def test_zmuv_rows():
arr = np.random.randn(10000, 100).astype(np.float32)
# -- numpy reference
gt = (arr - arr.mean(0)) / arr.std(0)
# -- cython
rows_mean, rows_std = zmuv_rows(arr.T)
gv = (arr - rows_mean) / rows_std
assert_allclose(gt, gv, rtol=RTOL, atol=ATOL)
# inplace
zmuv_rows_apply_inplace(arr.T, rows_mean, rows_std)
assert_allclose(gt, arr, rtol=RTOL, atol=ATOL)
def partial_fit(self, X):
_check_X(X)
rows_mean, rows_std = zmuv_rows(X.T)
if self._rows_mean is None:
self._rows_mean = rows_mean
self._rows_std = rows_std
self._n_samples = len(X)
else:
# XXX: change this, we should use sum and ssq in
# cython-based helpers
n_samples_new = self._n_samples + len(X)
rows_mean_new_n = (
self._n_samples * self._rows_mean
+
len(X) * rows_mean
)
rows_mean_new = rows_mean_new_n / n_samples_new
rows_ssq_new = (
self._n_samples * ((self._rows_std ** 2.)
+ self._rows_mean ** 2.)
+
len(X) * ((rows_std ** 2.) + rows_mean ** 2.)
)
rows_std_new = np.sqrt(
rows_ssq_new / n_samples_new
-
rows_mean_new ** 2.
)
self._rows_mean = rows_mean_new.astype('float32')
self._rows_std = rows_std_new.astype('float32')
self._n_samples = n_samples_new
return self
def fit(self, X):
_check_X(X)
self._rows_mean, self._rows_std = zmuv_rows(X.T)
return self