class DkuLassoLarsRegressor(BaseEstimator):
def __init__(self, max_var=0):
self.max_var = max_var
self.lars = None
self.X_offset = None
self.y_offset = None
self.X_scale = None
self.coef_ = None
self.current_index = None
self.intercept_ = None
self.coef_path_ = None
def fit(self, X, y):
# note: for now we perform rescaling. While this requires some more computation on our part, it has better
# numerical stability (could test with or without)
self.lars = LassoLars(alpha=0.0).fit(X, y)
# we recreate the rescaling
_, _, self.X_offset, self.y_offset, self.X_scale = self.lars._preprocess_data(
X, y, True, True, True)
# we normalize the coef path here
self.coef_path_ = [x / self.X_scale for x in self.lars.coef_path_.T]
self.coef_ = self.lars.coef_
self.intercept_ = self.lars.intercept_
self.alphas = self.lars.alphas_
if self.max_var > 0:
self._perform_cut(self.max_var)
return self
def _perform_cut(self, n):
n = min(n, self.lars.coef_path_.shape[1] - 1)
self.current_index = n
# note: not normalized, this is normal since the _set_intercept will normalize it
coef = self.lars.coef_path_[:, n]
self.lars.coef_ = coef
# recompute the intercept and normalize coefficients using scikit private method
self.lars._set_intercept(self.X_offset, self.y_offset, self.X_scale)
self.coef_ = self.lars.coef_
def post_process(self, user_meta):
if self.current_index is not None:
n = self.current_index
else:
n = self.max_var
n = user_meta.get("lars_cut", n)
if n > 0:
self._perform_cut(n)
def predict(self, X):
return self.lars.predict(X)
