def revert(self, ds):
for name, mins, maxs in zip(self.var_names, self.mins, self.maxs):
stacked, stack_info = util.to_stacked_array(ds[[name]])
reverted = .5 * (np.tanh(stacked) + 1.)
ds_unnormalized = reverted * (maxs - mins) + mins
unstacked = util.to_unstacked_dataset(ds_unnormalized.values,
stack_info)
ds = ds.assign({name: unstacked[name]})
return ds
def test_stacking_roundtrip():
a = np.arange(12).reshape(2, 6)
b = np.arange(12, 24).reshape(2, 2, 3)
c = np.arange(2)
ds = xr.Dataset({
'a': (('sample', 'foo'), a),
'b': (('sample', 'ff', 'bb'), b),
'c': (('sample'), c)
})
arr, info = to_stacked_array(ds)
ds_u = to_unstacked_dataset(arr.values, info)
assert ds.identical(ds_u)
def revert(self, ds):
for name, lmbda, shifting_factor_per_feature, boundary_location in \
zip(self.var_names, self.lmbdas, self.shifting_factors, self.boundary_locations):
stacked, stack_info = util.to_stacked_array(ds[[name]])
# This does not guarantee that the generated samples are above zero.
reverted = inv_boxcox(stacked, lmbda) - shifting_factor_per_feature
unstacked = util.to_unstacked_dataset(reverted.values, stack_info)
ds = ds.assign({name: unstacked[name]})
if boundary_location == 'right':
ds = ds.assign({name: -ds[name]})
return ds
def apply(self, ds):
assert not self.mins, 'This function cannot be called twice.'
ds = ds.astype('float64')
for name in self.var_names:
sample_dim = ds[name].dims[0]
stacked, stack_info = util.to_stacked_array(ds[[name]])
mins = stacked.min(sample_dim) - NUMERICAL_OFFSET # feature
maxs = stacked.max(sample_dim) + NUMERICAL_OFFSET # feature
self.mins.append(mins)
self.maxs.append(maxs)
ds_normalized = (stacked - mins) / (maxs - mins)
transformed = np.arctanh(2 * ds_normalized - 1)
unstacked = util.to_unstacked_dataset(transformed.values,
stack_info)
ds = ds.assign({name: unstacked[name]})
return ds
def apply(self, ds):
assert not self.shifting_factors, 'This function cannot be called twice.'
ds = ds.astype('float64')
for name, lmbda, boundary_location in \
zip(self.var_names, self.lmbdas, self.boundary_locations):
if boundary_location == 'right':
ds = ds.assign({name: -ds[name]})
sample_dim = ds[name].dims[0]
stacked, stack_info = util.to_stacked_array(ds[[name]])
mins = stacked.min(sample_dim) # feature
shifting_factor_per_feature = abs(
mins) + NUMERICAL_OFFSET # feature
shifting_factor_per_feature.load()[mins >= NUMERICAL_OFFSET] = 0.
self.shifting_factors.append(shifting_factor_per_feature)
transformed = boxcox(stacked + shifting_factor_per_feature, lmbda)
unstacked = util.to_unstacked_dataset(transformed.values,
stack_info)
ds = ds.assign({name: unstacked[name]})
return ds