(ind, 1)
nY = Y[ind_shift1, :]
new_label = None if self.pdata.label is None else self.pdata.label + '_shuf'
return PairedData(nX, nY, label=new_label)