I am using LassoLars to solve a multi-dimensional sparse coding problem. After upgrading from 0.12.1 to 0.13, I have encountered a bug where the coef_ attribute of LassoLars is a list, not a numpy array. This breaks the decision_function method, since it takes the transpose of coef_:

/export/disk0/wb/python2.6/lib/python2.6/site-packages/sklearn/linear_model/base.pyc in decision_function(self, X)
138 """
139 X = safe_asarray(X)
--> 140 return safe_sparse_dot(X, self.coef_.T) + self.intercept_
141
142 def predict(self, X):

AttributeError: 'list' object has no attribute 'T'

This seems to be due to this change that made it into the 0.13 release by @GaelVaroquaux:
e18465d